A cultivated land intelligent supervision operation and maintenance optimization method based on low-altitude unmanned aerial vehicle remote sensing network

Abstract

This invention relates to the field of operation and maintenance optimization technology, and discloses a method for intelligent monitoring and operation and maintenance optimization of cultivated land based on a low-altitude UAV remote sensing network. This method includes deploying a hexagonal topology UAV remote sensing network, fusing multi-source heterogeneous data to construct a four-dimensional spatiotemporal cube, generating high-confidence change patches and performing fine-grained classification through a dual-temporal neural network, establishing a three-level dynamic early warning mechanism, linking task assignment and rectification acceptance, forming a monitoring record chain, and iteratively optimizing the detection model based on localized factors and causal analysis. This application can achieve highly timely, high-precision, and closed-loop intelligent monitoring of cultivated land changes.

Description

Technical Field

[0001] This invention relates to the field of operation and maintenance optimization technology, and more specifically, to a method for intelligent monitoring and operation and maintenance optimization of farmland based on a low-altitude unmanned aerial vehicle (UAV) remote sensing network. Background Technology

[0002] With the deepening of farmland supervision, especially in remote sensing monitoring and intelligent perception, the demand for high-frequency, high-precision, and closed-loop supervision of farmland use changes is becoming increasingly urgent. Currently, farmland supervision mainly relies on a multi-source data system encompassing air, space, and ground, with satellite remote sensing as the backbone and manual inspections as a supplement, and ground sensors introduced in some areas for auxiliary monitoring. Standardized operating procedures have been established in land surveys and inspections. However, this model still has room for improvement in terms of timeliness, identification precision, and collaborative handling: satellite remote sensing is constrained by revisit cycles and weather conditions, resulting in time delays in image acquisition; manual inspections have limited coverage in complex terrain areas, and interpretation consistency is easily affected by subjective factors; multi-source data has not yet achieved efficient fusion in terms of spatiotemporal benchmarks, resolution, and semantic levels; the identification of change patches largely relies on empirical rules or simple index models, with limited ability to distinguish complex violation scenarios; and the problem handling process largely relies on offline methods, lacking a full-process traceability and image verification mechanism.

[0003] Therefore, how to integrate low-altitude UAV remote sensing networks, construct a unified spatiotemporal data structure, improve the semantic granularity of change recognition, and establish an intelligent closed-loop supervision mechanism for early warning, dispatching, verification, acceptance, and optimization has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a method for intelligent monitoring and operation optimization of arable land based on low-altitude UAV remote sensing network, which solves the technical problem that existing technologies cannot achieve high-frequency, high-precision, and closed-loop monitoring of changes in arable land use.

[0005] This invention provides a method for intelligent monitoring and operation optimization of cultivated land based on a low-altitude unmanned aerial vehicle (UAV) remote sensing network, comprising: acquiring a network deployment parameter set containing equipment performance parameters and spatial topology configuration information to deploy a remote sensing network covering a target cultivated land area; using the remote sensing network, conducting regular patrols and emergency monitoring of the target cultivated land area based on the spatial coordinate information in the network deployment parameter set to acquire multi-source heterogeneous remote sensing data to form a remote sensing dataset; performing preprocessing and fusion processing on the remote sensing dataset under a unified spatiotemporal reference to construct a data structure and record the registration accuracy index during the fusion process; based on the data structure, utilizing dual-time... A phase-depth semantic segmentation and change detection model analyzes farmland use status, dynamically adjusts feature extraction weights based on the registration accuracy index to generate high-confidence farmland change patches, and outputs the confidence value of each patch. Multispectral index features, texture features, geometric features, and contextual semantic features of the farmland change patches are extracted from the data structure, and fine-grained semantic classification is performed in conjunction with the confidence values ​​to identify specific violation types and output classification results containing violation type labels and feature vectors. Based on the violation type labels and patch areas in the classification results, a three-level dynamic early warning mechanism is established in conjunction with the confidence values, and a preset early warning threshold is set. The system triggers corresponding warning signals for farmland violations of varying severity. These warning signals, along with the classification results, are transformed into structured farmland supervision tasks. Responsible entities are matched using administrative boundaries and ownership information from the remote sensing dataset. Tasks are then dispatched to enforcement terminals at the corresponding levels according to preset response time limits. The system receives on-site verification and rectification feedback information for the farmland supervision tasks. Based on the network deployment parameter set, the remote sensing network is scheduled to perform re-flight verification. The re-flight images are compared with the original change images in the data structure to output a rectification acceptance conclusion containing multi-dimensional evaluation indicators. This conclusion is then archived to construct a traceable supervision record. Based on historical map data in the regulatory record chain and multi-dimensional evaluation indicators in the rectification and acceptance conclusions, a clustering algorithm is used to classify farmland risk and divide regulatory sub-regions. Time series analysis is combined to identify farmland use trend characteristics in each sub-region. Localized influencing factors and remote sensing identification deviation data are integrated. After filtering out interference variables, real emerging influencing factors are extracted and added as new input features to the data structure to provide feedback to optimize the parameters of the dual-temporal deep semantic segmentation and change detection model and complete iterative training. At the same time, the inspection frequency configuration in the network deployment parameter set is dynamically adjusted according to the risk classification results.

[0006] Furthermore, this includes: planning the spatial locations and coverage areas of several fixed take-off and landing stations based on topography, accessibility, power and communication infrastructure, and the distribution of historical violation hotspots on farmland, to construct remote sensing network nodes for farmland monitoring; configuring at least one vertical take-off and landing fixed-wing UAV for each of the several fixed take-off and landing stations, determining its maximum endurance time to be no less than a first preset duration threshold, its cruising speed to be no less than a first preset speed threshold, and its payload including a visible light camera, a multispectral imager, and a lidar, and recording the equipment performance parameters of each UAV; wherein, the ground sampling distance of the visible light camera at a preset flight altitude is no greater than a first preset accuracy threshold, and the multispectral imager includes multiple preset... The laser radar is configured with a spectral band and a spectral resolution better than a second preset accuracy threshold. The point cloud density of the laser radar at the preset flight altitude is not lower than a first preset density threshold. Unmanned aerial vehicles (UAVs) are installed at the fixed take-off and landing site using a fixed support structure welded from steel of a preset material, connected to a concrete base via chemical anchors. The wind resistance level is higher than or equal to a preset wind resistance threshold based on the terrain features of the site. Each UAV is equipped with an independent communication unit and a self-organizing network antenna, establishing a long-term connection with a cloud-based farmland monitoring platform via a preset communication protocol. The system uploads equipment performance parameters such as battery power, door status, temperature and humidity, and flight logs in real time to update the network deployment parameter set. The UAVs are arranged in a cellular hexagonal topology. The field is deployed according to the aforementioned spatial location, with the distance between adjacent nodes controlled to not exceed a preset distance threshold based on the coverage area. The topology deployment spatial coordinate information, including the device performance parameters, is recorded in the network deployment parameter set. All UAVs are connected to a unified flight mission scheduling platform. Based on real-time device battery level and door status information in the network deployment parameter set, three farmland monitoring operation modes are supported: on-demand triggering, periodic patrol, and emergency response. When the battery level of an airport falls below a first preset battery threshold or is under maintenance, the farmland monitoring task is migrated to a nearby available airport based on the topology deployment spatial coordinate information. The flight mission scheduling platform periodically acquires data with resolutions better than a preset threshold from the National Remote Sensing Image Coordination Service Platform. The system acquires optical satellite images of cultivated land areas at preset time intervals. Through the low-altitude UAV remote sensing network, based on the topology deployment spatial coordinate information in the network deployment parameter set and the mission instructions of the flight mission scheduling platform, it calls upon the onboard visible light camera, multispectral imager, and lidar to perform routine patrols of key cultivated land monitoring areas at preset frequencies under clear weather conditions, and to perform emergency re-photographing within preset response time limits for cultivated land warning areas, thereby acquiring low-altitude UAV aerial images. Through soil moisture, meteorological, and video monitoring sensor nodes deployed around the fixed take-off and landing station, it collects real-time data on cultivated land soil moisture content, temperature, precipitation, wind speed, and fixed-point video streams synchronized with the low-altitude UAV aerial images.Simultaneously acquire vector data of administrative division boundaries, land use status maps, permanent basic farmland protection zone boundaries, and land ownership information corresponding to the coverage area; integrate the satellite imagery, the low-altitude UAV aerial imagery, the real-time data collected by the sensor nodes, and the vector data into the multi-source heterogeneous remote sensing dataset, and add timestamps, data source identifiers, and the spatial coordinate information of the topological layout.

[0007] Furthermore, this includes: integrating several drones into a unified scheduling task; supporting three operation modes—on-demand triggering, periodic patrol, and emergency response—based on the device status information and topology coordinate information; migrating the task to a neighboring node based on the topology coordinate information when a node's power or status is abnormal; periodically acquiring high-resolution satellite images from a remote sensing image service platform through the unified scheduling platform; and directing drones to utilize visible light cameras, multispectral imagers, and lidar based on the topology coordinate information to perform routine patrols and emergency monitoring in key and early warning areas, acquiring multiple aerial photography data to obtain an aerial photography dataset; collecting sensor data in real time that is spatiotemporally synchronized with the aerial photography dataset through soil moisture, meteorological, and video monitoring sensor nodes deployed around the fixed take-off and landing station, and simultaneously acquiring vector data of administrative division boundaries, land use status, and ownership information corresponding to the coverage area; integrating the satellite images, aerial photography data, sensor data, and vector data into a multi-source heterogeneous remote sensing dataset, and adding timestamps, data source identifiers, and the topology coordinate information as spatiotemporal identifiers. Furthermore, the remote sensing dataset undergoes spatiotemporal benchmark unification and fusion processing to construct a four-dimensional spatiotemporal cube data structure. This includes: uniformly converting several farmland area image data from the multi-source heterogeneous remote sensing dataset to a preset geodetic coordinate system and a preset elevation benchmark, resulting in an image dataset under a unified coordinate system; performing radiometric calibration and atmospheric correction on satellite images and low-altitude UAV aerial images in the image dataset to eliminate the influence of atmospheric scattering and sensor differences, resulting in a radiometrically calibrated and atmospherically corrected image data processing set; and employing a joint spatial solution method based on feature point matching and regional network adjustment to register farmland area images acquired at different times and platforms in the image data processing set to the same geographic grid, with the registration error controlled within a certain range. Within a preset registration error threshold, the actual registration error is recorded as the registration accuracy index to obtain a spatially registered image data registration set. Multiple ground sensor data points in the remote sensing dataset are interpolated by timestamps to align the interpolated ground sensor data with the time series of the image data registration set, resulting in a time-aligned sensor dataset. A four-dimensional spatiotemporal cube data structure is constructed based on the preset geodetic coordinate system and preset elevation datum. The three spatial dimensions of the data structure are east, north, and elevation, and the time dimension is the acquisition time. Each spatiotemporal cell stores the image data registration set, the sensor dataset, and metadata, including the registration accuracy index and data source identifier.

[0008] Further, the method includes: constructing a dual-temporal input encoder and decoder neural network architecture; wherein the encoder adopts an improved convolutional neural network backbone network; the decoder adopts an upsampling network with a skip connection structure, forming two parallel encoder branches; extracting spatiotemporal cube slices of cultivated land areas from the data structure for the current and previous temporal phases, and inputting them into the two parallel encoder branches of the dual-temporal input encoder and decoder neural network architecture for feature extraction; introducing channel attention and spatial attention mechanisms at the encoder high-level feature maps output by the two parallel encoder branches, and dynamically adjusting the attention weights according to the registration accuracy index in the data structure to obtain a weight-adjusted feature map; based on the feature map, fusing dual-temporal information through two strategies, feature difference and feature concatenation, to generate a preliminary change probability map of cultivated land; applying morphological closing operations and connected component analysis to the preliminary change probability map, removing noise patches with areas smaller than a preset area threshold to output a final set of cultivated land change patches, and calculating and recording the confidence value for each patch in the set of cultivated land change patches; and targeting the cultivated land change patch set... For each farmland change patch, based on the spatiotemporal cube slices of the farmland area in the current and previous time phases, multispectral index features, texture features, geometric features, and contextual semantic features are extracted from the data structure within the internal and surrounding buffer zones to form a feature vector. The multispectral index features are calculated based on multispectral band data in the spatiotemporal cube slices, including the normalized vegetation index, soil-regulated vegetation index, and building index. The texture features are calculated based on the gray-level co-occurrence matrix of the spatiotemporal cube slices, including contrast, correlation, and homogeneity. The geometric features are calculated based on the spatial morphology of the farmland change patch, including the aspect ratio, compactness, main direction, and concavity / convexity. The contextual semantic features are obtained based on the distribution statistics of other land cover types within the surrounding buffer zone of the farmland change patch. The feature vector and the confidence value are used as input to an ensemble learning classifier composed of a gradient boosting decision tree and a multilayer perceptron, classifying each farmland change patch in the set into a preset number of violation type categories, and outputting a classification result containing the violation type label and the feature vector.

[0009] Furthermore, based on the classification results and confidence values, a three-level dynamic early warning mechanism is established, triggering different levels of early warning signals according to preset thresholds, including: setting the first-level early warning threshold as suspected farmland abandonment behavior with a single identification area smaller than a first preset area threshold and no signs of hardening, based on the violation type label in the classification results; setting the second-level early warning threshold as farmland facility agriculture encroachment behavior with an area within the range of the first to second preset area thresholds or with temporary structures; setting the third-level early warning threshold as illegal occupation of farmland with an area larger than the second preset area threshold, a surface hardening rate exceeding a preset hardening rate threshold, or involving permanent buildings; when the classification results... When a farmland violation patch meets any of the first, second, or third warning threshold conditions, a structured warning message of the corresponding level is generated based on the violation type label and the patch area. The structured warning message includes the patch ID, center point latitude and longitude, area, confidence value, warning level, and a thumbnail of multi-temporal imagery obtained from a low-altitude UAV remote sensing network. The structured warning message is pushed to the mobile terminal of the Institute of Natural Resources through the government intranet API interface and simultaneously written into the farmland protection supervision database, recording the warning level, response time, check-in location of the inspectors, and timestamp of the rectification photo upload for each farmland warning. Furthermore, the early warning signals and classification results are transformed into structured regulatory tasks and dispatched to responsible entities according to response time limits. This includes: matching the relevant natural resources authorities based on the geographical location of the illegal farmland patches, combined with the administrative boundaries and ownership information in the multi-source heterogeneous remote sensing dataset, to determine the corresponding responsible entity; determining the urgency level and processing time limit of the farmland regulatory task based on the violation type label in the classification results and the area size of the illegal farmland patches; wherein, the default violation type is marked as a level one emergency label and responds within a first preset time threshold, and the regular type is marked as a level two label, requiring a second preset time threshold. The system sets a response time threshold; generates an electronic task order containing the location of the illegally cultivated land patch, an image screenshot obtained from a low-altitude UAV remote sensing network, the classification result, the warning signal level, the emergency level, the processing time limit, legal basis, and processing suggestions; pushes the electronic task order to the mobile enforcement terminal of the corresponding level of the natural resources authority according to the responsible entity, and records the dispatch time and reception status; if no processing confirmation for the electronic task order is received within the processing time limit, an alarm is escalated according to the level of the responsible entity and the superior regulatory department is notified.

[0010] Furthermore, the system receives on-site verification and rectification feedback information, and confirms the rectification effectiveness through the re-flight verification mechanism of the remote sensing network, outputting a rectification acceptance conclusion containing multi-dimensional evaluation indicators. This includes: receiving photos, videos, text descriptions, and GPS positioning information of the farmland rectification site uploaded by grassroots law enforcement personnel via mobile law enforcement terminals, forming on-site verification feedback information; comparing the GPS positioning information in the on-site verification feedback information with the original farmland patch location in the electronic task sheet; if the spatial location deviation exceeds a preset location deviation threshold, it is considered invalid feedback; if the deviation is within the threshold range, it is confirmed as valid verification feedback; for farmland supervision tasks marked as completed rectification in the valid verification feedback, based on the patch location information in the electronic task sheet and the spatial coordinate information in the network deployment parameter set, the remote sensing network is scheduled to perform multiple re-inspection flights on the rectified plots at preset re-inspection time nodes after the rectification task is issued; in each re-inspection flight, multispectral images of the rectified plots are acquired, the normalized vegetation index and bare soil index are calculated, and compared with the baseline value before rectification recorded in the data structure; if the normalized vegetation index increases for a consecutive preset number of times, the system will be considered valid feedback. If the increase in the soil mass exceeds a first preset threshold and the decrease in the bare soil index exceeds a first preset threshold, it is determined to be effective land reclamation. The latest multispectral imagery acquired during the re-inspection flight is compared pixel-level with the original change imagery in the data structure to calculate vegetation cover restoration rate, building demolition integrity, and surface disturbance index as multi-dimensional evaluation indicators. Based on the point cloud data acquired during the re-inspection flight, an oblique photogrammetry model is used to calculate the standard deviation of the cultivated land surface elevation of the remediated plots. If the standard deviation of the surface elevation is less than a preset standard deviation threshold, the terrain is considered to have been restored to flatness, and the surface elevation is... The elevation standard deviation is included in the multi-dimensional evaluation indicators. If the vegetation cover restoration rate is greater than the first preset restoration rate threshold, the building demolition integrity is greater than the second preset restoration rate threshold, the surface disturbance index is less than the preset proportion threshold of the initial value, and the surface elevation standard deviation meets the standard, then the farmland rectification is deemed qualified, and an electronic acceptance form containing the multi-dimensional evaluation indicators is generated as the rectification acceptance conclusion. After confirmation by the administrator, it is archived to the blockchain storage node. Otherwise, a second farmland rectification instruction is generated and reissued based on the non-compliant items in the multi-dimensional evaluation indicators.Furthermore, based on the historical plot data and rectification results in the aforementioned regulatory record chain, a clustering algorithm is used to classify the target cultivated land area into risk levels, dividing it into cultivated land supervision sub-regions. Time series analysis is then used to identify the cultivated land use trend characteristics of each sub-region. This includes: using the target cultivated land area as the basic unit, aggregating the rectification cycle, recurrence frequency, and trend direction of all historical cultivated land violation plots within the target cultivated land area from the cultivated land supervision digital ledger; combining this with the multi-dimensional evaluation indicators in the rectification acceptance conclusion to form a cultivated land risk feature dataset containing violation characteristic data; inputting the cultivated land risk feature dataset into a K-means clustering algorithm, and dividing the target cultivated land area into a preset number of risk level area categories based on the rectification cycle, recurrence frequency, trend direction, and multi-dimensional evaluation indicators. Each category corresponds to a different cultivated land violation characteristic pattern and is marked as high-risk cultivated land. The risk level classification includes high-risk, medium-risk, and low-risk farmland areas. For each sub-region within the risk level classification, differentiated inspection frequencies are set based on the remote sensing network. Specifically, high-risk farmland areas are inspected at a fourth preset frequency, medium-risk farmland areas at a fifth preset frequency, and low-risk farmland areas at a sixth preset frequency. A dedicated AI identification strategy is configured for each sub-region within the risk level classification. In high-risk farmland areas, a temporal deep learning model is used to capture subtle, gradual changes in farmland. In low-risk farmland areas, soil moisture and vegetation succession sequences are analyzed based on the violation characteristic patterns. The identification strategy parameters corresponding to each risk level classification are fed back to a dual-temporal deep semantic segmentation and change detection model. Simultaneously, the differentiated inspection frequency configuration is updated to the network deployment parameter set.

[0011] Furthermore, based on the farmland use trend characteristics of the sub-regions corresponding to the risk level categories and the violation pattern characteristics in the regulatory record chain, on-site farmland surveys are conducted for the high-risk, medium-risk, and low-risk farmland areas to collect a set of localized influencing factors, including land ownership status, farmers' planting intentions, and the frequency of surrounding construction activities. Analysis of variance is performed on each influencing factor in the set of localized influencing factors. Influencing factors whose target variance is less than a preset variance threshold are removed according to a preset proportion threshold, resulting in a set of non-constant influencing factors. Pearson correlation coefficients are calculated on the remaining factors in the set of non-constant influencing factors. When the absolute value of the correlation coefficient between two factors is greater than a preset correlation coefficient threshold, influencing factors whose difference in correlation coefficient with the area of ​​farmland violation patches in the regulatory record chain is less than a preset difference are retained, resulting in a set of deredundant influencing factors. Policy factors in the set of deredundant influencing factors are then analyzed... The event window is truncated, retaining only the observations within a preset time window before and after the event, resulting in a set of influence factors filtered by the time window. This set of influence factors is then input into a feature importance ranking algorithm to remove weak contribution factors with importance scores below a preset importance threshold, resulting in a final set of high contribution influence factors. This set of high contribution influence factors is then input into a causal inference model to identify core driving variables that have a causal effect on farmland changes in each sub-region of the risk level category, and these core driving variables are added as new input features to the data structure. Using the data structure with the added core driving variables, the dual-temporal deep semantic segmentation and change detection model is retrained to update and iteratively optimize model parameters. The remote sensing network is then scheduled to perform differentiated inspections using differentiated inspection frequency configurations in the network deployment parameter set to verify the model optimization effect after adding the core driving variables.

[0012] In a second aspect, a terminal includes a processor and a storage medium; characterized in that: the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the intelligent monitoring and maintenance optimization method for cultivated land based on a low-altitude unmanned aerial vehicle remote sensing network as described in the first aspect of the present invention.

[0013] Thirdly, a computer-readable storage medium is characterized in that it is used to store computer-readable instructions, which, when read by a computer, enable the execution of the intelligent monitoring and maintenance optimization method for cultivated land based on a low-altitude unmanned aerial vehicle remote sensing network as described in the first aspect of this invention.

[0014] The beneficial effects of this invention are as follows: By integrating low-altitude UAV remote sensing networks, constructing a unified four-dimensional spatiotemporal cube data structure, adopting deep semantic segmentation and change detection models, and establishing an intelligent closed-loop supervision mechanism of early warning-dispatch-verification-acceptance-optimization, this invention forms a complete closed-loop optimization system for intelligent supervision and operation of cultivated land based on low-altitude UAV remote sensing networks. It achieves a comprehensive technological leap from passive response to proactive prediction, from offline handling to online closed-loop, from experience-based interpretation to intelligent identification, from single data source to multi-source fusion, and from result supervision to process traceability. This effectively solves the technical problems of high-frequency, high-precision, and full-process closed-loop supervision of cultivated land use changes. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a method for intelligent monitoring, operation and maintenance optimization of cultivated land based on a low-altitude unmanned aerial vehicle remote sensing network, provided in an embodiment of the present invention. Detailed Implementation

[0016] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0017] like Figure 1 As shown, at least one embodiment of the present invention discloses a method for intelligent monitoring and operation optimization of cultivated land based on a low-altitude unmanned aerial vehicle (UAV) remote sensing network, comprising the following steps: Step 1: Obtain multiple network deployment parameters to deploy a remote sensing network covering the target cultivated land area; conduct patrols and emergency monitoring of the target cultivated land area through the remote sensing network to obtain multi-source heterogeneous remote sensing data and form a remote sensing dataset; Step 2: Perform spatiotemporal benchmark unification and fusion processing on the remote sensing dataset to construct a four-dimensional spatiotemporal cube data structure; based on the data structure, analyze the farmland use status to generate high-confidence farmland change patches and output the corresponding confidence values; Step 3: Perform fine-grained semantic classification on the changed patches, identify the violation type, and output the classification result with labels and feature vectors; Step 4: Based on the classification results and confidence values, establish a three-level dynamic early warning mechanism and trigger different levels of early warning signals according to preset thresholds; Step 5: Transform the early warning signals and classification results into structured regulatory tasks and distribute them to the responsible entities according to the response time limit; receive on-site verification and rectification feedback, verify the rectification effectiveness through remote sensing re-flight and form multi-dimensional acceptance conclusions, and build a regulatory record chain after archiving; Step 6: Based on the regulatory record chain, conduct farmland risk classification and farmland regional trend analysis, integrate localized factors and identification bias to extract emerging influencing factors, and use the feedback to optimize the change detection model parameters and complete iterative training.

[0018] In this embodiment, the deployment phase of the remote sensing network is first described. The target area is a typical hilly-plain transition zone in the central part of a province, with a total area of ​​approximately 1200 square kilometers. Three types of terrain parameters—slope, aspect, and elevation variation coefficient—were extracted using a digital elevation model. A layer depicting the distribution of transportation infrastructure was overlaid, and spatial coordinate data of historical farmland violations over the past three years were combined. A density-based spatial clustering algorithm was used to identify three high-density violation hotspots. Based on the above spatial data, a hexagonal cellular topology model was used to plan and lay out fixed take-off and landing stations. The effective coverage radius of a single station was set to 8 kilometers, and the distance between adjacent nodes was approximately 13.856 kilometers, ensuring seamless coverage of the monitoring range by ensuring overlapping edges within the coverage area. A total of 27 fixed take-off and landing stations were ultimately deployed, and their geographical coordinates were encoded using the 2000 geodetic coordinate system and stored in the network deployment parameter set. Each fixed take-off and landing station was equipped with a vertical take-off and landing quadcopter-fixed-wing hybrid UAV with a maximum cruising radius of 45 kilometers, an endurance of 98 minutes, and a cruising speed of 18 to 25 meters per second, meeting the task relay requirements between adjacent stations. The UAV carries three types of payloads: the first is a visible light camera with a ground sampling distance better than 2 centimeters at a flight altitude of 120 meters; the second is a five-band multispectral imager, including blue, green, red, red-edge, and near-infrared bands, used to calculate quantitative indicators of vegetation health such as the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EDI); the third is a 16-line lidar with a three-dimensional point cloud density of more than 30 points per square meter at a flight altitude of 100 meters. The specific models, spatial resolutions, spectral band center wavelengths, and installation deflection angles of these payloads are all included in the network deployment parameter set, serving as the benchmark for subsequent task scheduling and data fusion. Each fixed take-off and landing site ground facility includes a UAV airfield cabin with a wind resistance rating of level 8. The cabin is a cylindrical aluminum alloy structure, 3.2 meters in diameter and 2.8 meters high, with an automatically opening and closing hatch on the top. The cabin integrates a meteorological sensing module, a battery quick-change mechanism, and a dual-mode communication unit for fourth-generation and fifth-generation mobile communication. The communication unit establishes a persistent connection with the cloud-based monitoring platform deployed on the government cloud platform via a message queue telemetry transmission protocol, uploading a heartbeat data packet every 30 seconds. The data packet includes information on remaining battery power, payload online status, door opening / closing status, and real-time dynamic differential positioning accuracy. The topology table for all 27 unmanned aerial vehicle (UAV) sites is stored in JavaScript object notation, with fields including a unique node identifier, latitude and longitude coordinates, coverage radius, operating status, and last heartbeat timestamp, and is synchronized to the task scheduling engine database of the cloud-based monitoring platform. During the remote sensing data acquisition phase, all 27 UAVs are connected to a unified scheduling platform. This platform has a built-in task allocation algorithm that dynamically generates daily patrol paths based on the device status information and topology coordinate information of each node.In periodic cruise mode, the platform automatically triggers a full-network coverage mission at 9:00 AM daily, with each drone executing a serpentine reciprocating flight path along a preset hexagonal boundary. In on-demand trigger mode, administrators manually define regions of interest via a web interface, and the platform automatically matches the nearest available node to that region based on Euclidean straight-line distance to execute a single flight mission. Emergency response mode is triggered in conjunction with the early warning system. When a violation patch is identified as a Level 3 warning by the intelligent recognition algorithm, the platform immediately dispatches drones from the drone airport corresponding to the hexagonal coverage area containing the patch and its two adjacent nodes to complete a re-flight verification and photography within a 2-hour time limit. If a node's remaining battery power falls below 20% or the message queue telemetry transmission protocol heartbeat data packet is interrupted for more than 300 seconds, the scheduling engine migrates the pending mission to the nearest operational node with the closest Euclidean distance and recalculates the flight path to avoid airspace conflicts between different drone routes. Simultaneously, the scheduling platform retrieves the latest quarterly satellite imagery from the remote sensing image service platform via an application programming interface, with a spatial resolution of 1.8 meters, covering the target area. Once the warning patch is generated, the platform analyzes its 1984 geodetic coordinates and converts them to the 2000 geodetic coordinate system. Then, it directs the corresponding UAV to perform refined data collection in key areas: visible light cameras capture images in orthographic projection mode, with a forward overlap rate set to 80% and a lateral overlap rate set to 70% to ensure stereoscopic mapping accuracy; multispectral imagers simultaneously acquire data from five bands to generate a normalized vegetation index (NZD) raster map; and lidar scans at a frequency of 10 times per second to generate laser scan format point cloud files. In addition, soil moisture sensors, miniature weather stations, and high-definition video monitoring PTZ cameras are deployed within a 500-meter radius of each fixed take-off and landing station. These ground-based sensing devices aggregate data to the station's edge computing gateway via a long-range, low-power wireless network. The gateway timestamps all data with Coordinated Universal Time (UTC) and aligns it with the start and end times of the day's aerial photography missions, controlling the time error within ±2 seconds. Simultaneously, the platform retrieved administrative division boundaries, land use status maps from the Third National Land Survey, and vector data of rural homestead ownership from the Department of Natural Resources' geographic information system server. All vector data underwent topological relationship checks and were spatially registered with remote sensing imagery. Ultimately, satellite imagery, UAV aerial photography data, ground sensor data, and vector data were integrated into a multi-source heterogeneous remote sensing dataset. Each record includes a unique data identifier, data source type, Coordinated Universal Time (UTC) timestamp, acquisition node identifier, well-known text format of geometric objects, and file storage path fields.

[0019] In the data preprocessing stage, a four-dimensional spatiotemporal cube was constructed. First, all image data were transformed to the 2000 geodetic coordinate system and the 1985 elevation datum using a seven-parameter Bursa-Robbins transformation model. Then, radiometric calibration was performed on the satellite imagery, converting the raw digital quantization values ​​into the apparent reflectance physical quantity. For UAV imagery, a 6S atmospheric correction model was used, inputting the daily aerosol optical thickness, and the atmospheric scattering effect was eliminated through the radiative transfer equation to obtain the true surface reflectance. Next, a scale-invariant feature transform algorithm was used to extract stable feature points from the images, and multi-source image registration was performed using a regional network adjustment method, unifying images taken at different times and from different platforms into a regular geographic grid with a grid cell size of 0.1 meters. During the registration process, if the root mean square error of corresponding points between a pair of images exceeded 0.3 meters, the image was marked as low quality and removed from the main data stream, but its metadata information was retained to record the actual registration error value for quality traceability. For time-series data from soil moisture sensors and meteorological station thermometers, cubic spline interpolation is used for time alignment to ensure that the temporal resolution of the sensor data is synchronized with the image acquisition time. The final constructed four-dimensional spatiotemporal cube is stored in the network general data format version 4, with four dimensions: spatial dimension X, spatial dimension Y, spatial dimension Z, and temporal dimension T. Each spatiotemporal cell stores three types of data: registered multispectral image blocks, interpolated sensor values, and metadata dictionary. In the change detection stage, a dual-temporal input encoder-decoder neural network architecture is constructed. The encoder part adopts a 34-layer residual network backbone architecture, extracting abstract features at different scales through progressive downsampling. The decoder part adopts a U-shaped network structure, fusing the feature maps of the current layer of the decoder with the corresponding layer of the encoder through skip connections to preserve spatial detail information. The entire network contains two parallel encoding branches, receiving farmland sub-cube data from the current time phase T and the previous time phase, respectively, with a time interval of 7 days and a data size of 256 pixels by 256 pixels by 5 bands. At the output of each coding branch, a convolutional block attention module is introduced. This module applies attention weights simultaneously in both the channel and spatial dimensions, strengthening important features and suppressing irrelevant features. The channel attention weights are dynamically adjusted based on the image registration error; the larger the registration error, the smaller the weight, thus suppressing feature contributions from low-quality images. The spatial attention mask uses the Canny edge detection operator to pre-generate binary maps of patch edges, guiding spatial attention to focus on changing boundary regions. The feature maps output from the two parallel coding branches are subjected to element-wise differencing and concatenated along the channel dimension to form a fused feature map. This fused feature map is input to the decoder for progressive upsampling to generate an initial change probability map. Subsequently, morphological closing operations are performed on this probability map to fill small holes inside the patches; then, connected component analysis is used to identify all connected changing regions, removing noise patches with an area smaller than 200 square meters.The remaining polygons are calculated by averaging the confidence scores of all pixels within them, which is used as the overall confidence score of the polygon. The final output is in the format of a geographic JavaScript object representation, which includes three fields: a well-known text format representation of polygon geometry, a confidence score, and the area of ​​the polygon in square meters.

[0020] In the semantic classification and early warning stage, fine-grained classification is performed on the changed patches. First, for each changed patch, multidimensional features within its interior and a 200-meter buffer zone are extracted from the four-dimensional spatiotemporal cube data structure. The multispectral index features include the following four categories: The first is the Normalized Difference Vegetation Index (NDVI), calculated using the characteristic that vegetation has high reflectivity in the near-infrared band and low reflectivity in the red band. The index value ranges from -1 to +1, with a larger positive value indicating better vegetation growth. The second is the Enhanced Vegetation Index (EVI), which improves sensitivity to high-vegetation-cover areas by introducing the blue band to correct for atmospheric scattering. The third is the Soil-Modified Vegetation Index (SMVEI), which corrects for the influence of soil background on the vegetation index by adjusting the parameter L (set to 0.5), making the index more accurate in low-vegetation-cover areas. The fourth is the Building Index, which identifies hardened surfaces by utilizing the high reflectivity of buildings in the short-wave infrared band. These four indices are calculated using five bands of data from a multispectral imager and are used to quantitatively identify land cover change characteristics. Texture features are calculated using the gray-level co-occurrence matrix (GLCM) method: continuous gray values ​​of near-infrared images are quantized into 32 discrete gray levels. A GLCM is constructed in four directions: 0°, 45°, 90°, and 135°. Based on this matrix, four texture indices—contrast, correlation, energy, and homogeneity—are calculated. These indices reflect the degree of gray-level difference between adjacent pixels, the degree of linear correlation, the uniformity of texture, and the smoothness of texture, respectively. The arithmetic mean of the texture index values ​​in the four directions is used as the final texture feature to identify changes in land surface structure. Geometric features are calculated based on the coordinate data of the vector boundaries of the patches and include four indices: aspect ratio reflects the elongation of the patch; compactness reflects the regularity of the patch shape, with a value closer to 1 indicating a more regular shape; principal direction reflects the main extension direction of the patch, with an angle range of 0° to 180°; and concavity / convexity reflects the degree of depression at the patch boundary, used to identify changes in land parcel morphology. Contextual semantic features are obtained through statistical analysis of the surrounding environment information of the map patch. The specific calculation method is as follows: First, a buffer zone with a radius of 200 meters is established with the map patch as the center. Then, a land use classification layer is superimposed within the buffer zone. The areas of four types of land features, namely roads, buildings, water bodies, and forest land, are counted respectively. The percentage of the area of ​​each type of land feature to the total area of ​​the buffer zone is calculated as the area proportion feature. At the same time, all road vectors within the buffer zone are searched, and the Euclidean plane straight-line distance (in meters) from the centroid of the map patch to the nearest road line is calculated. This distance value reflects the traffic accessibility of the map patch. These contextual features are used to identify the correlation between changes in the surrounding environment (such as newly built roads, new buildings, etc.) and changes in cultivated land.

[0021] The aforementioned multispectral index features (including 4 dimensions: Normalized Difference Vegetation Index, Enhanced Vegetation Index, Soil-Regulated Vegetation Index, and Building Index), texture features (including 16 dimensions, namely, contrast, correlation, energy, and homogeneity, calculated as 4 x 4 equals 16 values ​​in the four directions of 0 degrees, 45 degrees, 90 degrees, and 135 degrees respectively), geometric features (including 4 dimensions: aspect ratio, compactness, main direction, and concavity / convexity), and contextual semantic features (including 6 dimensions: road area ratio, building area ratio, water area ratio, forest area ratio, other land cover area ratio, and distance to the nearest road), totaling 30 dimensions, are concatenated with the map patch confidence value (1 dimension) output from the aforementioned change detection stage to form a 31-dimensional feature vector. The 31-dimensional feature vector is input into an ensemble learning classifier for violation type discrimination. This classifier employs a fusion architecture of two machine learning models: the first is a gradient boosting decision tree model (containing 120 decision trees, each with a maximum depth of 8 layers, and a learning rate of 0.05; the model iteratively improves prediction accuracy by fitting the residuals of the previous model); the second is a multilayer perceptron neural network model (containing 3 hidden layers with 128, 64, and 32 neurons respectively, using the modified linear unit function ReLU, where the output value equals the maximum of the input value and zero). The prediction results from the two models are fused using a weighted voting method (the weight ratio is 3:1, i.e., the vote of the gradient boosting decision tree). The voting weight is 3, and the voting weight of the multilayer perceptron is 1. The final output is five types of violation labels: the first is seasonal abandonment (farmland is left uncultivated for a long time, resulting in vegetation degradation); the second is facility agriculture exceeding the boundary (agricultural facilities such as greenhouses occupy farmland beyond the planned scope); the third is temporary soil piling (temporarily piling soil or building materials on farmland); the fourth is illegal construction (illegally building houses or other permanent buildings on farmland); and the fifth is damage caused by natural disasters (farmland is damaged by natural disasters such as floods and landslides). At the same time, the 32-dimensional neuron activation values ​​of the last hidden layer of the multilayer perceptron neural network (the third hidden layer contains 32 neurons) are retained as the depth feature vector of the patch for use in the subsequent cluster analysis in the risk classification stage.

[0022] The specific implementation of the three-level dynamic early warning mechanism is as follows: Based on the violation type label and the area value of the land parcel in the classification results, three-level early warning judgment conditions are set: The first-level early warning threshold condition is that all three judgment conditions must be met simultaneously: the area of ​​a single identified land parcel is less than 500 square meters, the building index is less than 0.1, and the surface hardening rate (i.e., the percentage of hardened surface pixels in the land parcel divided by the total number of pixels in the land parcel) is less than 10%. Land parcels meeting these conditions are judged as suspected seasonal abandonment behavior, and the early warning risk level is low. The second-level early warning threshold condition is that one of the following three judgment conditions must be met: Condition 1 is that the area of ​​the land parcel is between 500 square meters and 2000 square meters; Condition 2 is that the building index is greater than or equal to 0.1 but the surface hardening rate is less than 10%. Within the range of 10% to 30%, condition three is that temporary structures (with a maximum height of less than 2 meters and a base area of ​​less than 50 square meters) are detected through lidar point cloud data. If any of these conditions is met, the plot is judged as an encroachment on the boundaries of facility agriculture, with a medium warning risk level. The third warning threshold condition is that one of the following three judgment conditions must be met: condition one is that the area of ​​the plot is greater than 2,000 square meters, condition two is that the surface hardening rate exceeds 30%, and condition three is that permanent buildings are detected through lidar point cloud data (three-dimensional structures with a height of greater than 2 meters and a base area of ​​greater than 50 square meters are identified through point cloud elevation data). If any of these conditions is met, the plot is judged as an illegal occupation of cultivated land for construction, with the highest warning risk level. When a farmland violation patch in the classification results meets any one of the above-mentioned Level 1, Level 2, or Level 3 early warning threshold conditions, the system automatically generates a structured early warning message of the corresponding level (low, medium, or high) based on the violation type label and the patch area. This message uses a data dictionary structure and includes the following fields: a unique identifier for the patch (generated using the Universal Unique Identifier 4 format, a 128-bit random identifier), the center point's latitude and longitude coordinates (expressed using the Coordinate System 1984, with latitude and longitude values ​​accurate to 6 decimal places), the patch area value (in square meters), and a confidence value (range 0.0). Up to 1.0, retaining 2 decimal places), warning level indicators (three levels: low, medium, or high), violation type labels (one of five categories: seasonal abandonment, facility agriculture beyond boundaries, temporary soil piling, illegal construction, or damage from natural disasters), three-phase image thumbnails (using the Joint Image Experts Group compression format, with an image size of 256 pixels by 256 pixels, and the three thumbnails corresponding to three different time points: phase T minus 14 days, phase T minus 7 days, and the current time of phase T), and multispectral index curves (using portable network graphics format, the curves show the daily changing trends of the normalized vegetation index and building index over the past 30 days).The structured early warning message is transmitted via a dedicated line on the Ministry of Natural Resources' internal network. It utilizes Hypertext Transfer Protocol Security (HTTP) (encrypting data using Transport Layer Security 1.3 to ensure data integrity during transmission) and an Open License Protocol 2.0 token authentication mechanism (issuing time-limited access tokens for authentication and access control). The message is then pushed to the corresponding enforcement terminal device (running Android OS version 10 with a mobile enforcement application version greater than or equal to 3.2.0) by calling the application programming interface of the Institute of Natural Resources' mobile enforcement terminal. Simultaneously, the early warning message is written to a relational database deployed on the government cloud platform (using PostgreSQL database management system version 13.5, with PostGIS spatial database extension version 3.1 installed for storing and querying geospatial data). The database record contains the following fields: Early Warning Identifier (primary key to ensure record uniqueness), Polygon Geometry (using a well-known text format to store polygon coordinates), Early Warning Level (low, medium, or high level), Confidence Level (floating-point value between 0 and 1), Violation Type (one of five violation types), Area (square meters), and Coordinated Universal Time (UTC) timestamp. The data consists of several fields: (i.e., Greenwich Mean Time stamp accurate to the second), terminal reception confirmation time field (stores the timestamp of the mobile terminal receiving and confirming the warning message), administrative division code field (stores the administrative division code), assigned enforcement officer employee number field (stores the unique employee number of the enforcement officer), on-site check-in GPS latitude and longitude fields (store the latitude and longitude coordinates of the enforcement officer when arriving at the site), response deadline field (stores the latest timestamp at which the enforcement officer must respond), and first upload time field (stores the timestamp of the first rectification photo uploaded by the enforcement officer). These fields constitute a complete data record of the warning and handling process, serving as the basic data source for subsequent evaluation of farmland supervision and maintenance efficiency. During the task assignment and closed-loop handling phases, the warning message is transformed into a structured regulatory task. The specific process is as follows: First, the system, based on the geographical location of the illegally cultivated land parcels (represented by the latitude and longitude coordinates of the center point of the parcel in the 1984 World Geodetic Coordinate System), combines the administrative boundary vector data (stored in shape file format, containing five levels of administrative boundary polygons) and ownership information vector data (including parcel number, right holder's name, land use classification, and boundary coordinate range) from the multi-source heterogeneous remote sensing dataset. Then, through spatial overlay analysis (using spatial inclusion relation predicates to determine whether a point is inside a polygon, i.e., whether the center point of the parcel is located within the boundary range of a certain administrative unit polygon), it matches the code of the natural resources authority to which the parcel belongs, thereby determining the corresponding responsible entity.If a land parcel crosses the boundaries of multiple administrative units (i.e., the polygon of the land parcel intersects with the polygons of multiple administrative units), the intersection area of ​​the land parcel within each administrative unit is calculated. The administrative unit with the largest percentage of intersection area is designated as the primary responsible unit, and the remaining intersecting administrative units are recorded as co-responsible units in the co-assignment field. Secondly, based on the violation type label in the classification results and the area size of the land parcel with violations, the urgency level and processing time limit for the land supervision task are determined. The specific judgment rules are as follows: If the violation type label is illegal construction or facility agriculture exceeding the boundary and the area of ​​the plot is greater than or equal to 2000 square meters, the task will be marked as Level 1 Emergency (level code L1), requiring a response within 24 hours (i.e., law enforcement personnel must arrive at the scene and upload a GPS check-in record via mobile terminal within 24 hours from the task dispatch time); if the violation type label is temporary soil piling or other types of violations with an area between 500 and 2000 square meters, the task will be marked as Level 2 Emergency (level code L2), requiring a response within 48 hours; if the violation type label is seasonal abandonment or minor violations with an area less than 500 square meters, the task will be marked as Level 3 Emergency (level code L3), requiring a response within 72 hours. The processing time limit starts from the task dispatch time (this time is stored in the task dispatch time field of the relational database). The system periodically checks whether the task exceeds the response time limit through a scheduled task mechanism (using the scheduled task expression: asterisk slash 30 spaces asterisk space asterisk space asterisk space asterisk space asterisk, indicating execution every 30 minutes).Next, the system automatically generates an electronic task sheet. This task sheet uses a JavaScript object representation architecture (a standard format for defining data structure specifications) and includes the following data fields: Task Unique Identifier: stores a 128-bit task identifier generated using the Universal Unique Identifier 4 (UUID) format; Associated Warning Message Identifier: stores the unique identifier of the corresponding warning message to establish the association; Patch Geometry: stores the coordinates of all vertices of the patch polygon using a well-known text format; Center Point Latitude and Longitude Coordinates: stores two floating-point numbers in array format, with the longitude value first and the latitude value last; Patch Area: stores the area value in square meters; Violation Type Label: stores a text label of one of the five violation types; Confidence Score: stores a floating-point value between 0 and 1; Warning Level: stores a low, medium, or high warning level; Emergency Level: stores an L1, L2, or L3 level code; Response Deadline: stores the date and time in ISO 8601 format (format: year-month-day T hour:minute:second Z, e.g., 2024-01-15T09:00:00Z). The rectification completion deadline field is automatically set to a date within the range of 7 to 30 days after the response, depending on the type of violation (7 days for minor violations, 30 days for serious violations). The main responsibility code field stores a 12-digit administrative division code, a 9-digit administrative division code, or a 6-digit administrative division code. The collaborating responsible unit code field stores the administrative division codes of multiple collaborating units in an array format. The three-phase image thumbnail unified resource locator array field stores the access URLs of three image files. The normalized vegetation index curve graph unified resource locator field stores the access URL of the curve graph file. The legal basis field stores the text of relevant laws and regulations. The handling suggestion field stores multiple handling suggestions in a text array format (e.g., the first suggestion is to conduct an on-site inspection and take photos of the four boundaries, the second suggestion is to question the parties involved and make a written record of the questioning, the third suggestion is to order the cessation of illegal activities, and the fourth suggestion is to demolish illegal structures within a time limit). The required evidence type field stores multiple evidence type requirements in a text array format (e.g., the first item is on-site photos, the second item is the written record of the questioning, and the third item is a scanned copy of the order to rectify notice).

[0023] Electronic task orders are pushed to mobile enforcement terminals at the corresponding levels via a government intranet messaging middleware system. This middleware is built on the Apache Kafka distributed stream processing platform. The message subject is named "Farmland Supervision Task," and the number of data partitions is set to 12 (distributing messages across 12 partitions enables parallel processing and improves throughput). The replication factor is set to 3 (each message is stored as a copy on 3 different nodes to ensure data reliability). The push strategy is differentiated based on the administrative level of the responsible entity: for village-level responsible entities, the task order is pushed to the village-level grid member, who receives the task through the government WeChat mini-program; for a specific responsible entity, the task order is pushed to the enforcement personnel of the natural resources bureau, who receive the task through the mobile enforcement application client; for any responsible entity, the task order is pushed to the supervisors of the natural resources bureau, who receive the task through the web-based management backend. The system records the task dispatch time field (the timestamp of the task order being sent from the server to the terminal) and the reception status field (this field uses integer encoding: 0 indicates not delivered, meaning the message has not yet arrived at the terminal device; 1 indicates delivered, meaning the message has been delivered to the terminal but the user has not viewed it; 2 indicates read, meaning the user has opened and viewed the message; 3 indicates accepted, meaning the user clicked the accept task button to confirm execution) for each terminal. After law enforcement personnel click the accept task button, the system automatically reads the current location coordinates and current timestamp of their mobile device's GPS and writes the three field values ​​of GPS latitude coordinates, longitude coordinates, and task acceptance time into the database as the audit basis for the task flow process and the evidence for tracing responsibility. If no confirmation of processing for the electronic task order is received within the response time limit (i.e., before the response deadline) (i.e., the task acceptance time field value is still empty, indicating that no one has accepted the task), the system automatically triggers an escalation alarm mechanism: The first level escalation involves sending an SMS reminder to the person in charge of the task. The SMS is sent through the mobile cloud messaging application service SMS gateway, and the message content adopts a preset template format: square brackets farmland supervision square brackets closed task number curly brackets task_id curly brackets closed timeout no response comma please urge law enforcement personnel to handle it in a timely manner, where task_id within the curly brackets is automatically replaced with the actual task number; if the timeout period reaches 1.5 times the response time limit (e.g., if there is still no response after 36 hours if the response time limit is 24 hours), and no response confirmation is received, the alarm message is escalated to the second level. At the same time, a serious timeout alarm SMS is sent to the person in charge of the relevant department, and the timeout task is marked with a flashing red icon on the large screen visualization interface of the farmland protection supervision dashboard (the icon flashes between red and dark red every 1 second to attract attention). Through the multi-level alarm mechanism, timeout issues are ensured to be detected and handled in a timely manner.

[0024] During the rectification and acceptance phase, law enforcement officers uploaded rectification feedback information via mobile enforcement terminals, including the following four categories of data: The first category is on-site photos (using the Joint Image Experts Group compression format, with an image resolution greater than or equal to 1920 pixels by 1080 pixels, i.e., full HD resolution; the exchangeable image file format information of the photo file must include a shooting timestamp and GPS coordinates to ensure the authenticity of the photo); the second category is on-site video (using the Dynamic Image Experts Group Part 4 format, with video encoding using the H.264 advanced video coding standard compression algorithm, video length greater than or equal to 30 seconds, and video content must include a sweeping view of the four boundaries of the illegal plot to comprehensively record the on-site situation); the third category is written descriptions (the text description must be no less than 100 Chinese characters, and the description must include an overview of the violation, a record of the statements made by the parties involved, a description of the rectification measures taken, and the expected completion time of rectification); and the fourth category is GPS positioning information (the positioning accuracy error of longitude and latitude coordinates must be less than or equal to 10 meters to ensure that law enforcement officers have indeed arrived at the scene). These four categories of data constitute complete on-site verification feedback information. By comparing the GPS coordinates in the on-site verification feedback with the original farmland patch locations in the electronic task sheet, the geodetic distance between the two points is calculated based on the principles of spherical trigonometry. If the calculated spatial location deviation exceeds 50 meters, the system marks the feedback as invalid and sends a location deviation alert message to law enforcement personnel via mobile terminal; if the deviation is within 50 meters, it is confirmed as valid verification feedback, and the system automatically extracts the rectification status tag from the text description of the feedback and writes it into the rectification status field for subsequent processes. For farmland supervision tasks marked as completed in the valid verification feedback, the system calculates the Euclidean distance (the straight-line distance between two points in a plane coordinate system) between the centroid of the plot and each unmanned aerial vehicle (UAV) node based on the plot location information in the electronic task sheet (reading the well-known text format polygon coordinates stored in the plot geometry field of the task sheet data structure) and the topology layout spatial coordinate information in the network deployment parameter set (storing the latitude and longitude coordinates of 27 UAV nodes). It then iterates through all nodes and selects the node that is closest to the plot and has normal equipment status (remaining power greater than or equal to 20%, online communication link, and the door is in an openable state) as the re-flight task execution node. After the rectification task is issued, at the preset review time nodes (the review time is planned to be three times: the first review is arranged to be carried out 48 hours after the rectification is completed and reported, the second review is arranged to be carried out 7 days after the first review, and the third review is arranged to be carried out 14 days after the second review, and multiple reviews are carried out to ensure the continuous and stable rectification effect), the system automatically generates a re-flight mission instruction data packet and pushes it to the mission scheduling platform through the message queue. The scheduling platform controls the selected low-altitude UAV remote sensing network node to carry out multiple review flights to collect the latest remote sensing data of the rectified plot.During each re-inspection flight, the UAV executes its flight mission according to a pre-planned flight path. This path covers a 100-meter buffer zone extending beyond the map patch, with a forward overlap of 80% and a lateral overlap of 70%. Following this path, the UAV acquires the latest multispectral imagery and lidar point cloud data of the remediated area. After orthorectification, the multispectral imagery is extracted to calculate the latest Normalized Difference Vegetation Index (NDI) and the latest Bare Soil Index. The system reads the baseline values ​​before remediation from the four-dimensional spatiotemporal cube data structure. If the vegetation restoration conditions are met for a preset number of consecutive re-inspections (preset number 2), specifically, if the NDI increase exceeds a first preset increase threshold of 0.25 and the bare soil index decreases exceeds a first preset decrease threshold of 0.15, the system determines it as effective re-cultivation and sets the re-cultivation verification pass field value to 1. Subsequently, the system performs pixel-level precise comparison and analysis between the latest multispectral imagery acquired during the re-inspection flight and the original change imagery stored in the four-dimensional spatiotemporal cube data structure. The two images were rigorously registered to ensure complete spatial correspondence, with a registration error of less than 0.1 meters. Three types of difference indices were calculated pixel by pixel: normalized difference vegetation index pixel difference, building index pixel difference, and Euclidean distance in red-green-blue color space. This distance value represents the degree of color change of that pixel.

[0025] Based on the aforementioned pixel-level difference data, three multi-dimensional evaluation indicators were calculated: the first indicator is the vegetation cover restoration rate, which is calculated by comparing the number of pixels with restored vegetation with the number of pixels with lost vegetation, reflecting the completeness of vegetation restoration; the second indicator is the integrity of building demolition, which is calculated by comparing the number of pixels with demolished buildings with the number of detected buildings, reflecting the thoroughness of illegal building demolition; the third indicator is the surface disturbance index, which is calculated by the average Euclidean distance of all pixels in the color space. This index represents the average degree of color change of all pixels before and after rectification, and the smaller the value, the less surface disturbance and the higher the rectification quality. In addition, based on the lidar point cloud data obtained from the review flight, the system calls the oblique photogrammetry 3D reconstruction model to calculate the surface elevation standard deviation of the rectified area. All point cloud data within the map area are spatially divided into 10m x 10m grid units, and the elevation standard deviation of all points within each grid unit is calculated. The arithmetic mean of the elevation standard deviations of all grids is taken as the overall surface elevation standard deviation, which reflects the flatness of the terrain undulation of the entire map area. If the overall surface elevation standard deviation is less than the preset threshold of 0.15 meters, the system determines that the terrain has been restored to a flat state, sets the field value of the terrain flattening completion to 1, and incorporates the surface elevation standard deviation value into the multi-dimensional evaluation index system for the determination of rectification qualification. Based on the above four multi-dimensional evaluation indicators, the system automatically determines whether the rectification is qualified: if all four of the following conditions are met simultaneously, the rectification is deemed qualified: Condition 1 is that the vegetation coverage restoration rate is greater than the first preset restoration rate threshold (this threshold is set to 70%, meaning that at least 70% of the damaged vegetation has been restored); Condition 2 is that the integrity of the demolished buildings is greater than the second preset restoration rate threshold (this threshold is set to 95%, meaning that at least 95% of the illegal buildings have been demolished); Condition 3 is that the surface disturbance index is less than the preset percentage threshold of the initial surface disturbance index (this percentage threshold is set to 20%, meaning that the surface disturbance index is less than 0.2 times the initial surface disturbance index, indicating that the change in surface color after rectification does not exceed 20% of the initial change); Condition 4 is that the standard deviation of surface elevation meets the standard (less than 0.15 meters). When all four conditions are met, the farmland rectification is deemed qualified (the value of the qualified rectification field is set to 1).This generates an electronic acceptance form containing detailed data on the aforementioned multi-dimensional evaluation indicators. The form is formatted in portable document style and includes the following six parts: Part 1: Basic task information (including unique task number, geographic coordinates of the map patch, violation type classification label, and area in square meters); Part 2: Timeline of the rectification process (including task assignment timestamp, law enforcement personnel response timestamp, rectification completion reporting timestamp, and number of re-inspection flights); Part 3: Detailed table of multi-dimensional evaluation indicators (listing the actual values ​​and compliance status of four indicators: vegetation cover restoration rate, building demolition integrity, surface disturbance index, and surface elevation standard deviation); Part 4: Comparison of re-flight images (including comparison of normalized vegetation index false-color composite images before and after rectification, building index heat map comparison, and surface elevation shading 3D visualization comparison); Part 5: Evaluation conclusion text (qualified or unqualified); Part 6: Digital signature area (administrators sign using an electronic seal, generated based on the commercial cryptographic algorithm SM2 elliptic curve digital signature algorithm to ensure security).

[0026] As the final rectification and acceptance conclusion document, the electronic acceptance form, after being digitally signed by the person in charge of the Natural Resources Institute on a mobile device (the signature data structure includes four elements: the signer's name, the signer's employee number, the signing timestamp, and the digital certificate serial number), is submitted to the blockchain evidence storage node through the application programming interface of the Hyperledger Consortium Blockchain Network's software development kit (SDK). (This blockchain network deploys four consensus nodes using the Raft consensus algorithm to ensure data consistency, and two endorsement nodes using an endorsement strategy that allows endorsement from either the service provider point of Organization 1 or the service provider point of Organization 2. The strategy expression is OR (single quotes: Org1MSP.peer, single quotes, comma, single quotes; Org2MSP.peer, single quotes, closed brackets). After verifying the signature, the blockchain node generates an immutable transaction hash value (using the SHA256 hash algorithm to generate a 256-bit unique identifier) ​​and returns the transaction identifier field value, completing the on-chain evidence storage process for the acceptance form and ensuring that the data is immutable and traceable. Otherwise, if any one of the four evaluation indicators fails to meet the standard, the system determines that the farmland rectification is unqualified (setting the value of the qualified rectification field to 0). The system analyzes each of the specific items that fail to meet the standards in the multi-dimensional evaluation indicators: if the vegetation coverage restoration rate is less than 70%, it means that the vegetation has not been fully restored to a healthy state; if the building demolition integrity is less than 95%, it means that the buildings have not been completely demolished; if the surface disturbance index is greater than or equal to 0.2 times the initial surface disturbance index, it means that the surface disturbance is still too large; if the standard deviation of surface elevation is greater than or equal to 0.15 meters, it means that the terrain has not been restored to flatness. The system automatically generates a secondary farmland rectification instruction document based on these unqualified items. This secondary rectification instruction adds four new fields to the original electronic task sheet data structure: Rework Round Number (marked as 2 to indicate the second round of rectification; if it fails again, it increments to 3, 4, etc.), Failure Indicator Field (an array of names of all non-compliant indicators, such as vegetation cover restoration rate, building demolition integrity, surface disturbance index, and surface elevation standard deviation), Rework Requirement Field (providing specific rectification requirements for each non-compliant item; for example, if building demolition is incomplete, the requirement is to further clean up the remaining building foundations and deeply till and level the surface to restore arable conditions), and New Rectification Completion Deadline Field (set to a date within 7 to 15 days depending on the complexity of the problem and the number of non-compliant items; 7 days for simple problems and 15 days for complex ones). The system will redistribute the generated secondary rectification instruction to the enforcement terminal of the original responsible entity through a message middleware, initiating the secondary rectification process and repeating the above verification and acceptance process to ensure that the rectification work is truly implemented and meets the compliance standards.

[0027] During the digital ledger archiving and regulatory record chain construction phase, the system archives all data from the entire process of early warning generation to rectification and acceptance, along with the final rectification and acceptance conclusions, into the farmland supervision digital ledger database. Specifically, the digital ledger uses a distributed time-series database (named TimescaleDB, a database system specifically designed for time-series data, developed based on PostgreSQL relational database extensions). The main table, named "Farmland Supervision Ledger," contains the following 26 fields: Record Identifier (using an auto-incrementing integer type as the primary key to ensure uniqueness for each record); Early Warning Message Identifier (storing the unique identifier of the associated early warning message); Task Identifier (storing the unique identifier of the associated task); Patch Geometry (storing polygon coordinates in a well-known text format); Patch Area (storing square meter values); Violation Type (storing one of five violation type labels); Early Warning Level (storing low, medium, and high level labels); Confidence Score (storing a floating-point value between 0 and 1); Early Warning Generation Time (storing the timestamp of early warning triggering); Task Assignment Time (storing the timestamp of task assignment); and Law Enforcement Personnel Response Time (storing the timestamp of law enforcement personnel response). The system stores the following time stamps: Task Acceptance Time, Rectification Completion Time (storing the time stamp of the time stamp of the time stamp of the time stamp of the first re-flight verification), Initial Re-inspection Time (storing the time stamp of the last acceptance), Final Acceptance Time (storing the time stamp of the last acceptance), Rectification Status (storing one of four statuses: in progress, completed, unqualified, or overdue), Vegetation Restoration Rate (storing a percentage value), Building Demolition Integrity (storing a percentage value), Surface Disturbance Index (storing a floating-point value), Surface Elevation Standard Deviation (storing a floating-point value in meters), Acceptance Conclusion (storing one of two conclusions: qualified or unqualified), Responsibility (storing the village-level administrative division code), Responsibility (storing a specific administrative division code), Law Enforcement Personnel Number (storing the unique identifier of the law enforcement personnel), Blockchain Transaction Hash (storing the transaction identifier returned after being uploaded to the blockchain), Record Creation Time (storing the time stamp of the first time the record was written to the database), and Record Update Time (storing the time stamp of the last time the record was modified). The digital ledger implements an automatic partitioning strategy based on time (creating an independent partition table for each calendar month, for example, data for January 2024 is stored in partition table 202401), and establishes two types of indexes to accelerate queries: spatial indexes (using the GIST general search tree index method provided by the PostGIS geographic information system extension to create spatial indexes on the geometric fields of map features to support fast spatial queries) and time indexes (using the BRIN block range index method to create indexes on the time fields to improve query efficiency by utilizing the sequential nature of time data).The regulatory record chain is built on the Hyperledger consortium blockchain network, which is deployed in a government cloud computing environment. The network topology includes the following components: 4 organizations (the first organization is the Provincial Department of Natural Resources, and the second to fourth organizations are the natural resources-related departments of the three pilot cities), 8 peer nodes (each organization deploys 2 peer nodes to store the ledger and execute smart contracts), 4 sorting nodes (using the Raft consensus algorithm for transaction sorting and block generation, with a leader election timeout set to 2 seconds and a heartbeat data packet sending interval set to 500 milliseconds, or half a second), 2 certificate authority nodes (responsible for issuing and managing digital certificates for network members), and 1 CouchDB document-oriented state database cluster (containing 3 nodes for storing the current state data of the blockchain). Each farmland monitoring task, from early warning generation to acceptance, triggers chaincode (i.e., smart contract code deployed on the blockchain network to automatically execute business logic) calls at the following four key nodes, writing the hash values ​​of key status data to the blockchain: The first key node is early warning generation, where a chaincode function named "Create Early Warning" is called. Input parameters include the early warning message identifier, a well-known text format string representing the geometries, a violation type label, and a confidence score. Internally, the chaincode function concatenates these four fields, calculates a SHA256 hash value, and writes it to the blockchain ledger. Upon successful execution, the function returns a transaction identifier denoted as "tx underscore alert". The second key node is task assignment, where a chaincode function named "Assign Task" is called. Input parameters include the task identifier, the associated early warning message identifier, the assigned responsible entity code, and the assignment timestamp. The first key node is when rectification is completed. It calls a chaincode function named "Submit Rectification," which takes three parameters: a task identifier, a rectification completion timestamp, and an array of Uniform Resource Locators (URLs) for evidence files (the hash values ​​of the rectification photos and videos uploaded to the InterPlanetary File System's distributed storage network). The second key node is when acceptance is successful. It calls a chaincode function named "Complete Acceptance," which takes six parameters: a task identifier, an acceptance conclusion (pass or fail), a percentage of vegetation restoration rate, a percentage of building demolition integrity, a standard deviation of ground elevation in meters, and digital signature data from law enforcement personnel (signatures generated by law enforcement personnel using their private keys to sign the acceptance data). The third key node is when rectification is completed. The third key node is when acceptance is successful. The fourth key node is when acceptance is successful. It calls a chaincode function named "Complete Acceptance," which takes six parameters: a task identifier, an acceptance conclusion (pass or fail), a percentage of vegetation restoration rate, a percentage of building demolition integrity, a standard deviation of ground elevation in meters, and digital signature data from law enforcement personnel (signatures generated by law enforcement personnel using their private keys to sign the acceptance data). The fifth key node is when acceptance is successful.

[0028] Each block generated by the blockchain network contains the following four core data structures: the hash value of the previous block (a 256-bit hash value calculated using the SHA256 hash algorithm to calculate all data in the previous block, linking all blocks into a chain structure with hash pointers), a timestamp (using a Unix epoch-level millisecond timestamp to record the precise time of block generation), the Merkle tree root hash value of the transaction set (constructing all transactions in the block into a Merkle tree binary structure, with the root node hash value verifying the integrity of any transaction in the tree), and the digital signature of the current block (using the commercial cryptographic algorithm SM2 elliptic curve digital signature algorithm, with the sorting node signing the block data to ensure the credibility of the block's origin). Once a block is generated and confirmed by consensus, it is written into the ledger and cannot be tampered with or deleted, ensuring the traceability (traceability of historical operations), auditability (auditability of operational processes), and backtrackability (backtrackability of historical states) of the regulatory records. It provides a web-based visual interface for a blockchain explorer (built on the Hyperledger Explorer open-source project, offering user-friendly query and display functions). Regulatory personnel can query the complete on-chain evidence record of a task by entering a task identifier or warning message identifier. The query results include the timestamp of each state change operation (displaying the operation time down to the second), the operator's identity information (displaying the operator's name and employee number), the original data hash value (displaying a 256-bit hexadecimal hash string), and the block height (displaying the position number of the transaction in the blockchain). The blockchain explorer enables full-process auditing and accountability. In addition, the digital ledger database automatically performs a scheduled backup task at 3:00 AM every day (the backup strategy consists of two parts: using the pg_dump tool to perform a full data backup and export a complete database snapshot, and using WAL write-ahead log incremental backup to continuously back up the database change log). The backup files are encrypted (using AES advanced encryption standard 256-bit key CBC cipher block chaining mode encryption to ensure that the backup files cannot be cracked even if leaked) and then stored in the object storage service (this service provides an application programming interface compatible with S3 simple storage service and supports massive data storage). The backup data retains the backup history data of the most recent 90 days (i.e. 3 months) for disaster recovery, ensuring data security and recoverability.

[0029] During the implementation phase of risk grading and differentiated inspection strategies, the system performs risk analysis using basic units. The specific process is as follows: First, the system aggregates and groups data from the farmland supervision digital ledger database according to code fields, extracting complete data records of all historical farmland violation patches within the village. For each unit, the following eight statistical indicators are calculated for risk assessment: The first indicator is the total number of violation patches, which counts the number of all violation patches that occurred in the village within the past 12 months (1 year); the second indicator is the average rectification cycle, in hours, which calculates the rectification cycle for each completed rectification patch in the village and then calculates the average, reflecting the village's rectification response speed; the third indicator is the recurrence frequency, which counts the ratio of the number of recurring patches to the total number of violation patches, reflecting the persistence of violations in the village; the fourth indicator is the trend direction, which uses the Theil-Sen robust slope estimation method to calculate the median slope of the normalized vegetation index time series data of all farmland pixels in the village over the past 12 months. If the trend direction slope is greater than 0.01, it indicates that vegetation cover is improving. The first indicator is the average vegetation restoration rate, calculated by summing the vegetation restoration rates of all accepted and qualified map patches and dividing by the number of accepted and qualified map patches. The second indicator is the average building demolition integrity, calculated by summing the building demolition integrity of all accepted and qualified map patches and dividing by the number of accepted and qualified map patches. The third indicator is the average surface disturbance index, calculated by summing the surface disturbance indices of all accepted and qualified map patches and dividing by the number of accepted and qualified map patches. The fourth indicator is the average surface elevation standard deviation, calculated by summing the surface elevation standard deviations of all accepted and qualified map patches and dividing by the number of accepted and qualified map patches. The aforementioned eight village-level statistical indicators, combined with four multi-dimensional evaluation indicators extracted from the rectification and acceptance conclusions (vegetation restoration rate, building demolition integrity, surface disturbance index, and surface elevation standard deviation), collectively form a village-level cultivated land risk feature dataset containing 12 dimensions. Next, the village-level cultivated land risk feature dataset is input into a K-means clustering algorithm for unsupervised automatic classification. Before clustering calculation, all 12-dimensional features are standardized using Z-scores to eliminate the influence of dimensions. The standardized feature values ​​have a mean of 0 and a standard deviation of 1, facilitating comparison of features with different dimensions. Subsequently, the elbow rule is used to determine the optimal number of clusters: the sum of squared errors within groups corresponding to k values ​​varying from 2 to 10 is calculated, a curve of sum of squared errors minus k is plotted, and the location of a significant inflection point on the curve is observed. The k value corresponding to this inflection point is selected as the optimal number of clusters; in this case, k equals 4 as the optimal number of clusters.

[0030] The K-means algorithm uses the K-means addition initialization method, setting the maximum number of iterations to 300 and the convergence threshold to 0.0001. After the clustering algorithm is executed, four cluster centers and cluster labels for each village sample are obtained. Based on the characteristics of each cluster center in three key dimensions—recurrence frequency, average rectification period, and trend direction—the four clusters are labeled with risk levels: clusters with a cluster identifier equal to 0 are labeled as high-risk farmland areas, characterized by a recurrence frequency greater than 0.3 (i.e., more than 30% of the patches recurred), an average rectification period greater than 120 hours (i.e., rectification took more than 5 days to complete, indicating a slow response), and a trend slope less than -0.01 (i.e., vegetation cover is deteriorating); clusters with a cluster identifier equal to 1 are labeled as medium-risk farmland areas, characterized by a recurrence frequency between 0.1 and 0.3 (i.e., 10% to 30% of the patches recurred), and an average rectification period of 60 hours. The risk level of cultivated land is defined as follows: Rectification is completed within 2.5 to 5 days, with a trend slope between -0.01 and +0.01 (meaning stable vegetation cover). Clusters with a cluster identifier of 2 are classified as low-risk cultivated land areas, characterized by a recurrence frequency of less than or equal to 0.1 (i.e., a recurrence rate below 10%), an average rectification period of less than or equal to 60 hours (i.e., rapid response with rectification completed within 2.5 days), and a trend slope greater than +0.01 (meaning vegetation cover shows an improving trend). Clusters with a cluster identifier of 3 are classified as special cultivated land areas, characterized by a small number of illegal patches but a particularly large area of ​​a single illegal patch, or the village being located within the ecological protection red line area requiring special attention. Each sample is assigned a risk level category label (the risk level field stores one of four labels: high risk, medium risk, low risk, or special risk). This label, along with village code and other information, is written into the village-level risk profile dimension table for subsequent differentiated inspection strategies. Next, for farmland supervision sub-areas with different risk levels, the system sets differentiated inspection frequency strategies based on low-altitude UAV remote sensing networks.The specific inspection frequency strategy is as follows: For high-risk farmland areas (for example, a 500-mu plot of land south of a village is identified as a high-risk area through cluster analysis, and the recurrence frequency of this area is 2.3 times per year, indicating a high incidence of violations), a daily inspection frequency is set (this frequency is recorded as the fourth preset frequency, meaning the task scheduling platform automatically generates a drone aerial photography task for this high-risk area at 9:00 AM every day to ensure timely detection of new violations); for medium-risk farmland areas, an inspection frequency of once every 3 days is set (this frequency is recorded as the fifth preset frequency, meaning an inspection task is performed every 3 days); for low-risk farmland areas, an inspection frequency of once every 7 days is set. (This frequency is recorded as the sixth preset frequency, which means that the inspection task is carried out once a week to save monitoring resources); For special cultivated land areas, the inspection time window is dynamically adjusted according to the environmental impact factors in the violation characteristic data (such as the agricultural production season including the spring plowing period, the harvest season, and the high incidence of illegal construction before major holidays). For example, during the spring plowing period (the three months from March to May), the inspection frequency is temporarily increased to once every 2 days (because illegal occupation of cultivated land is prone to occur during the spring plowing period), and during the non-busy farming period (the four months from November to February of the following year), the inspection frequency is reduced to once every 10 days (because there are relatively fewer violations during the non-busy farming period, the monitoring frequency can be appropriately reduced). The differentiated inspection frequency configuration information is stored in the inspection plan field of the network deployment parameter set database in JavaScript object notation format. This field adopts a nested object structure and contains the following subfields: Village Code field stores the village-level administrative division code, such as 123456; Risk Level field stores the risk level identifier of the village, such as high risk; Inspection Frequency Days field stores the number of days between inspections, such as 1 indicating daily inspection; Next Inspection Time field stores the precise timestamp of the next planned inspection, using the International Organization for Standardization 8601 format, such as 2024-01-15T09, colon 00, colon 00Z, indicating 9:00 AM on January 15, 2024; and Assigned Node Identifier field stores the node number of the unmanned airport responsible for performing the inspection task, such as node underscore 03. The task scheduling platform reads the inspection plan configuration data from the network deployment parameter set at 1:00 AM daily. Based on the next inspection time field value, it determines all inspection tasks that need to be executed that day, automatically generates a list of inspection tasks for the day, and distributes it to the corresponding unmanned aerial vehicle (UAV) nodes. After the inspection tasks are completed, the system automatically updates the next inspection time field value to a future timestamp obtained by adding the current time to the number of days of inspection frequency, ensuring that inspection tasks are executed periodically according to plan. Furthermore, a dedicated AI recognition strategy is configured for each risk level category of farmland supervision sub-area to improve the accuracy of targeted detection.

[0031] For high-risk farmland areas, the system uses a temporal deep learning prediction model (this model is based on a convolutional long short-term memory network architecture, which contains two layers of convolutional long short-term memory units, with each layer having 128 hidden state channels). The model input data is the daily normalized vegetation index image sequence of the high-risk area over the past 30 days (the data tensor shape is square brackets 30 comma H comma W comma 1 closed, where 30 represents a time step of 30 days, H represents the number of pixels in image height, W represents the number of pixels in image width, and 1 represents a single channel). The model output data is the predicted normalized vegetation index value and the confidence score of abnormal behavior for the next 7 days. This time-series model is specifically trained on historical data from high-risk areas to capture subtle, gradual changes in cultivated land (such as slow surface hardening and gradual abandonment and degradation of vegetation). The model's hyperparameters include: a learning rate of 5 x 10^-5 (0.00005, using a small learning rate to fine-tune the model parameters), a batch size of 8 (processing 8 samples per training session), and a loss function consisting of mean squared error and anomaly detection loss (where mean squared error measures the deviation between predicted and true values, and anomaly detection loss is denoted as L-index anomaly, which uses a single-class support vector machine method to identify anomalous samples that deviate from the normal pattern). The training data comes from a time-series dataset of historical violation patches in high-risk areas (with a sample size of 5000 or more to ensure sufficient model training).

[0032] For special arable land areas, the system customizes identification strategies based on the violation characteristics of the area (for example, if the main violation type in a village is determined to be seasonal abandonment, data analysis shows that this type of violation is strongly correlated with environmental factors such as soil moisture and precipitation). The system focuses on analyzing the relationship between soil moisture changes and vegetation succession sequences. The specific implementation method is as follows: First, extract the soil moisture content time series data of this special area from the four-dimensional spatiotemporal cube data structure (this data comes from the deployed soil moisture sensors, which measure the volumetric moisture content at three soil depths of 10 cm, 20 cm, and 30 cm, with a data sampling interval of 4 hours), and simultaneously extract the normalized vegetation index time series data of this area (data sampling interval of 1 day); Second, perform Pearson correlation analysis on the soil moisture content time series data and the normalized vegetation index time series data (this method calculates the linear correlation coefficient r between the two variables, with the correlation coefficient r ranging from -1 to +1, and the closer the absolute value is to 1, the stronger the correlation). If the absolute value of the calculated correlation coefficient is greater than 0.6 (indicating a moderate or higher correlation), then the soil moisture data is introduced as an auxiliary input channel into the dual-temporal change detection model during the change detection stage to enhance the model's sensitivity to changes related to soil moisture. Furthermore, the system analyzes the characteristics of vegetation growth cycles by fitting vegetation phenological curves and uses a double Logistic function model to fit the annual change curve of the vegetation index. The function expression of this model is: St = (a1 + b1) / (1 + natural logarithm base e) * (c1) * (t - d1) * (c2) * (t - d2 ...c2) * (t - d2) * (c2) * (c2) * (c2) * (c2) * (c2) * (c2) * (c2) * (c2) * (c2) * (c2) * (c2) * (c2) * (c The system identifies key phenological periods based on fitted curves: the start of the growing season (SOS, the time point corresponding to the maximum growth rate in the rising segment of the curve), the peak growth period (POS, the time point corresponding to the maximum value of the curve and that maximum value), and the end of the growing season (EOS, the time point corresponding to the maximum decline rate in the falling segment of the curve). If the SOS at the start of the growing season is delayed by more than 15 days for two consecutive years (indicating delayed vegetation greening, possibly due to lack of cultivation) or the maximum vegetation index (POS) at the peak period is less than 0.5 (indicating poor vegetation growth, possibly due to abandonment), the system automatically triggers a potential abandonment risk warning message to remind managers to pay close attention.All parameter configuration information for the dedicated AI recognition strategy (including the storage path of the model weight file in the server file system, the number and type of model input feature channels, the numerical settings of various judgment thresholds, and the weight coefficients of each item in the composite loss function) and differentiated inspection frequency configuration data are fed back to the configuration file of the dual-temporal deep semantic segmentation and change detection model (which uses YAML format, a highly readable data serialization format) through a RESTful expressive state transition application programming interface (a web application interface design style based on the Hypertext Transfer Protocol). The change detection model reads this configuration file when the system starts up and loads, and dynamically selects the corresponding recognition strategy branch and model parameters based on the risk level category of the currently processed patch (obtained by performing a spatial query operation on the center point coordinates of the patch to retrieve the risk level identifier of the village to which the location belongs from the village-level risk profile dimension table), thus achieving differentiated and accurate detection for different risk areas. Simultaneously, the differentiated inspection frequency configuration data is updated in real time to the inspection plan table in the network deployment parameter set database. The task scheduling platform automatically generates a daily inspection task list based on the inspection frequency and planned time stored in this table, ensuring that high-risk areas receive denser monitoring coverage (daily inspections promptly detect new violations) and low-risk areas reduce unnecessary flight times (weekly inspections save resources), thereby optimizing drone hardware resource configuration and battery life management to maximize monitoring efficiency. During the localized influencing factor extraction and model iteration training phase, the system further deepens model optimization to improve predictive capabilities. Based on the farmland use trend characteristics of sub-regions corresponding to risk level categories (including trend direction slope, violation type distribution statistics, and rectification effectiveness evaluation results) and the violation characteristic patterns stored in the regulatory record chain (including recurrence frequency and average rectification cycle), in-depth farmland field surveys are conducted for villages identified as high-risk, medium-risk, and special farmland areas.The survey content includes the following three aspects: The first aspect is the land transfer situation: Historical records of land ownership changes for all cultivated land parcels in the area were retrieved from the real estate registration center (records include fields such as transfer time, transfer price, transferee's entity or individual status, existence of mortgages and mortgage amount, etc.). Based on these records, a land transfer activity index was calculated. The calculation formula is derived as follows: First, the number of land ownership transfers in the area during the observation period is counted, the total number of cultivated land parcels in the area is counted, and the observation period span is recorded. Then, the land transfer activity index equals the number of transfers divided by the total number of parcels (in brackets) multiplied by the observation period (in brackets). The larger the index value, the more active the land transfer. The second aspect is farmers' planting intentions: A planting intention questionnaire was distributed to farmers in the area (using stratified random sampling to ensure sample representativeness; the sample size should be greater than or equal to 10% of the total number of farmers in the village and no less than 50 questionnaires; the questionnaire design includes a 5-point Likert scale, with the question being whether you are willing to continue cultivating this parcel for the next 3 years; the answer option is 1, which means complete acceptance). The survey categorized planting intentions into five levels: 1 (unwilling), 2 (somewhat unwilling), 3 (uncertain), 4 (somewhat willing), and 5 (very willing). After collecting the questionnaires, the average planting intention score (calculated by summing all questionnaire scores and dividing by the number of questionnaires) and standard deviation (reflecting the degree of dispersion of intentions) were calculated. The third aspect of the survey focused on surrounding construction activities: approval information for all approved construction projects within a 3-kilometer radius of the area in the past three years was extracted from the Housing and Urban-Rural Development Bureau's construction permit database (information included project type classification, project land area in square meters, project start time, and project expected completion time). Based on this information, the frequency index of surrounding construction activities was calculated. The calculation formula is derived as follows: first, sum the land area of ​​all projects within the range, then divide by the area of ​​the statistical range (the area of ​​the circular range is equal to pi multiplied by the square of the radius R, where R equals 3 kilometers). Finally, the frequency of surrounding construction activities is equal to the sum of the land areas of all projects divided by the sum of the Greek letter pi multiplied by the square of R (closed brackets). The unit is square meters per square kilometer. This index reflects the level of pressure from surrounding construction and development in the area.The aforementioned three aspects of survey data, along with other socioeconomic data, constitute a localized influencing factor set, which includes 15 variables: land transfer activity, average land transfer price (yuan per mu per year), average planting intention score, outflow rate of young and middle-aged labor force (calculated as the reciprocal of the proportion of permanent residents aged 16 to 59 to the registered population; the higher the value, the more severe the outflow), frequency of surrounding construction activities, distance from the county center (kilometers), distance from the nearest highway exit (kilometers), distance from the nearest industrial park (kilometers). The following are included: per capita cultivated land area (unit: mu per person), agricultural mechanization rate (calculated as the percentage of mechanized farming area to total cultivated land area), proportion of high-standard farmland (percentage of high-standard farmland area to total cultivated land area), irrigation facility coverage rate (percentage of cultivated land with irrigation facilities to total cultivated land area), soil fertility grade (a rating of 1 to 5 by the agricultural department, with 1 being the worst and 5 being the best), average grain subsidy amount in the past 3 years (unit: yuan per mu), and participation rate in policy-based agricultural insurance (percentage of insured cultivated land area to total cultivated land area).

[0033] The 15 variables in the localized influencing factor set were progressively screened to retain truly influential factors. The first step was an analysis of variance (ANOVA): the variance of each variable across all surveyed village samples was calculated, and constant factors with variances less than a threshold of 0.05 were removed. This first step resulted in a non-constant influencing factor set containing 12 discriminative variables. The second step involved a Pearson correlation test to remove redundant variables: the Pearson correlation coefficient between any two variables in the non-constant influencing factor set was calculated. If the absolute value of the calculated correlation coefficient was greater than a preset threshold of 0.85, it indicated severe collinearity redundancy between the two variables. In this case, one variable needed to be removed, with the rule being to retain the variable more strongly correlated with the area of ​​violations and remove the variable with a weaker correlation. This second step resulted in a de-redundant influencing factor set containing 9 variables. The third step involves truncating policy-related factors within a time window to avoid interference from long-term trends: Policy-related factors are identified in the redundant influencing factor set. The specific implementation time of the policy event is retrieved, and only observations within a preset time window before and after the policy event are retained for analysis. The time window is set to 60 days, 30 days before and after the policy implementation time. Historical data points outside the time window are set as missing values, and forward imputation is used to fill in the missing values. The purpose of this step is to avoid interference from long-term trend changes before and after policy implementation on the causal inference of policy effects. After the third step, the influencing factor set after time window filtering is obtained. This set still contains 9 variables, but the observations of some variables have been truncated. The fourth step involves feature importance ranking to eliminate weak contributing factors: The influencing factor set after time window filtering is constructed into a feature matrix, and the total area of ​​illegally occupied cultivated land patches is constructed into a target vector. The feature matrix and target vector are input into a feature importance ranking algorithm based on the permutation importance method. The algorithm execution flow is as follows: First, a baseline prediction model is trained using the gradient boosting regression tree algorithm. The model is trained using a feature matrix to predict the target vector. After training, the model performance is evaluated on the test set to obtain the baseline R-squared score. Then, for each of the nine features, the value order of the feature in all samples is randomly shuffled while keeping the other eight features unchanged. The R-squared score of the model on the test set is recalculated using the shuffled feature matrix. Next, the importance score of the feature is calculated. The larger the importance score, the greater the decrease in model performance after shuffling the feature, indicating that the feature contributes more to the model's predictive ability. To obtain a stable importance assessment, steps two and three are repeated 10 times. The arithmetic mean of the importance scores obtained from the 10 calculations is taken as the final stable importance score of the feature. Finally, weak contribution factors with importance scores below a preset threshold of 0.01 are removed.After screening, a final set of high-contribution impact factors was obtained, containing five core variables and their importance scores: land transfer price, outflow rate of young and middle-aged labor force, frequency of surrounding construction activities, distance to the nearest industrial park, and average planting intention. Then, the core driving variables were identified for the causal inference model: the final set of high-contribution impact factors was input into a causal forest model, which was implemented based on a dual machine learning framework. The specific implementation principle of the causal forest model is as follows: for each of the five core variables, two auxiliary prediction models were constructed to remove confounding: Auxiliary model one is a propensity score model, which estimates the probability of a sample being processed given other characteristics, with the processing variable being a binary indicator variable of that variable, estimated using the Logistic regression method; Auxiliary model two is an outcome residual model, which estimates the expected area of ​​violations given other characteristics, with the target variable being the area of ​​illegally cultivated land patches, estimated using the random forest regression method. The residuals of the two auxiliary models mentioned above were orthogonalized to eliminate confounding factors. The average treatment effect was calculated, which represents the average increase in the area of ​​illegally cultivated land patches by one unit increase in the variable. This represents the magnitude of the causal effect of the variable on the area of ​​illegal land. Simultaneously, a 95% confidence interval and p-value were calculated to test whether the causal effect was statistically significant. The five core variables were sorted from largest to smallest by the absolute value of their average treatment effect and identified as the core driving variables with significant causal effects on the changes in cultivated land in each sub-region of the risk level category. These five core variables are: land transfer price, outflow rate of young and middle-aged labor force, frequency of surrounding construction activities, distance to the nearest industrial park, and average planting intention. Finally, the identified core driving variables were added as new input features to the four-dimensional spatiotemporal cube data structure to enhance the model input information. The specific addition method is as follows: For each spatiotemporal cell (i.e., a voxel unit in three-dimensional space) in the four-dimensional spatiotemporal cube, firstly, based on the spatial coordinates of the voxel (reading the X coordinate of the voxel to represent the number of meters to the east and the Y coordinate to represent the number of meters to the north), determine the code to which the spatial location belongs through a spatial connection query operation (i.e., judging the point within the polygon). Then, read the values ​​of the five core driving variables of the village (including land transfer price, outflow rate of young and middle-aged labor force, frequency of surrounding construction activities, distance to the nearest industrial park, and average planting intention) from the village-level socio-economic data table. Finally, write these five scalar values ​​as new bands (i.e., five new channels from band 6 to band 10) into the multidimensional data array of the voxel.The original four-dimensional spatiotemporal cube data structure has the following dimensions: square brackets X comma Y comma Z comma T comma 5 (representing the spatial dimension of the X-axis, the spatial dimension of the Y-axis, the elevation dimension of the Z-axis, the time dimension of the T-axis, and 5 multispectral image bands). After adding core driving variables, it is expanded to square brackets X comma Y comma Z comma T comma 10 (representing the addition of 5 socioeconomic characteristic bands to the original 5 multispectral bands, for a total of 10 bands). Using the four-dimensional spatiotemporal cube data structure with added core driving variables, the dual-temporal deep semantic segmentation and change detection model is retrained (this process is called fine-tuning, i.e., continuing to train and optimize parameters based on the original model). The training dataset contains all accepted violation patch samples from the past 12 months, and the dataset is divided into training and validation sets at an 80% to 20% ratio. The model's neural network architecture remains unchanged (still using a 34-layer residual network encoder plus a U-shaped network decoder), but the number of input channels has been expanded from the original 5 channels to 10 channels (adding 5 core driving variable channels as additional inputs).

[0034] The training hyperparameters were configured as follows: the learning rate was set to 0.0001 (using the Adam adaptive moment estimation optimizer, with its two momentum parameters set to 0.9 and 0.999 respectively), the batch size was set to 16 (processing 16 patch samples simultaneously during each training iteration), the training epochs were set to 20 (i.e., traversing the training set 20 times), and the loss function was a weighted binary cross-entropy function. The loss function formula was derived as follows: summing all pixels, positive sample weight multiplied by the true label multiplied by the logarithm (log), predicted value, negative sample weight multiplied by 1 minus the true label, logarithm (log), predicted value, positive sample weight multiplied by 1 minus the predicted value, positive sample weight multiplied by 1 minus the predicted value, positive sample weight multiplied by 1 minus the predicted value, positive sample weight multiplied by 1 minus the predicted value, positive sample weight multiplied by 1 minus the predicted value. The positive sample weight was set to 3.0 (because the proportion of illegal patch pixels to all pixels is less than 10%, indicating a severe imbalance between positive and negative samples, the positive sample weight needs to be increased), and the negative sample weight was set to 1.0. After each training epoch, four evaluation metrics are calculated on the validation set: F1 score (the harmonic mean of precision and recall), precision (the proportion of actual positive samples among those predicted as positive), recall (the proportion of actual positive samples correctly predicted), and IoU (Intersection over Union) (the area of ​​the intersection of predicted and actual patches divided by the area of ​​their union). The model weights with the highest F1 score are selected and saved. The performance improvement is evaluated after training. The results show that adding five core driving variables significantly improves the model's accuracy in identifying farmland use changes in different risk level categories, especially significantly enhancing its ability to capture and detect subtle, gradual farmland change features in high-risk areas (such as early surface hardening and slow vegetation abandonment and degradation). The updated model metadata is written to the model registry system for subsequent tasks. The model registry uses the tracking server component of the MLflow machine learning lifecycle management platform, which uniformly manages model versions, performance metrics, and deployment status. When generating new change detection tasks, the task scheduling platform automatically reads the latest model (i.e., the best-performing and validated model) marked with the production environment tag in the model registry, ensuring that the system always uses the best-performing model for inference and prediction. Furthermore, the system uses differentiated inspection frequency configuration data centrally stored in the network deployment parameters to schedule low-altitude UAV remote sensing networks to execute differentiated inspection flight tasks, verifying the model optimization effect after adding core driving variables in the field. The specific verification process is as follows: high-risk villages with a high historical recurrence frequency are selected as pilot verification areas. After the model is officially deployed and launched, intensive inspections are carried out on these villages according to the differentiated inspection frequency to obtain new temporal violation patch samples for detection, and the performance indicators of the model detection results are statistically analyzed. The verification results show that the model optimization strategy of adding core driving variables is effective, significantly improving the model's detection accuracy, and significantly reducing both the false alarm rate and the missed detection rate.The verification results were compiled into a formal assessment report and submitted to the regulatory department of the Department of Natural Resources for approval. After approval, the optimized model was promoted and deployed throughout the province, realizing a technological leap from passive response-based regulation (i.e., passively discovering violations after they occur) to proactive predictive regulation (i.e., identifying high-risk areas in advance based on driving factors). Ultimately, a complete closed-loop optimization system for intelligent monitoring and operation of cultivated land based on low-altitude UAV remote sensing network was formed.

[0035] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention is further explained below in conjunction with a specific application scenario.

[0036] When a suspected illegal occupation incident occurs on a plot of farmland measuring 860 square meters on the south side of the target area, the system first acquires UAV multispectral imagery and LiDAR point cloud data of the plot at the current time through periodic patrol missions, and retrieves historical archived data from 7 days prior. The two sets of data are then cropped from the four-dimensional spatiotemporal cube into 256-pixel sub-cubes according to the bounding rectangle of the map features, and then input into two parallel encoding branches of the dual-temporal change detection neural network. Since the registration error of the previous time-phase image in this area is 0.18 meters and the registration error of the current time-phase image is 0.22 meters, both error values ​​are less than the quality threshold of 0.3 meters. Therefore, both images are retained for change detection calculation. The channel attention module calculates attenuation factors based on the registration errors of 0.22 meters and 0.18 meters respectively: the attenuation factor for the current time-phase image is equal to the base of the natural logarithm e^(-2.0) multiplied by 0.22, which equals 0.644, and the attenuation factor for the previous time-phase image is equal to the base of the natural logarithm e^(-2.0) multiplied by 0.18, which equals 0.698. The attenuation factor dynamically suppresses and adjusts the feature channel weights of the corresponding coding branches, thereby reducing the interference of registration position uncertainty on the extraction of high-level abstract semantic features. At the same time, the spatial attention mask is guided by the binary image of the patch edges pre-generated by the Canny edge detection operator, so that the spatial attention of the neural network focuses on the outline region of newly added structures at the boundary of the plot. After element-wise difference operations, the feature maps output from the two encoding branches highlight the abrupt changes in elevation and differences in spectral reflectance response at the boundaries of the land parcels. The decoder generates a change probability map by progressively upsampling based on the difference features. After morphological closure operations (dilation followed by erosion) fill the small holes inside the probability map, a single connected patch is identified through connected component analysis (8-neighborhood connectivity rule). This patch has an area of ​​860 square meters (860 square meters divided by 0.01 square meters equals 86,000 pixels because the grid cell size is 0.1 meters). Since the patch area is greater than the noise removal threshold of 200 square meters, it is retained. The arithmetic mean of the confidence values ​​of all pixels within the patch is calculated to be 0.87, and the output is a high-confidence change patch. Subsequently, this change patch enters the fine-grained semantic classification stage. The efficient network B3 version classification model receives the visible light image patch corresponding to the patch, the time-series change curve of the normalized vegetation index (the curve shows that the normalized vegetation index dropped sharply from 0.62 to 0.15 in the past 7 days, indicating that the vegetation disappeared rapidly), lidar elevation profile data (the profile shows that the maximum elevation difference is 2.3 meters, indicating that there is a significant protrusion on the ground surface), and building index heat map (a building index of 0.38 indicates the presence of obvious building features).The classification model comprehensively judged that the map patch met the conditions for a Level 3 warning: the surface hardening rate was greater than or equal to 30% (calculated by the percentage of pixels with a standard deviation of laser point cloud elevation values ​​less than 0.05 meters and a normalized vegetation index less than 0.2 reaching 42%, indicating large-area surface hardening) and the patch area was greater than or equal to 500 square meters, triggering the highest level of Level 3 warning. The system automatically generated a structured warning message data packet, including a unique identifier in the format of Universal Unique Identifier Version 4, the latitude and longitude coordinates of the patch center point in the 1984 geodetic coordinate system (114.2356 degrees east longitude and 30.5678 degrees north latitude), the patch area of ​​860 square meters, a confidence score of 0.87, a warning level of "high", and attached image thumbnails of three time phases (corresponding to 14 days ago, 7 days ago, and the current time, respectively). The warning message was pushed to the law enforcement terminal of the natural resources office to which the map patch belongs via the government intranet application programming interface. Law enforcement officers arrived at the scene within 1.5 hours (timely response) to conduct verification, and uploaded on-site photos, scanned copies of the interrogation records made by the parties involved, and scanned copies of the order to stop the illegal act via mobile terminal.

[0037] Within the 24-hour rectification deadline after task assignment, the task scheduling engine automatically assigned drones from the original unmanned aerial vehicle (UAV) airport node to perform a re-flight verification task in the area, collecting new temporal remote sensing data after rectification. The change detection model re-analyzed the rectified imagery, identifying a remaining patch area of ​​120 square meters, a reduction of 740 square meters (86.0%) from the original patch area of ​​860 square meters (a reduction greater than the effective rectification threshold of 80%). Furthermore, the normalized vegetation index (NDI) rose to 0.53 after rectification (greater than the healthy vegetation threshold of 0.5), meeting the preliminary criteria for effective rectification. The acceptance module simultaneously calculated multi-dimensional evaluation indicators: vegetation recovery rate equals (0.53 - 0.15 / 0.8 - 0.15) equals 0.585 (58.5%), structure volume change rate equals 92.1%, indicating that 92.1% of buildings have been demolished, and the standard deviation of ground elevation decreased from 0.85 meters before rectification to 0.12 meters after rectification, indicating that the ground surface has basically returned to a flat state. All the aforementioned assessment indicator data were packaged into JavaScript object representation format and written into the Hyperledger Consortium Chain blockchain network, forming an immutable and permanent regulatory record. During subsequent model iteration and optimization, system analysis revealed that the land parcel triggered three Level 3 high-risk warnings within the past six-month time window, with a recurrence frequency of 2.4 times per year. The village to which the land parcel belongs was classified as a high-risk arable land area using the K-means clustering algorithm. Information retrieved from the real estate registration center by the field research team showed that the land ownership of the parcel had recently been transferred, with the transferee being a company located in a nearby industrial park. A survey of farmers' planting intentions indicated that local land transfer prices were higher than the income from grain cultivation, leading to farmers' reluctance to cultivate the land. A search of the Housing and Urban-Rural Development Bureau's construction permit database showed that a new logistics and warehousing project had received construction approval within a 2-kilometer radius of the land parcel. After filtering out constant factors such as the uniform subsidy amount across villages using analysis of variance, a causal forest model analysis was conducted on variables such as land transfer price, the outflow rate of young and middle-aged laborers, and the expansion rate of surrounding industrial parks. The analysis results showed that the average treatment effect of the land transfer price variable was 0.73, with a p-value less than 0.01, reaching statistical significance. This variable is the strongest driving factor affecting violations of cultivated land regulations. This land transfer price factor was encoded as the sixth band feature and injected into a four-dimensional spatiotemporal cube data structure for the next round of model fine-tuning training. This will improve the sensitivity of the change detection neural network to detecting early surface hardening behavior in areas with similar high land transfer prices, thereby achieving a technological leap in the regulatory model from passively responding to violations to proactively predicting violation risks.

[0038] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent monitoring and operation optimization of cultivated land based on low-altitude unmanned aerial vehicle (UAV) remote sensing networks, characterized in that, include: Obtain multiple network deployment parameters to deploy a remote sensing network covering the target cultivated land area; The remote sensing network is used to conduct patrols and emergency monitoring of the target cultivated land area, and multi-source heterogeneous remote sensing data is acquired to form a remote sensing dataset. Spatiotemporal benchmark unification and fusion processing of remote sensing datasets are performed to construct a four-dimensional spatiotemporal cube data structure; Based on the data structure, the farmland use status is analyzed to generate high-confidence farmland change patches and output the corresponding confidence values. Fine-grained semantic classification is performed on the changed patches to identify violation types and output classification results with labels and feature vectors; Based on the classification results and confidence values, a three-level dynamic early warning mechanism is established, which triggers different levels of early warning signals according to preset thresholds; The early warning signals and classification results are transformed into structured regulatory tasks and distributed to the responsible entities according to the response time limit; Receive on-site verification and rectification feedback, verify the rectification effectiveness through remote sensing re-flight and form multi-dimensional acceptance conclusions, and build a regulatory record chain after archiving; Based on the regulatory record chain, we conduct risk classification and regional trend analysis of arable land, integrate localized factors and identification biases to extract emerging influencing factors, and use the feedback to optimize the parameters of the change detection model and complete iterative training.

2. The method for intelligent monitoring and operation optimization of cultivated land based on low-altitude unmanned aerial vehicle (UAV) remote sensing network according to claim 1, characterized in that, include: Based on topography, infrastructure, and historical violation records, the spatial distribution and coverage of several fixed take-off and landing stations are planned to form a remote sensing network node layout. Vertical take-off and landing fixed-wing UAVs are configured for each of several fixed take-off and landing stations. The flight time, cruising speed, and performance indicators of the visible light camera, multispectral imager, and lidar payloads carried by the UAVs are determined according to the layout of the remote sensing network nodes and recorded in the network deployment parameter set. Wind-resistant unmanned aerial vehicles are set up at the spatial distribution locations of several fixed take-off and landing stations. The wind resistance level is obtained according to the terrain and landform characteristics. The communication unit maintains a long connection with the cloud monitoring and transmits equipment status information, including the load performance indicators, back in real time. Each unmanned airport is deployed according to the spatial distribution location in a hexagonal topology, the spacing between adjacent nodes is controlled according to the coverage range, and the topology coordinate information containing the device status information is recorded in the network deployment parameter set.

3. The method for intelligent monitoring and operation optimization of cultivated land based on low-altitude unmanned aerial vehicle (UAV) remote sensing network according to claim 2, characterized in that, include: Several drones are integrated into a unified scheduling task. Based on the device status information and topology coordinate information, three operation modes are supported: on-demand triggering, periodic patrol, and emergency response. When the node's power or status is abnormal, the task is migrated to a neighboring node according to the topology coordinate information. The system acquires high-resolution satellite imagery and, based on the topological coordinate information, directs the UAV to utilize the visible light camera, multispectral imager, and lidar to perform routine patrols and emergency monitoring in key and early warning areas, acquiring multiple aerial data sets to obtain an aerial data set. By deploying soil moisture, meteorological, and video monitoring sensor nodes around the fixed take-off and landing station, real-time sensor data that is spatiotemporally synchronized with the aerial photography dataset is collected, and vector data of administrative division boundaries, land use status, and ownership information corresponding to the coverage area are acquired simultaneously. The satellite imagery, aerial data, sensor data, and vector data are integrated into a multi-source heterogeneous remote sensing dataset, and timestamps, data source identifiers, and topological coordinate information are added as spatiotemporal identifiers.

4. The method for intelligent monitoring and operation optimization of cultivated land based on low-altitude unmanned aerial vehicle (UAV) remote sensing network according to claim 3, characterized in that, The remote sensing dataset undergoes spatiotemporal benchmark unification and fusion processing to construct a four-dimensional spatiotemporal cube data structure, including: The image data of several cultivated land areas in the multi-source heterogeneous remote sensing dataset are uniformly converted to a preset geodetic coordinate system and a preset elevation datum to obtain an image dataset under a unified coordinate system. Radiometric calibration and atmospheric correction are performed on the satellite images and low-altitude UAV aerial images in the image dataset to eliminate the effects of atmospheric scattering and sensor differences, resulting in a radiometrically calibrated and atmospherically corrected image data processing set. A joint spatial solution method based on feature point matching and regional network adjustment is adopted to register farmland area images acquired at different times and on different platforms in the image data processing set to the same geographic grid. The registration error is controlled within a preset registration error threshold. The actual registration error is recorded as a registration accuracy index to obtain a spatially registered image data registration set. Interpolate multiple ground sensor data in the remote sensing dataset according to timestamps to align the interpolated multiple ground sensor data with the time series of the image data registration set, thus obtaining a time-aligned sensor dataset; A four-dimensional spatiotemporal cube data structure is constructed based on the preset geodetic coordinate system and preset elevation benchmark. The three spatial dimensions of the data structure are eastward, northward and elevation, and the time dimension is the acquisition time. Each spatiotemporal cell stores the image data registration set, the sensor dataset and metadata. The metadata includes the registration accuracy index and the data source identifier.

5. The method for intelligent monitoring and operation optimization of cultivated land based on low-altitude unmanned aerial vehicle (UAV) remote sensing network according to claim 1, characterized in that, include: Construct a neural network architecture for encoders and decoders with dual temporal inputs; The encoder employs an improved convolutional neural network backbone network. The decoder employs an upsampling network with a skip connection structure, forming two parallel encoder branches; Extract spatiotemporal cube slices of cultivated land areas from the current time phase and the previous time phase from the data structure, and input them into the two parallel encoder branches of the encoder and decoder neural network architecture of the dual-time phase input for feature extraction. Channel attention and spatial attention mechanisms are introduced into the high-level feature map of the encoder output from the two parallel encoder branches. The attention weights are dynamically adjusted according to the registration accuracy index in the data structure to obtain the feature map after weight adjustment. Based on the feature map, dual-temporal information is fused using two strategies: feature difference and feature splicing, to generate a preliminary probability map of farmland change. Morphological closing operations and connected component analysis are applied to the preliminary change probability map to remove noise patches with an area smaller than a preset area threshold, so as to output the final set of cultivated land change patches, and the confidence value is calculated and recorded for each patch in the set of cultivated land change patches; For each farmland change patch in the set of farmland change patches, based on the spatiotemporal cube slices of farmland areas in the current time phase and the previous time phase, multispectral index features, texture features, geometric features and contextual semantic features in the internal and surrounding buffers are extracted from the data structure to form a feature vector; The multispectral index features are calculated based on the multispectral band data in the spatiotemporal cube slice, including the normalized vegetation index, soil-regulated vegetation index, and building index. The texture features are calculated based on the spatiotemporal cube slices using a gray-level co-occurrence matrix, including contrast, correlation, and homogeneity. The geometric features are calculated based on the spatial morphology of the cultivated land change patches, including the aspect ratio, compactness, main direction, and concavity / convexity of the patches; The contextual semantic features are obtained based on the distribution statistics of other land cover types within the buffer zone surrounding the cultivated land change patch; The feature vector and the confidence value are used as inputs to an ensemble learning classifier composed of a gradient boosting decision tree and a multilayer perceptron. Each farmland change patch in the set of farmland change patches is classified into a preset number of violation type categories, so as to output a classification result containing violation type labels and the feature vector.

6. The method for intelligent monitoring and operation optimization of cultivated land based on low-altitude unmanned aerial vehicle (UAV) remote sensing network according to claim 2, characterized in that, Based on classification results and confidence scores, a three-level dynamic early warning mechanism is established, triggering different levels of early warning signals according to preset thresholds, including: Based on the violation type label in the classification results, the first-level warning threshold is set as suspected farmland abandonment behavior where the single identification area is smaller than the first preset area threshold and there are no signs of hardening. The second-level warning threshold is set as farmland facility agriculture encroachment behavior with an area within the first preset area threshold to the second preset area threshold or with temporary structures. The third-level warning threshold is set for illegal occupation of farmland with an area larger than the second preset area threshold, a surface hardening rate exceeding the preset hardening rate threshold, or involving permanent buildings. When the farmland violation patches in the classification results meet any of the first-level warning threshold, the second-level warning threshold, or the third-level warning threshold, a structured warning message of the corresponding level is generated based on the violation type label and the patch area; the structured warning message includes the patch ID, center point latitude and longitude, area, the confidence value, warning level, and a multi-temporal image thumbnail obtained based on the low-altitude UAV remote sensing network. The structured early warning messages are pushed to the mobile terminal of the Institute of Natural Resources through the government intranet API interface and simultaneously written into the farmland protection supervision database, recording the early warning level, response time, check-in location of the verification personnel, and timestamp of the uploaded rectification photos for each farmland early warning.

7. The method for intelligent monitoring and operation optimization of cultivated land based on low-altitude unmanned aerial vehicle (UAV) remote sensing network according to claim 6, characterized in that, The early warning signals and classification results are transformed into structured regulatory tasks and dispatched to the responsible entities according to the response time limits, including: Based on the geographical location of the illegally cultivated land patches, combined with the administrative boundaries and ownership information in the remote sensing dataset, the relevant natural resources authorities are matched to determine the corresponding responsible entities. Based on the violation type label in the classification results and the area size of the illegal farmland patches, the urgency level and processing time limit of the farmland supervision task are determined; Among them, the default violation type is marked as a first-level emergency label and a response is required within a first preset time threshold; the regular type is marked as a second-level label and a response is required within a second preset time threshold. Generate an electronic task sheet containing the location of the illegally cultivated land patches, image screenshots obtained from a low-altitude UAV remote sensing network, the classification results, the warning signal level, the emergency level, the processing time limit, the legal basis, and processing suggestions; The responsible entity will push the electronic task order to the mobile enforcement terminal of the corresponding level of the natural resources authority, and record the dispatch time and receipt status. If no confirmation of processing for the electronic task order is received within the processing time limit, based on the dispatch time and processing time limit, an alarm will be escalated according to the level of the responsible entity and the superior regulatory department will be notified.

8. The method for intelligent monitoring and operation optimization of cultivated land based on low-altitude unmanned aerial vehicle remote sensing network according to claim 7, characterized in that, Receive on-site verification and rectification feedback information, and confirm the rectification effectiveness through the re-flight verification mechanism of the remote sensing network, so as to output a rectification acceptance conclusion containing multi-dimensional evaluation indicators, including: It receives photos, videos, text descriptions, and GPS location information of farmland rectification sites uploaded by grassroots law enforcement personnel through mobile law enforcement terminals, and forms on-site verification feedback information. Compare the GPS location information in the on-site verification feedback information with the original farmland patch location in the electronic task sheet. If the spatial location deviation exceeds the preset location deviation threshold, it is considered invalid feedback. If the deviation is within the threshold range, it is confirmed as valid verification feedback. For farmland supervision tasks marked as completed in the valid verification feedback, the remote sensing network is scheduled to perform multiple re-inspection flights on the rectified plots based on the location information of the map patches in the electronic task sheet and the spatial coordinate information in the network deployment parameter set at the preset re-inspection time node after the rectification task is issued. During each of the aforementioned review flights, multispectral images of cultivated land in the remediation plots are acquired, the normalized vegetation index and bare soil index are calculated, and compared with the baseline values ​​before remediation recorded in the data structure. If the increase of the normalized vegetation index exceeds the first preset increase threshold and the decrease of the bare soil index exceeds the first preset decrease threshold for a consecutive preset number of times, it is determined to be an effective remediation. The latest multispectral images acquired during the re-examination flight are compared with the original change images in the data structure at the pixel level, and the vegetation cover restoration rate, building demolition integrity and surface disturbance index are calculated as multi-dimensional evaluation indicators. Based on the point cloud data obtained from the re-inspection flight, the oblique photogrammetry model is called to calculate the standard deviation of the surface elevation of the cultivated land in the rectified area. If the standard deviation of the surface elevation is less than the preset standard deviation threshold, the terrain is considered to have been restored to flatness, and the standard deviation of the surface elevation is included in the multi-dimensional evaluation index. If the vegetation cover restoration rate is greater than the first preset restoration rate threshold, the building demolition integrity is greater than the second preset restoration rate threshold, the surface disturbance index is less than the preset proportion threshold of the initial value, and the surface elevation standard deviation meets the standard, then the farmland rectification is deemed qualified, and an electronic acceptance form containing the multi-dimensional evaluation indicators is generated as the rectification acceptance conclusion. After confirmation by the administrator, it is archived to the blockchain evidence storage node; otherwise, a second farmland rectification instruction is generated and reissued based on the non-compliant items in the multi-dimensional evaluation indicators.

9. A method for intelligent monitoring and operation optimization of cultivated land based on low-altitude unmanned aerial vehicle (UAV) remote sensing network according to claim 8, characterized in that, Based on historical plot data and rectification results in the aforementioned regulatory record chain, risk classification of the target cultivated land area is performed using a clustering algorithm, dividing the cultivated land into sub-regions for regulatory purposes. Furthermore, time series analysis is used to identify the cultivated land use trend characteristics of each sub-region, including: Using the target cultivated land area as the basic unit, the rectification cycle, recurrence frequency and trend direction of all historical cultivated land violation patches in the target cultivated land area are aggregated from the cultivated land supervision digital ledger. Combined with the multi-dimensional evaluation indicators in the rectification and acceptance conclusion, a cultivated land risk feature dataset containing violation feature data is formed. The farmland risk feature dataset is input into the K-means clustering algorithm. Based on the rectification cycle, recurrence frequency, trend direction and multi-dimensional evaluation indicators, the target farmland area is divided into a preset number of risk level area categories. Each category corresponds to a different farmland violation feature pattern and is marked as a high-risk farmland area, a medium-risk farmland area and a low-risk farmland area. For different categories of farmland supervision sub-areas within the risk level category, differentiated inspection frequencies are set based on the remote sensing network. Specifically, high-risk farmland areas are inspected at a fourth preset frequency, medium-risk farmland areas at a fifth preset frequency, and low-risk farmland areas at a sixth preset frequency. A dedicated AI identification strategy is configured for each type of farmland supervision sub-region in the risk level category. In the high-risk farmland area, a temporal deep learning model is enabled to capture small and gradual changes in farmland. In the risk farmland area, the focus is on soil moisture and vegetation succession sequence analysis based on the violation feature pattern. The identification strategy parameters corresponding to each risk level category are fed back to the dual-temporal deep semantic segmentation and change detection model. At the same time, the differentiated inspection frequency configuration is updated to the network deployment parameter set.

10. A method for intelligent monitoring and operation optimization of cultivated land based on a low-altitude unmanned aerial vehicle (UAV) remote sensing network according to claim 9, characterized in that, Based on the characteristics of farmland use trends in the sub-regions corresponding to the risk level categories and the violation patterns in the regulatory record chain, field surveys of farmland are conducted in the high-risk, medium-risk, and low-risk farmland areas to collect a set of localized influencing factors, including land ownership status, farmers' planting intentions, and the frequency of surrounding construction activities. A variance analysis is performed on each impact factor in the localized impact factor set. Impact factors whose target variance is less than a preset variance threshold are removed according to a preset proportion threshold to obtain a non-constant impact factor set. The Pearson correlation coefficient is calculated for the remaining factors in the set of non-constant influencing factors. When the absolute value of the correlation coefficient between two factors is greater than the preset correlation coefficient threshold, the influencing factors whose difference in correlation coefficient with the area of ​​illegal cultivated land patches in the regulatory record chain is less than the preset difference are retained, thus obtaining the set of redundant influencing factors. The policy factors in the deredundant impact factor set are truncated by an event window, retaining only the observations within a preset time window before and after the event, thus obtaining the impact factor set after time window filtering. The set of impact factors filtered by the time window is input into the feature importance ranking algorithm to remove weak contribution factors with importance scores below a preset importance threshold, thus obtaining the final set of high contribution impact factors. The set of high-contribution influencing factors is input into the causal inference model to identify the core driving variables that have a causal effect on the changes in cultivated land in each sub-region of the risk level category, and the core driving variables are added as new input features to the data structure. Using the data structure after adding the core driving variables, the dual-temporal deep semantic segmentation and change detection model is retrained to complete model parameter updates and iterative optimization. The remote sensing network is then scheduled to perform differentiated inspections through the differentiated inspection frequency configuration in the network deployment parameter set to verify the model optimization effect after adding the core driving variables.

Images