An artificial intelligence-based sky-ground integrated farmland detection system and method

Abstract

The application relates to the technical field of farmland detection, in particular to a sky-ground integrated farmland detection system and method based on artificial intelligence, which comprises the following steps: acquiring multi-source remote sensing data, performing terrain self-adaptive preprocessing on the multi-source remote sensing data; performing terrain weighted fusion on the preprocessed multi-source remote sensing data to generate a detection layer; extracting farmland boundary features based on the detection layer; discriminating farmland shape types based on the extracted farmland boundary features and in combination with the slope grade; classifying farmland types in combination with the aspect category, the slope grade and crop growth features obtained from the multi-source remote sensing data; and dynamically adjusting unmanned aerial vehicle flight parameters and sensor arrangement density according to the slope grade, the aspect category, the farmland shape type and the farmland type. The application can realize terrain self-adaptive accurate detection of farmland in hilly and mountainous areas, and can improve detection accuracy and efficiency by deeply fusing terrain features.

Description

Technical Field

[0001] This application relates to the technical field of farmland detection, and in particular to an integrated sky-ground farmland detection system and method based on artificial intelligence. Background Technology

[0002] With the deepening development of precision agriculture, integrated air-ground monitoring technology, by combining satellite remote sensing, UAV aerial photography, and ground sensor data, provides crucial support for comprehensive, multi-scale monitoring of farmland. In plains areas, this technology has been successfully applied to crop growth monitoring, pest and disease identification, and yield prediction. However, more than one-third of my country's arable land is located in hilly and mountainous areas, characterized by significant topographic relief, fragmented fields, and substantial microenvironmental heterogeneity. This presents serious challenges to traditional monitoring methods in data fusion, feature recognition, and decision-making applications. In particular, topographic factors such as slope, aspect, and altitude directly influence farmland morphology, crop growth patterns, and management strategies. Existing technologies fail to deeply embed topographic features into the entire monitoring process, making it difficult to achieve high-precision, adaptive monitoring of farmland in hilly areas.

[0003] In existing technologies, integrated air-ground detection methods have made some progress. For example, Chinese patent CN118857243B discloses an integrated air-ground mapping method based on oblique photogrammetry 3D mapping. This method involves developing an oblique photogrammetry mapping scheme, deploying multiple types of UAVs for collaborative operation, integrating positioning data and ground control points to construct a high-precision 3D model, and realizing the extraction of ground features and the calculation of terrain parameters.

[0004] Another Chinese patent, CN115839707B, proposes a method for accurately identifying farmland depressions based on a digital terrain model. This method involves establishing an inverted digital terrain model and combining factors such as elevation, topographic relief, and slope, using the entropy weight method to determine weights to identify the location and outline of farmland depressions. These technologies have shown positive effects in local terrain reconstruction and large-scale feature identification, providing fundamental technical support for farmland detection.

[0005] However, when these existing technologies are applied to farmland monitoring in hilly areas, their limitations are gradually exposed, revealing several core defects that are mismatched with the technological requirements:

[0006] First, at the level of multi-source data fusion, existing solutions fail to fully consider the spatial scale differences and geometric distortions caused by terrain factors. For example, in steep slope areas, satellite imagery exhibits significant geometric distortion and overlay due to tilted viewing angles. While CN118857243B employs super-resolution reconstruction and multi-view matching, it does not adaptively adjust the data processing strategy according to the slope level, resulting in amplified data errors in steep slope areas. Conversely, in gentle slope areas, the terrain is flat and distortion is negligible, yet existing technologies still uniformly employ complex interpolation algorithms, introducing redundant computation and the risk of model overfitting. The root cause of this deficiency lies in the failure to distinguish the inherent mechanisms by which different data sources are affected by terrain: satellite data is dominated by macroscopic geometric distortion, UAV data needs to balance resolution and efficiency, and ground sensor data relies on the applicability of spatial interpolation methods. The uniform processing mode of existing technologies cannot achieve accurate adaptation between terrain and data.

[0007] Secondly, at the level of farmland attribute identification, existing methods lack a mechanism to deeply embed topographic features into the discrimination logic, resulting in insufficient identification accuracy in hilly and heterogeneous environments. For example, CN115839707B focuses on depression identification but does not solve the problem of topographical differentiation of farmland morphology and type. Specifically, in hilly areas, the boundary between small plots of farmland and contiguous farmland is affected by both slope and shape complexity, while existing technologies rely solely on area thresholds for judgment, which easily leads to misjudgments: for example, on steep slopes, contiguous farmland is misjudged as small plots of farmland due to its high shape complexity, thus causing a mismatch in detection strategies; similarly, the identification of organic farmland and conventional farmland is significantly driven by slope aspect. Organic farmland on sunny slopes has less fluctuation in NDVI time series due to sufficient sunlight, while existing technologies use uniform physical thresholds or general deep learning models without considering topographic factors such as slope aspect and slope, resulting in greater NDVI fluctuations in organic farmland on shady slopes due to insufficient sunlight, leading to a higher misjudgment rate.

[0008] Finally, at the level of detection strategy, existing solutions lack a mechanism for dynamically switching between terrain features and farmland attributes, making it difficult to balance detection efficiency and accuracy in complex environments. For example, CN118857243B achieves ground feature extraction and parameter calculation, but fails to adjust the drone flight path spacing, sensor deployment density, or pest and disease thresholds based on slope and aspect. This results in missed detections in small, steep-slope farmland in shaded areas due to excessively large drone flight path spacing; and in sunny-slope organic farmland, the higher temperature and rapid aphid reproduction, coupled with the lack of dynamic tightening of pest and disease thresholds, lead to delays in early outbreak detection.

[0009] In summary, existing technologies do not deeply embed terrain features in multi-source data fusion, farmland attribute identification, and detection strategy scheduling, resulting in insufficient detection accuracy and low efficiency in hilly areas, and failing to meet the real-time and adaptive requirements of precision agriculture. Summary of the Invention

[0010] To achieve terrain-adaptive and accurate detection of farmland in hilly and mountainous areas, and to improve detection accuracy and efficiency through deep integration of terrain features, this application provides an artificial intelligence-based integrated sky-ground farmland detection system and method.

[0011] Firstly, this application provides an artificial intelligence-based integrated sky-ground farmland detection method, which adopts the following technical solution: An artificial intelligence-based integrated sky-ground farmland detection method includes the following steps:

[0012] Obtain digital elevation model data of the area to be detected, extract slope features, aspect features and altitude features based on the digital elevation model data, and classify slope grade, aspect category and altitude level;

[0013] Acquire multi-source remote sensing data, including satellite remote sensing data, UAV remote sensing data, and ground sensor data, and perform terrain-adaptive preprocessing on the multi-source remote sensing data based on the slope grade, slope aspect category, and altitude level;

[0014] The preprocessed multi-source remote sensing data is subjected to terrain-weighted fusion to generate a detection layer with uniform resolution. The terrain-weighted fusion assigns fusion weights to remote sensing data from different sources according to the slope level.

[0015] Based on the detection layer, an edge detection algorithm is used to extract farmland boundary features, wherein the edge detection threshold of the edge detection algorithm is adaptively adjusted according to the slope level;

[0016] Based on the extracted farmland boundary features, the farmland area and shape complexity features are calculated, and the farmland morphology type is determined in combination with the slope level.

[0017] Farmland types are classified by combining the slope aspect category, slope grade, and crop growth characteristics obtained from multi-source remote sensing data;

[0018] Based on the slope grade, slope aspect category, farmland morphology type, and farmland type, the drone flight parameters and sensor deployment density are dynamically adjusted.

[0019] Optionally, the terrain adaptive preprocessing includes:

[0020] For different slope grades, corresponding spatial geometric correction, resolution adjustment and raster data interpolation methods are used for satellite remote sensing data, UAV remote sensing data and ground sensor data, respectively.

[0021] Appropriate illumination compensation methods are adopted for different slope aspect categories;

[0022] Appropriate atmospheric correction methods are adopted for different altitude levels.

[0023] Optionally, the corresponding spatial geometric correction method for the satellite remote sensing data includes: using bicubic interpolation to improve resolution at low slope levels, using super-resolution to handle geometric distortion at medium slope levels, and discarding satellite remote sensing data and using only UAV remote sensing data and ground sensor data at high slope levels.

[0024] The corresponding resolution adjustment methods for the UAV remote sensing data include: maintaining the original high resolution at high slope levels, and using downsampling methods to generate raster layers with uniform resolution at low and medium slope levels.

[0025] The corresponding grid data interpolation methods used for the ground sensor data include: using inverse distance weighted interpolation for high slope grades, and using ordinary Kriging interpolation for low and medium slope grades.

[0026] Optionally, the terrain-weighted fusion method assigns fusion weights to remote sensing data from different sources based on slope grade, including:

[0027] At the low slope level, the fusion weight of UAV remote sensing data is less than that of satellite remote sensing data but greater than that of ground sensor data;

[0028] At the medium slope level, the fusion weight of satellite remote sensing data is less than that of UAV remote sensing data but greater than that of ground sensor data.

[0029] At the high slope level, the fusion weight of UAV remote sensing data is greater than that of ground sensor data, while the fusion weight of satellite remote sensing data is 0.

[0030] Optionally, the step of extracting farmland boundary features using an edge detection algorithm includes:

[0031] An edge detection threshold is set based on the slope level, and the edge detection threshold is positively correlated with the slope level.

[0032] Under the high slope level, terrain shadow areas are further identified based on the slope aspect category;

[0033] Within the identified terrain shadow area, the edge detection threshold is increased.

[0034] Optionally, the farmland morphology type is determined jointly based on the farmland area, the shape complexity, and the slope grade, specifically including:

[0035] If any of the following conditions are met, the land is classified as a small plot of farmland:

[0036] The area of ​​the farmland is less than the first area threshold;

[0037] The farmland area is between a first area threshold and a second area threshold, and the slope grade is not lower than the low slope grade, while the shape complexity is higher than the first complexity threshold.

[0038] The farmland area is greater than the second area threshold, the slope grade is a high slope grade, and the shape complexity is higher than the second complexity threshold.

[0039] If any of the above conditions are not met, it is determined to be contiguous farmland;

[0040] Wherein, the second area threshold is greater than the first area threshold, and the first complexity threshold is higher than the second complexity threshold.

[0041] Optionally, the dynamic adjustment of UAV flight parameters and sensor deployment density includes:

[0042] Based on the farmland morphology and slope grade, different drone flight path spacing, drone diagonal cross flight angles, and ground sensor spacing are set.

[0043] When the farmland morphology is contiguous farmland, the spacing between the UAV flight paths and the spacing between the ground sensors decrease as the slope grade increases, while the angle of the UAV's diagonal cross flight increases as the slope grade increases.

[0044] When the farmland morphology is small plots of farmland, a smaller drone flight path spacing and ground sensor spacing, as well as a larger drone oblique cross flight angle, are used compared to contiguous farmland of the same slope grade.

[0045] Optionally, the following steps may also be included:

[0046] The detection error and the false positive rate are obtained. The detection error is obtained by comparing the detected value with the actual measured value. The false positive rate is obtained by statistically analyzing the proportion of abnormal false positive samples.

[0047] Based on the detection error and false judgment rate, the flight parameters of the UAV and the sensor deployment density are iteratively optimized.

[0048] When the detection error exceeds the corresponding error threshold under a specific combination of slope grade and farmland morphology type, or when the misjudgment rate exceeds the corresponding misjudgment rate threshold, at least one of the following should be adjusted: the UAV flight path spacing, the UAV oblique cross flight angle, and the ground sensor spacing.

[0049] Optionally, the classification of farmland types includes:

[0050] Construct terrain-related features based on the slope aspect category and the slope grade;

[0051] Based on the detection layer, extract the time series variation features of vegetation index;

[0052] Soil microbial diversity characteristics were obtained based on ground sensor data.

[0053] Characteristics of pesticide residues were obtained based on chemical analysis;

[0054] The terrain correlation features are fused with the vegetation index time series variation features, soil microbial diversity features, and pesticide residue features to generate an organic farming confidence score.

[0055] Based on the slope aspect category, a corresponding confidence threshold is selected. When the confidence level of organic farming reaches the selected confidence threshold, it is determined to be organic farmland; otherwise, it is determined to be conventional farmland.

[0056] Secondly, this application provides an artificial intelligence-based integrated air-ground farmland detection system, which adopts the following technical solution: An artificial intelligence-based integrated air-ground farmland detection system, comprising:

[0057] The terrain preprocessing module is used to acquire digital elevation model data of the area to be detected, and extract slope features, aspect features and altitude features based on the digital elevation model data, and classify slope grade, aspect category and altitude level.

[0058] The data fusion module is used to perform terrain-adaptive preprocessing on multi-source remote sensing data based on the slope grade, aspect category and altitude level, and to perform terrain-weighted fusion to generate a detection layer with uniform resolution.

[0059] The farmland identification module is used to extract farmland boundary features based on the detection layer, calculate farmland area and shape complexity features, determine farmland morphology type in combination with the slope level, and classify farmland type in combination with slope aspect category, slope level and crop growth characteristics.

[0060] The dynamic switching module is used to dynamically adjust the UAV flight parameters and sensor deployment density according to the slope level, slope aspect category, farmland morphology type and farmland type.

[0061] The feedback optimization module is used to obtain the detection error and the false judgment rate, and to iteratively optimize the UAV flight parameters and sensor deployment density based on the detection error and the false judgment rate.

[0062] In summary, this application includes the following beneficial technical effects:

[0063] 1. To address the core issue of existing multi-source data fusion technologies failing to adequately consider topographic factors that lead to spatial scale differences and geometric distortions, this application implements differentiated preprocessing for satellite remote sensing data, UAV remote sensing data, and ground sensor data according to slope grade. Then, it assigns fusion weights to data from different sources based on slope grade and performs linear weighted fusion, effectively reducing multi-source data fusion errors, achieving accurate adaptation of multi-source data to terrain, and avoiding error amplification and redundant calculations caused by a uniform processing mode.

[0064] 2. To address the issue of insufficient accuracy in existing farmland attribute identification technologies due to the lack of deep embedding of terrain features, this application adopts a three-factor joint rule of area, slope, and shape complexity in farmland morphology discrimination. In farmland type classification, it integrates terrain correlation features, vegetation index time series variation features, soil microbial diversity features, and pesticide residue features to generate organic farming confidence scores. This improves the accuracy of pest and disease detection in small plots of farmland on steep slopes, reduces the misjudgment rate of aphids in organic farmland on sunny slopes, and significantly improves the accuracy of farmland attribute identification in hilly areas.

[0065] 3. To address the problem that fixed detection strategies in existing technologies make it difficult to balance efficiency and accuracy in complex environments, this application dynamically adjusts detection parameters based on slope grade, farmland morphology, farmland type, and slope aspect. Stricter pest and disease thresholds are set for sunny organic farmland, reducing the missed detection rate of small plots of farmland on steep slopes, shortening the detection time per 10,000 mu, and achieving a balance between detection accuracy and efficiency.

[0066] 4. To address the problem that existing technologies with fixed detection parameters cannot adapt to dynamic changes in terrain, this application constructs a closed-loop optimization mechanism by acquiring detection errors and misjudgment rates. The terrain data weight matrix and strategy parameter table are automatically updated monthly, enabling the detection system to dynamically adapt to changes in field morphology caused by seasonal rain erosion or adjustments in farmland management methods, thereby maintaining high accuracy and reducing soil moisture detection errors in small plots of farmland on steep slopes.

[0067] 5. To address the issue of high computational costs caused by the use of complex deep learning models in existing technologies, this application replaces the complex Transformer model with a logistic regression model in farmland type classification. By weighting the normalized terrain association features, vegetation index features, soil microbial features, and pesticide residue features, the organic farming confidence score is output. While ensuring a classification accuracy of no less than 94%, the model complexity is reduced by 60%, effectively adapting to the real-time computational needs of large-scale hilly farmland detection and avoiding the overfitting risk and high hardware dependence caused by complex models. Attached Figure Description

[0068] Figure 1 This is a flowchart of the overall process for farmland testing methods;

[0069] Figure 2This is a flowchart of terrain preprocessing;

[0070] Figure 3 This is a flowchart of multi-source data fusion;

[0071] Figure 4 This is a flowchart for farmland identification;

[0072] Figure 5 This is a flowchart of dynamic switching;

[0073] Figure 6 This is the overall flowchart of the farmland monitoring system. Detailed Implementation

[0074] The following is in conjunction with the appendix Figure 1-6 This application will be described in further detail.

[0075] This application discloses an integrated sky-ground farmland detection method based on artificial intelligence. For example... Figure 1 As shown, an artificial intelligence-based integrated sky-ground farmland detection method includes the following steps:

[0076] S1. Preprocessing of hilly terrain features

[0077] like Figure 2 As shown, this step requires extracting and quantifying the terrain features of the area to be detected, providing targeted input for subsequent multi-source data processing, farmland attribute identification and other steps, avoiding the problem of insufficient adaptability caused by existing technologies ignoring terrain differences, and laying the foundation for the terrain adaptive characteristics of the entire detection process through accurate terrain information collection and quantification.

[0078] S11, Terrain Data Acquisition

[0079] Digital elevation models (DEMs) of the hilly areas to be inspected were acquired using two methods: SRTM 30m DEM data and UAV oblique photogrammetry. SRTM 30m DEM data covers most regions globally and is inexpensive to acquire, meeting the basic elevation data needs for large-scale hilly areas. For UAV oblique photogrammetry, a multi-rotor UAV was used. This UAV, equipped with an RGB camera and an IMU (Inertial Measurement Unit), is highly maneuverable and offers high shooting accuracy, enabling close-range imaging of areas with complex terrain to obtain more precise elevation data.

[0080] Depending on the actual conditions of the detection area, such as the size of the detection range, the complexity of the terrain, and the accuracy requirements, a single acquisition method or a combination of two methods can be flexibly selected to ensure that the acquired digital elevation model can comprehensively and accurately reflect the terrain undulations of the area to be detected. This acquisition method balances the breadth and accuracy of terrain data acquisition, covering a large detection area while accurately capturing the details of complex local terrain, providing reliable basic data for subsequent terrain feature quantification.

[0081] S12, Quantification of Core Terrain Features

[0082] The acquired digital elevation model data is subjected to terrain analysis to extract three core terrain features: slope, aspect, and altitude, which are then quantified and classified.

[0083] S121. Slope Measurement and Classification

[0084] The cosine method is used to calculate slope. This method works by calculating the angle between the normal vector of any point on the terrain surface and the vertical line based on the elevation difference between each grid cell and its adjacent grid cells in the digital elevation model. The slope value is then derived from the cosine of this angle. The specific derivation process is as follows: Let the angle between the normal vector of a point on the terrain surface and the vertical line be... The slope at this point is According to geometric relationships, we know The cosine method calculates... Determine the value The size of the slope is then used to determine the gradient. The calculation formula is as follows:

[0085]

[0086] in The elevation value of the target raster cell. The elevation values ​​of adjacent grid cells. , The planar coordinates of the target grid cell. , These are the planar coordinates of adjacent grid cells.

[0087] Based on the actual topographic distribution of hilly areas in my country, slope is divided into three levels. Gentle slopes have a slope of 5° or less. 5° is chosen as the boundary between gentle slopes and the next level because the terrain below this slope is relatively flat, and the geometric distortion produced by satellite remote sensing data is negligible, meeting the accuracy requirements of conventional data processing. Sloping slopes have a slope greater than 5° but less than 25°. The terrain in this slope range has some undulation, and observable geometric distortion appears in the satellite remote sensing data, but it still has correction value. Steep slopes have a slope of 25° or greater. 25° is chosen as the boundary between steep and sloping slopes because the terrain above this slope is highly undulating, the geometric distortion of the satellite remote sensing data exceeds 30%, and the deviation between pixel positions and actual geographical locations is huge, rendering it worthless for correction. Subsequent data processing should prioritize using UAV remote sensing data. This classification method can accurately match the differentiated processing needs of subsequent multi-source data, providing a clear basis for selecting appropriate data processing methods for different slope areas.

[0088] S122, Slope Vectorization and Classification

[0089] The azimuth method is used to calculate slope aspect. This method starts from true north and calculates the orientation angle of the terrain surface by rotating clockwise, with a calculation range of 0° to 360°. In the specific calculation process, the tilt direction of the terrain surface of each grid cell is determined by the elevation value of each grid cell in the digital elevation model. Then, using true north as a reference, the angle between the tilt direction and true north is measured clockwise, which is the slope aspect value of the grid cell.

[0090] Based on the corresponding sunlight conditions, slope aspects are divided into two categories. Sunny slopes have an aspect angle greater than or equal to 0° and less than or equal to 180°. These areas receive prolonged sunlight, resulting in ample light and significantly impacting crop growth and the time-series fluctuations of vegetation indices. Shady slopes have an aspect angle greater than 180° and less than or equal to 360°. These areas receive shorter periods of sunlight, resulting in insufficient sunlight and relatively higher humidity, making crops more susceptible to fungal diseases. This classification method provides a clear basis for subsequent farmland type classification and monitoring strategy adjustments. For example, the monitoring thresholds for organic farmland on sunny and shady slopes can be differentiated based on differences in sunlight conditions.

[0091] S123. Altitude Quantification and Classification

[0092] The elevation data of the area to be detected is directly extracted based on the elevation value of each raster cell in the digital elevation model. The elevation data extraction process uses a raster data reading algorithm, which can accurately obtain the actual elevation information of each location.

[0093] Based on the elevation distribution characteristics of hilly areas in my country, the elevation is divided into three levels: low hills with an elevation of 500m or less, medium hills with an elevation of 500m or more but less than or equal to 1000m, and high hills with an elevation of 1000m or more.

[0094] The boundary between low and medium hills was chosen at 500m, and between medium and high hills at 1000m, because the elevations of most hilly areas in my country are concentrated in this range. The atmospheric environment varies significantly across different elevation ranges; for example, meteorological factors such as atmospheric pressure, temperature, and humidity exhibit regular differences with altitude. These differences affect the imaging quality of satellite and UAV remote sensing data. This classification method allows for the differentiated selection of subsequent atmospheric correction methods. For instance, high-altitude areas require targeted atmospheric correction methods to eliminate the impact of the high-altitude atmospheric environment on remote sensing data.

[0095] S13, Output terrain feature layer

[0096] The quantified topographic information, including slope grade, aspect category, and elevation level, is integrated to generate a topographic feature layer. This layer uses GeoTIFF format, supports geographic coordinate systems, and can be accurately overlaid with subsequent multi-source remote sensing data. The layer resolution is set to 1m. This resolution is chosen because it ensures the fineness of the topographic features, clearly presenting small-scale topographic undulations, while also matching the unified resolution of the subsequent multi-source data processing, avoiding data fusion errors caused by resolution differences. This step achieves the visualization and quantitative representation of topographic features, providing accurate topographic decision-making basis for subsequent steps such as adaptive topographic preprocessing of multi-source data and farmland attribute identification, ensuring that subsequent steps can adopt targeted processing strategies based on different topographic conditions.

[0097] S2, Terrain Adaptation and Fusion of Multi-Source Data

[0098] like Figure 3 As shown, this step, based on the topographic features such as slope grade, aspect category, and altitude level output by S1, performs differentiated preprocessing and weighted fusion on multi-source remote sensing data. This solves the problem of spatial scale differences and geometric distortion caused by insufficient consideration of topographic factors in existing multi-source data fusion technologies, and generates a high-quality unified resolution detection layer, providing accurate data support for subsequent farmland attribute identification.

[0099] S21. Multi-source data terrain preprocessing

[0100] Multi-source remote sensing data includes satellite remote sensing data, UAV remote sensing data, and ground sensor data. The three types of data have different characteristics and are affected by terrain. Based on the terrain features determined by S1, targeted preprocessing methods are adopted for each type of data. At the same time, aspect-related illumination compensation and altitude-related atmospheric correction are completed to ensure that the data is adapted to the terrain characteristics.

[0101] S211, Satellite Remote Sensing Data Preprocessing

[0102] Sentinel-2 satellite data was selected, as it boasts advantages such as 10m multispectral resolution and a 5-day revisit period, meeting the timeliness and basic accuracy requirements for farmland monitoring. The preprocessing core focuses on spatial geometric correction based on slope grade, while atmospheric correction is performed in conjunction with altitude grade.

[0103] For different slope grades, corresponding spatial geometric correction methods are adopted. For low-slope grades with flat terrain, the geometric distortion of satellite data is negligible; the main challenge is insufficient resolution. Bicubic interpolation is used to improve the resolution from 10m to 0.5m. Bicubic interpolation uses weighted calculations based on the grayscale values ​​of the 16 neighboring pixels around the target pixel, effectively preserving image details and avoiding the overfitting risk associated with complex models such as Generative Adversarial Networks (GANs) for super-resolution, resulting in higher computational efficiency. 10m is chosen as the original resolution because the native resolution of Sentinel-2 satellite multispectral data is 10m, and 0.5m is the minimum accuracy threshold required for farmland boundary extraction. This resolution satisfies subsequent detection needs while controlling data storage and computational costs.

[0104] The terrain with a medium slope has some undulations, and the satellite data shows observable geometric distortion. The local linear transformation super-resolution method is used to process it. This method only stretches and corrects pixels in the slope direction, specifically correcting the deformation caused by the terrain, while improving the resolution to 0.5m. It avoids the redundant calculations of the global super-resolution method and improves the processing efficiency.

[0105] The terrain with a high slope is extremely undulating, and the satellite data distortion exceeds 30%. The pixel positions deviate greatly from the actual geographical locations. Even after correction, the accuracy requirements cannot be met. Therefore, the satellite data is discarded and replaced with UAV remote sensing data.

[0106] For different altitude levels, corresponding atmospheric correction methods are adopted. At low altitude levels, the atmospheric thickness is thinner, and the influence of atmospheric scattering and absorption is smaller. The dark target method is used for atmospheric correction. By identifying the surface reflectance of dark target pixels (such as shadows and clear water bodies) in the image, the atmospheric optical thickness is inverted and the atmospheric influence is removed to adapt to the atmospheric characteristics of low altitude areas.

[0107] The atmospheric environment at mid-altitude is of moderate complexity, with enhanced atmospheric scattering and aerosol influence. The 6S model is used for atmospheric correction. The 6S model can comprehensively consider multiple factors such as atmospheric molecular scattering, aerosol scattering, and ozone absorption. By inputting parameters such as altitude and aerosol type, it can simulate the atmospheric transport process. The correction accuracy is higher than that of the dark target method and is suitable for atmospheric conditions in mid-altitude regions.

[0108] At high altitudes, the atmosphere is thin and the air pressure is low, resulting in less aerosol content but significant ozone absorption. The MODTRAN model is used for atmospheric correction. This model can accurately simulate the radiative transfer process of the atmosphere at high altitudes. By refining the ozone concentration profile and atmospheric molecular distribution parameters, the influence of ozone absorption and atmospheric molecular scattering on remote sensing data is effectively eliminated, ensuring the radiometric accuracy of satellite data in high-altitude areas.

[0109] S212, UAV remote sensing data preprocessing

[0110] RGB / multispectral drone data was selected, and the drone adopted a low-altitude vertical overhead shooting mode, resulting in minimal geometric distortion in the image. Its core advantage is its high native resolution of 0.1m, which can capture fine details of farmland. The preprocessing core focuses on resolution adjustment based on slope level, while also incorporating illumination compensation based on slope aspect.

[0111] For different slope grades, corresponding resolution adjustment methods are adopted. High slope grades are key and difficult areas for detection, and satellite data has been discarded. The original resolution of 0.1m for UAV data is maintained to make full use of its high precision advantage, accurately depict the boundaries of small farmland and crop growth details on steep slopes, and avoid information loss due to reduced resolution.

[0112] For low and medium slope grades, satellite data of acceptable quality serves as the primary information source, while UAV data only serves as a supplementary verification function. A mean downsampling method is used to adjust its resolution to 1m. Mean downsampling calculates the average grayscale value of 2×2 adjacent pixels as the target pixel value, which can reduce the amount of data while retaining core information. The 1m resolution is consistent with the scale required for subsequent fusion, facilitating subsequent calculations.

[0113] For different slope aspects, corresponding illumination compensation methods are adopted. Sunny slopes receive ample sunlight, and some farmland areas are prone to overexposure due to strong light, resulting in blurred boundary details. Histogram equalization is used for illumination compensation. By adjusting the distribution range of image grayscale values, the contrast of overexposed areas is enhanced, restoring the farmland boundaries and crop texture details.

[0114] In areas with insufficient sunlight, the overall image brightness is low, and crop features are poorly distinguishable. Gamma correction is used for illumination compensation, with the gamma value set to 1.5. This value has been verified through extensive experiments on hilly farmland samples: when the gamma value is less than 1.5, the image is not bright enough; when it is greater than 1.5, noise in dark areas is amplified. A value of 1.5 can brighten the image while avoiding noise interference, ensuring that the farmland features in shady slope areas are clearly presented.

[0115] S213, Ground Sensor Data Preprocessing

[0116] The collected ground sensor data includes discrete point data such as soil moisture and pH value. These need to be interpolated to generate a continuous raster layer. The preprocessing core performs raster data interpolation around the slope level to ensure the spatial continuity and accuracy of the data.

[0117] For different slope grades, corresponding raster data interpolation methods are adopted. High slope grades have drastic terrain undulations, and soil properties (such as moisture and nutrients) are affected by topographic erosion and gravity, resulting in weak spatial distribution continuity. Ordinary Kriging interpolation, based on the assumption of "spatial autocorrelation," is prone to error amplification in this scenario due to spatial structure disruption. Therefore, an inverse distance-weighted interpolation method is used. The core logic of the inverse distance-weighted interpolation method is that the influence of a sample point on the target point decreases with increasing distance. Its calculation formula is:

[0118]

[0119] in, For the interpolation result of the target point, For the first Observations of each sample point For the target point and the first Euclidean distance of sample points This is the distance attenuation coefficient (set to 2 here, which has been verified by experiments to balance the influence of near and far sample points). The number of sample points to be interpolated (set to 10 here to ensure coverage of local terrain features). This method is less affected by complex spatial structures, and the interpolation results are more robust in steep slope areas, accurately reflecting the spatial heterogeneity of soil properties.

[0120] Low and medium slope grades have gentle terrain and continuous, stable spatial distribution of soil properties, meeting the assumptions of ordinary kriging interpolation. Therefore, ordinary kriging interpolation is adopted. This method fits the spatial variability structure of soil properties through a semi-variogram, quantifies spatial correlation, and reduces random errors. Its interpolation accuracy is higher than that of inverse distance weighted interpolation. The calculation formula is as follows:

[0121]

[0122] in, For the first The weights of each sample point (satisfying) The method, determined by solving the Kriging equations, ensures that the interpolation results are unbiased and have minimal variance. This method accurately reconstructs the spatial distribution patterns of soil properties, providing reliable ground data support for subsequent farmland type classification.

[0123] Regardless of the interpolation method used, the ground sensor data is generated as a 1m resolution raster layer. The 1m resolution is consistent with the resolution of the UAV remote sensing data (low / medium slope) and the subsequent adjusted satellite data, which can avoid fusion errors caused by resolution differences, ensure the consistency of the three types of data in spatial scale, and lay the foundation for subsequent weighted fusion.

[0124] S22, Terrain-weighted multi-source data fusion

[0125] After completing the multi-source data topographic preprocessing, the satellite remote sensing data (low / medium slope) is first adjusted from 0.5m resolution to 1m resolution through mean downsampling to ensure that all data sources have a uniform scale. Then, based on the slope level of S1, the fusion weights are assigned, and a uniform resolution detection layer is generated through linear weighted fusion to give full play to the advantages of various data sources under different terrain conditions.

[0126] S221. Construct the terrain-data weight matrix

[0127] Based on the slope grade, fusion weights are assigned to satellite remote sensing data, UAV remote sensing data, and ground sensor data. The weight allocation is based on "the information effectiveness and accuracy contribution of each data source under the corresponding terrain" and satisfies the constraint that "the sum of the weights is 1". The specific allocation is as follows:

[0128] Low-slope terrain is flat, and satellite data has a wide coverage and high geometric accuracy, making it the core information source. Its fusion weight is set at 60%. UAV data can supplement the lack of details in satellite data (such as field ridges and crop spacing), and its weight is set at 30%. Ground sensor data is only used to calibrate the radiometric bias of remote sensing data, and its weight is set at 10%.

[0129] Satellite data with medium slope exhibits slight geometric distortion, reducing its information effectiveness; therefore, its weight is reduced to 40%. UAV data, due to its low-altitude imaging advantages and enhanced terrain adaptability, can more accurately capture changes in farmland morphology caused by terrain undulations; therefore, its weight is increased to 45%. Ground sensor data has enhanced calibration capabilities (such as correcting remote sensing inversion errors in soil moisture caused by slope); therefore, its weight is adjusted to 15%.

[0130] High-slope satellite data has been discarded and its weight is set to 0%; UAV data has a weight of 70%; ground sensor data can provide accurate ground truth values ​​(such as actual soil moisture and pH value) and is used to verify the accuracy of UAV data, with a weight of 30%.

[0131] This weighting ensures that each data source is used to its maximum effect in adapting to the terrain, avoiding interference from invalid data.

[0132] S222, Linear Weighted Fusion Calculation

[0133] Satellite remote sensing data, UAV remote sensing data, and ground sensor data, all unified to a 1m resolution, are fused using a linear weighted fusion method. Linear weighted fusion, based on the principle that "the higher the accuracy of each data source, the greater its weight," efficiently integrates multi-source information. The derivation of its calculation formula is as follows:

[0134] Let the grayscale value of a certain pixel in the merged detection layer be... The corresponding pixel grayscale value of satellite remote sensing data The weight is The corresponding pixel grayscale value of the UAV remote sensing data is The weight is The pixel grayscale value corresponding to the ground sensor data is The weight is Since the three types of data have undergone radiometric normalization (calibrated using ground sensor data), and the physical meaning of the pixel grayscale values ​​is consistent, fusion can be achieved directly through weighted summation, as shown in the formula:

[0135]

[0136] in, This constraint ensures that the grayscale values ​​of the merged pixels are within a reasonable dynamic range (0-255), avoiding distortion of farmland features caused by abnormal brightness.

[0137] For example, in a pixel at a low slope level, , , Substitute , , , can be obtained This result combines the stability of satellite data, the detail of UAV data, and the accuracy of ground data, and can accurately reflect the actual condition of farmland.

[0138] S223, Output a uniform resolution detection layer

[0139] Through the above fusion calculations, a unified detection layer with a 1m resolution for terrain adaptation is finally output. This layer adopts the GeoTIFF format, supports the WGS84 geographic coordinate system, and can be accurately overlaid with the terrain feature layer output by S1, ensuring spatial matching between terrain features and farmland data in subsequent steps.

[0140] This detection layer addresses several key issues: First, it overcomes the shortcomings of existing technologies, such as inconsistent scale and uncorrected distortion in multi-source data. The multi-source data fusion error is reduced from 0.5m to 0.3m, significantly improving data accuracy. Second, it maximizes the advantages of various data sources through terrain-weighted fusion. For example, low-slope data relies on satellite data for wide coverage, while high-slope data relies on UAV data for high-precision detection. Ground data is calibrated throughout the process to ensure reliability. Third, the 1m resolution meets the accuracy requirements for farmland boundary extraction and area calculation while controlling the data volume, making it suitable for large-scale hilly farmland detection applications and providing a high-quality data foundation for subsequent farmland attribute identification.

[0141] S3. Terrain recognition of farmland morphology and type

[0142] like Figure 4 As shown, this step is based on the 1m resolution unified detection layer output by S2 and the terrain features such as slope level and slope aspect category extracted by S1. It deeply embeds terrain factors into the farmland attribute recognition logic, solves the problem of insufficient accuracy in farmland morphology and type recognition in hilly areas, and provides accurate farmland attribute basis for subsequent dynamic adjustment of detection strategies.

[0143] S31. Farmland boundary and topographic extraction

[0144] S311, Precise Identification of Farmland Boundaries

[0145] An improved Canny edge detection algorithm is used to extract farmland boundaries, which can effectively distinguish the boundary contours between farmland and non-farmland areas (such as roads and woodlands). The edge detection threshold is positively correlated with the slope level, and the threshold is adaptively adjusted according to the slope level determined by S1 (low slope ≤ 5°, medium slope 5°-25°, high slope ≥ 25°).

[0146] Low-slope terrain is flat, and farmland boundaries are less affected by terrain but their outlines are relatively blurred (e.g., the color difference between field ridges and soil is small in gentle slope areas). A low threshold can better preserve boundary details, so the threshold is set to 0.3. Medium-slope terrain has some undulations, and farmland boundaries are moderately clear. The threshold is set to 0.5 to balance detail preservation and noise suppression (e.g., false edges caused by filtering weeds). High-slope terrain is highly undulating, and farmland boundaries are more clearly delineated by terrain (e.g., fields in steep slope areas have obvious boundaries due to terrain differences). A high threshold can reduce non-boundary noise interference, so the threshold is set to 0.7.

[0147] These thresholds were verified by a sample test in hilly farmland: when the thresholds were below the corresponding thresholds, the noise interference rate in low-slope areas increased from 12% to 28%; when the thresholds were above the corresponding thresholds, the loss rate of boundary details in high-slope areas increased from 8% to 23%. The current values ​​can achieve the optimal balance between boundary extraction accuracy (≥96%) and noise resistance.

[0148] For areas with high slope gradients, terrain shadow regions are identified by combining the slope aspect categories determined by S1 (sunny slopes 0°-180°, shady slopes 180°-360°). Shady slopes, due to insufficient sunlight, easily form terrain shadows at the bottom, which can lead to "false edges in dark areas" in the image (e.g., the grayscale difference between the shadow and farmland is misjudged as a boundary), interfering with boundary recognition accuracy. Using a triangular relationship model of "solar azimuth angle-slope aspect-elevation difference," the shadow coverage area is calculated, and the edge detection threshold within the identified shadow region is increased from 0.7 to 0.8. Experimental results show that this adjustment reduces the false edge misjudgment rate in high-slope shady slope areas from 15% to below 5%, while avoiding the loss of true farmland boundaries.

[0149] After extracting the polygonal boundaries of the farmland using an edge detection algorithm, three main morphological parameters are calculated:

[0150] 1. Farmland area The formula for calculating the area of ​​a polygon is derived based on the principle of the cross product of vectors at boundary coordinate points. The formula is as follows:

[0151]

[0152] in, For continuous coordinate points of the polygon boundary, (Closed polygon) The unit is (Subsequent conversion to "mu"). This formula can accurately calculate the area of ​​any irregular polygon, making it suitable for the diverse shapes of fields in hilly areas.

[0153] 2. Farmland perimeter The length of each side of the polygon boundary is obtained by summing the lengths of each side using the distance formula between two points. Calculate, in meters.

[0154] 3. Shape complexity : Through formula It is deduced that, among which It is a dimensionless index. For the perimeter of the farmland, This represents the area of ​​farmland. The derivation of this formula is based on the geometric characteristics of farmland shape: the more irregular the boundaries (such as fields on steep slopes cut by gullies), the larger the ratio of perimeter to area. The higher the value, the more objectively it can quantify the degree of fragmentation of farmland morphology, providing a quantitative indicator for subsequent morphology type identification.

[0155] S312, Identification of Farmland Morphology and Type

[0156] A combined rule of area, slope and shape complexity is adopted to replace the existing single area threshold judgment method, avoiding shape misjudgment caused by terrain influence (such as large but fragmented farmland in steep slope areas being misjudged as contiguous farmland).

[0157] First, set the key threshold:

[0158] Area thresholds: The first area threshold is 1 mu (0.067 hectares), and the second area threshold is 2 mu (0.067 hectares). The second area threshold is greater than the first area threshold. 1 mu is chosen as the first threshold because in hilly areas of my country, plots smaller than 1 mu account for 42%, generally exhibiting a "small and scattered" characteristic. 2 mu is chosen as the second threshold because when the area is larger than 2 mu and the terrain is flat, the plots are often connected by field ridges to form contiguous areas, which meets the definition of contiguous farmland.

[0159] Shape complexity thresholds: The first complexity threshold is 12, and the second complexity threshold is 10, with the first complexity threshold being higher than the second complexity threshold. These thresholds were derived from statistical analysis of hilly farmland samples: 12 is the critical value for "morphological fragmentation" in farmland with an area of ​​1-2 mu (farms exceeding this value have boundary irregularities exceeding 60%), and 10 is the critical value for "morphological fragmentation" in farmland with an area of ​​more than 2 mu under high slope conditions (in high-slope areas, due to terrain fragmentation, even large areas with a complexity exceeding 10 still exhibit small-scale characteristics).

[0160] A farmland is considered a small plot of farmland if any of the following conditions are met:

[0161] 1. Farmland area is less than 1 mu;

[0162] 2. The farmland area is between 1 mu and 2 mu, and the slope grade is not lower than the low slope grade (i.e., medium slope or high slope), while the shape complexity is higher than 12.

[0163] 3. The farmland area is greater than 2 mu, and the slope grade is high slope grade (≥25°), and the shape complexity is higher than 10.

[0164] If none of the above conditions are met, the farmland is determined to be contiguous farmland. Specifically, one of the following two conditions must be met simultaneously:

[0165] 1. The farmland area is greater than or equal to 2 mu, and the slope grade is low slope grade (≤5°), while the shape complexity is less than or equal to 12;

[0166] 2. The farmland area is between 1 mu and 2 mu, and the slope grade is low slope grade (≤5°), while the shape complexity is less than or equal to 10.

[0167] This discrimination rule fully considers the influence of topography on farmland morphology. For example, farmland with an area of ​​3 mu but a shape complexity of 11 under a high slope grade is classified as such because "area > 2 mu, high slope, "It was determined to be a small plot of farmland, avoiding the problem of existing technology misclassifying it as contiguous farmland based solely on area; farmland with an area of ​​1.5 mu and a shape complexity of 9 under a low slope grade was classified as such." low slope "It was determined to be contiguous farmland, which is consistent with the actual distribution characteristics of fields in gentle slope areas."

[0168] S32. Classification of Farmland Types

[0169] A logistic regression model based on topographic association features and physical detection features was used to classify farmland types.

[0170] S321. Construction of Terrain-Related Features

[0171] Based on the slope aspect category and slope grade determined by S1, terrain-related features are constructed, and qualitative terrain features are transformed into quantitative indicators that the model can process through numerical encoding:

[0172] Slope aspect category coding: sunny slopes (0°-180°) are coded as 1, and shady slopes (180°-360°) are coded as 0. This rule intuitively reflects the impact of light conditions on organic farmland (sunny slopes are more conducive to photosynthesis of organic crops).

[0173] Slope grade coding: Low slope grade (≤5°) is coded as 3, medium slope grade (5°-25°) is coded as 2, and high slope grade (≥25°) is coded as 1. The higher the slope grade, the smaller the code value, which is in line with the actual rule that "the higher the slope, the more difficult it is to grow organic farmland" (such as the easy loss of organic fertilizer in steep slope areas).

[0174] S322. Extraction of Variation Features in Vegetation Index Time Series

[0175] Based on the unified detection layer output by S2, NDVI (Normalized Difference Vegetation Index) time series (12 consecutive periods, covering one crop growth cycle) are extracted at a frequency of "5 days / time". The formula for calculating NDVI is:

[0176]

[0177] in, For near-infrared reflectivity, For red light band reflectivity, The range of values ​​is A higher value indicates better crop growth.

[0178] Organic farmland does not use chemical ripening agents, resulting in stable crop growth rates. The time series shows small fluctuations; in conventional farmland, due to the use of chemical fertilizers, crop growth exhibits significant "fast-slow" fluctuations. (Calculation) Standard deviation of time series As a fluctuation indicator, a threshold is set for organic farmland on sunny slopes. ≤0.08, shady slope area ≤0.10. This threshold was determined through a comparative experiment of organic and conventional farmland: sunny slopes with ample sunlight are more conducive to organic crops. Smaller fluctuations (Average 0.06), shady slopes receive insufficient sunlight, even in organic farmland. The volatility was also slightly higher. Average value 0.09).

[0179] S323. Acquisition of Soil Microbial Diversity Characteristics

[0180] Based on ground sensor data (soil samples were collected at a depth of 20 cm, with one sampling point per 10 mu), the Shannon index was used to quantify soil microbial diversity. The formula is as follows:

[0181]

[0182] in, The Shannon index, This represents the total number of soil microbial species (such as bacteria and fungi). For the first The relative abundance of a species (the proportion of a particular species to the total number of microorganisms). This formula comprehensively reflects species richness and evenness; a higher index indicates greater microbial diversity.

[0183] Organic farmland, due to the long-term use of organic fertilizers, has a more stable soil ecosystem and higher microbial diversity. Threshold setting: Sloping organic farmland (slope ≥ 5°). Flat land (slope ≤ 5°) organic farmland The threshold was determined by soil sample testing: slope soils have better aeration, providing a more favorable environment for microbial survival. (Average value 3.3), flat soils are prone to compaction and have slightly lower microbial diversity. (Average value 3.1), the current threshold can accurately distinguish the soil ecological differences between organic and conventional farmland (conventional farmland) Average value 2.5).

[0184] S324. Acquisition of pesticide residue characteristics

[0185] GC-MS (Gas Chromatography-Mass Spectrometry) was used to detect pesticide residues in soil samples. This technology can identify common pesticide components such as organophosphates and pyrethroids. Since the use of chemical pesticides is prohibited in organic farmland, a pesticide residue threshold of 0 mg / kg (i.e., no chemical pesticide residues were detected) was set. This threshold is a key hard indicator for distinguishing organic from conventional farmland.

[0186] S325, Integration and Generation of Organic Farming Confidence Level

[0187] The above four types of features (terrain-related features, The fluctuation characteristics, Shannon index characteristics, and pesticide residue characteristics were normalized (mapped to) (Interval), input to the logistic regression model. The model outputs the organic farming confidence score through weighted summation and Sigmoid function transformation. The formula is:

[0188]

[0189] in, The weights of each feature (derived from sample training): pesticide residues and The volatility has the highest weight because it has the most significant impact on farmland type. These are the normalized eigenvalues. For bias terms , The range of values ​​is The higher the value, the greater the probability that the farmland is organic farmland.

[0190] S326. Classification and Determination of Farmland Type

[0191] The corresponding confidence threshold was selected based on the slope aspect category, with a value of 0.85 set for both sunny and shady slopes. This threshold was verified experimentally: when... At this time, the accuracy rate for identifying organic farmland is ≥94% (false negative rate ≤6%), while the recall rate is ≥92% (false positive rate ≤8%), which can balance the needs of "avoiding false negatives for organic farmland" and "preventing false negatives for conventional farmland". If a uniform threshold of 0.8 is used, the false negative rate for organic farmland on shady slopes increases from 8% to 14%; if 0.9 is used, the false negative rate for organic farmland on sunny slopes increases from 6% to 12%, and 0.85 is the optimal threshold.

[0192] When the confidence level for organic farming reaches 0.85, the farmland is classified as organic; otherwise, it is classified as conventional farmland. This classification method fully considers the influence of topography on farmland type. For example, a farmland on a shady slope... Fluctuation 0.09 (close to the threshold of 0.10), Shannon index 3.1, no pesticide residue, terrain-related feature encoding 0.5, model output. The land was identified as organic farmland, avoiding the limitations of existing technologies due to its "shady slope". The problem of "slightly high fluctuations" being mistakenly identified as regular farmland.

[0193] S4. Dynamic terrain switching of detection strategy

[0194] like Figure 5As shown, this step is based on the farmland morphology type (contiguous farmland, small plots of farmland), farmland type (organic farmland, conventional farmland) identified by S3, and the slope level (low slope ≤ 5°, medium slope 5°-25°, high slope ≥ 25°) and slope aspect category (sunny slope 0°-180°, shady slope 180°-360°) extracted by S1. It dynamically adjusts the UAV flight parameters (flight path spacing, oblique cross flight angle) and sensor deployment density (ground sensor spacing). At the same time, it sets differentiated detection thresholds to solve the problem that efficiency and accuracy are difficult to balance due to the fixed detection strategy of existing technologies, and ensures that the detection needs of different terrains and farmland types can be met.

[0195] S41. Parameter adjustment based on farmland morphology and slope grade

[0196] Based on the farmland morphology determined by S3 and the slope level determined by S1, the spacing between UAV flight paths, the angle of oblique cross-flight of UAVs, and the spacing between ground sensors are set accordingly. The core of this adjustment logic is: for contiguous farmland with concentrated distribution, the parameter design needs to balance efficiency and coverage; for small farmland with fragmented morphology and scattered distribution, the parameter design needs to prioritize accuracy to avoid missed detections; the higher the slope level, the more significant the terrain obstruction, and the blind spots need to be eliminated by reducing the spacing and increasing the cross-flight angle.

[0197] S411, Parameter settings for contiguous farmland

[0198] When the farmland is contiguous, the spacing between UAV flight paths and the spacing between ground sensors decrease as the slope grade increases, while the angle of UAV diagonal cross flight increases as the slope grade increases. This gradient adjustment adapts to the differences in terrain occlusion under different slopes.

[0199] The low-slope, contiguous farmland has flat terrain with no significant topographical obstructions. The drone flight path spacing and ground sensor spacing were both set to 10m, and the drone's diagonal cross-flight angle was set to 0° (no diagonal cross-flight required). Choosing 10m as the spacing parameter resulted in 98% detection coverage, only 5% data redundancy, and the detection time for a single 10,000-mu (approximately 667 hectares) was controlled within 5 hours, achieving a balance between "high coverage, low redundancy, and high efficiency." The 0° cross-flight angle further reduced invalid flight paths, improving detection efficiency.

[0200] The terrain of contiguous farmland with medium slopes is somewhat undulating, and there is a risk of occlusion in some areas. Therefore, the spacing between drone flight paths was reduced from 10m to 8m, and the spacing between ground sensors was simultaneously reduced to 8m. This reduction in spacing increases coverage density. The drone's diagonal cross-flight angle was set to 15° to supplement areas obstructed by the forward flight path. The rationale for this parameter adjustment is that in medium slope areas, using a low-slope 10m spacing results in a 6% blind spot rate, while the combination of 8m spacing and a 15° cross-angle reduces the blind spot rate to below 2%. Simultaneously, the detection time per 10,000 mu (approximately 667 hectares) only increases by 1 hour, still meeting the timeliness requirements for large-scale farmland detection.

[0201] High-slope contiguous farmland features dramatic terrain undulations and significant shading issues. To maximize coverage density, the spacing between drone flight paths and ground sensors was further reduced to 5 meters. The drone's diagonal cross-flight angle was increased to 30°, allowing for more comprehensive coverage of terrain shadows and obstructed areas on steep slopes. The 5-meter spacing and 30° cross-flight angle were chosen because this parameter combination achieves over 95% detection coverage in high-slope contiguous farmland, significantly improving upon the 65% coverage of existing technologies with fixed parameters (10-meter spacing, 0° cross-flight angle). This significantly addresses the issue of missed detections due to steep slope shading, while keeping data redundancy below 12%, avoiding excessive computational costs.

[0202] S412, Parameter settings for small farmland

[0203] When the farmland is small plots, smaller drone flight path spacing and ground sensor spacing, as well as larger drone diagonal cross-flight angles, are used compared to contiguous farmland with the same slope grade. Small plots of farmland are small in area and scattered, requiring higher density detection coverage to avoid missed detections.

[0204] For small plots of farmland with low slopes: the spacing between drone flight paths was reduced from 10m to 6m for contiguous farmland with the same slope, the spacing between ground sensors was reduced from 10m to 6m, and the angle of the drone's diagonal cross flight was increased from 0° to 10°. The 6m spacing can accurately cover the boundaries and internal areas of small plots of farmland with an area of ​​less than 1 acre, and the 10° cross angle can supplement the details of the field edges. Experiments have verified that under these parameters, the missed detection rate of small plots of farmland with low slopes has been reduced from 15% to less than 3%, and the detection time for a single 10,000 acres can be controlled within 8 hours, balancing accuracy and efficiency.

[0205] For small plots of farmland with moderate slopes: the spacing between UAV flight paths is reduced from 8m to 4m for contiguous farmland with the same slope, the spacing between ground sensors is reduced from 8m to 4m, and the angle of oblique cross-flight of the UAVs is increased from 15° to 25°. The 4m spacing can accommodate the irregular boundaries formed by terrain cutting in small plots of farmland in moderate slope areas, and the 25° cross-angle can cover the gaps between plots. This combination of parameters enables the detection integrity of small plots of farmland with moderate slopes to reach 97%, solving the problem of missed detection of scattered plots caused by excessive spacing in existing technologies.

[0206] For small plots of farmland with steep slopes: the spacing between drone flight paths was reduced from 5m to 3m for contiguous farmland with the same slope, the spacing between ground sensors was reduced from 5m to 3m, and the angle of the drone's diagonal cross flight was increased from 30° to 45°. 3m is the minimum effective spacing for detecting small plots of farmland with steep slopes. A spacing smaller than 3m will result in a data redundancy rate exceeding 20%, increasing the processing burden; a spacing greater than 3m will cause the false negative rate to rise back to over 8%; the 45° cross angle can fully cover the shaded area of ​​the steep slope, avoiding false negatives at the edge of the plots due to terrain obstruction.

[0207] S42. Adjustment of detection thresholds based on slope aspect category and farmland type

[0208] Based on the farmland type (organic farmland, conventional farmland) determined by S3 and the slope aspect category (sunny slope, shady slope) determined by S1, and combined with the vegetation index time series variation characteristics extracted by S3, differentiated pest and disease thresholds (specifically referring to aphid density, since aphids are the main pests of farmland in hilly areas) and upper limits of NDVI time series fluctuations are set. At the same time, the soil microbial detection frequency and pesticide residue detection methods are adjusted to ensure that the detection thresholds are accurately adapted to management needs and topographic microenvironment.

[0209] S421, Threshold setting for organic farmland

[0210] Organic farmland prohibits the use of chemical pesticides, relying instead on ecological methods (such as the introduction of natural enemies and biological pesticides) for pest and disease control. Stricter detection thresholds are needed for early warning. Furthermore, slope aspect affects light and temperature differences, impacting pest and disease reproduction and crop growth, necessitating targeted adjustments to the thresholds. Sunny slopes in organic farmland receive ample sunlight and have average daily temperatures 3-5°C higher than shady slopes, leading to faster aphid reproduction (a shortened generation cycle of 2-3 days). Therefore, a pest and disease threshold of ≤3 aphids / 100 plants is set. Experiments on sunny slopes have shown that when aphid density exceeds 3 aphids / 100 plants, the risk of an outbreak reaches 80% within 7 days. The 3 aphid / 100 plant threshold allows sufficient response time for ecological control measures (such as releasing aphid wasps) to prevent the spread of pests.

[0211] Based on the time series variation characteristics of vegetation index extracted by S3, the organic farmland on the sunny slope has stable light and regular crop growth. The upper limit of NDVI time series fluctuation is set at 0.08. Exceeding this upper limit indicates that crop growth may be affected by pests, diseases, or water stress, and further verification is required.

[0212] Organic farmland on shady slopes receives insufficient sunlight and has a daily average humidity 15%-20% higher than that on sunny slopes. Aphids reproduce relatively slowly, but fungal diseases (such as sheath blight) are more likely to occur. The pest and disease threshold is set at ≤5 aphids / 100 plants. The threshold of 5 aphids / 100 plants avoids false alarms caused by overly strict thresholds (the natural mortality rate of aphids is higher on shady slopes) while effectively controlling aphid damage. Based on the NDVI characteristics of S3, the light fluctuations on shady slopes are large, and crop growth is more significantly affected by light. The upper limit of NDVI time series fluctuation is set at 0.10. This value is suitable for the light conditions on shady slopes and reduces misjudgments of growth abnormalities caused by sudden changes in light (such as a temporary increase in NDVI due to short periods of sunny weather on shady slopes).

[0213] In addition, the frequency of soil microbial testing in organic farmland needs to be adapted to the influence of slope aspect on microbial activity: sunny slopes have higher temperatures and faster changes in microbial activity, so the testing frequency is set to once every two weeks; shady slopes have lower temperatures and more stable microbial activity, so the testing frequency is set to once a month. Pesticide residue testing uses gas chromatography-mass spectrometry for complete detection, with a detection limit of 0.001 mg / kg.

[0214] S422, Threshold setting for conventional farmland

[0215] Chemical pesticides (such as imidacloprid) can be used to control pests and diseases in conventional farmland. The detection threshold can be appropriately relaxed. At the same time, the slope aspect should be adjusted to match the actual occurrence pattern of pests and diseases, so as to avoid the waste of manpower and costs caused by excessive detection.

[0216] In conventional farmland on sunny slopes, the higher temperature leads to a 40% higher aphid incidence rate compared to shady slopes. A pest and disease threshold of ≤15 aphids per 100 plants is set. When the aphid density exceeds 15 aphids per 100 plants, the efficacy of chemical pesticides reaches over 90%, meeting conventional control needs while avoiding pesticide overuse due to premature application. Based on the NDVI characteristics of S3, crop growth rates in conventional farmland on sunny slopes fluctuate significantly due to fertilizer application (such as urea). Therefore, an upper limit for NDVI time series fluctuation is set at 0.15 to adapt to fertilizer-driven growth rhythm changes and reduce misjudgments.

[0217] In shady slopes of conventional farmland, high humidity and low temperature result in a lower frequency of aphid occurrences. Therefore, a pest and disease threshold of ≤20 aphids / 100 plants is set. This threshold aligns with the aphid occurrence patterns on shady slopes: aphid populations grow slowly on shady slopes, and pesticide application at this threshold can still effectively control the infestation. Furthermore, it represents a 30% wider threshold than for sunny slopes, reducing unnecessary pesticide use. Based on the NDVI characteristics of S3, shady slopes of conventional farmland experience insufficient sunlight and relatively slow crop growth. Therefore, the upper limit of NDVI time series fluctuation is set to 0.20 to match the growth characteristics under low light conditions and avoid misjudgments caused by slow growth.

[0218] The frequency of soil microbial testing in conventional farmland is set at once per quarter. Since conventional farmland uses chemical fertilizers and pesticides, the microbial diversity is relatively stable, and high-frequency testing is not required. Pesticide residue testing uses a rapid reagent kit with a detection limit of 0.01 mg / kg, which meets the screening requirements for pesticide residues in conventional farmland (the pesticide residue limit standard in conventional farmland is mostly above 0.05 mg / kg). Moreover, the detection efficiency is 5 times higher than that of gas chromatography-mass spectrometry, making it suitable for the rapid testing needs of large-scale conventional farmland.

[0219] This step addresses the issues of coverage and missed detection in complex terrain by adjusting flight parameters and sensor density based on farmland morphology and slope grade. High-density parameters for small farmland ensure no missed detections. By adjusting detection thresholds and frequencies based on slope aspect and farmland type, it resolves misjudgments caused by differences in terrain microenvironment and management needs. Strict thresholds are applied to organic farmland for ecological control, while wide thresholds are applied to conventional farmland for chemical control. Differentiated thresholds for sunny / shady slopes are adapted to microenvironmental patterns, comprehensively addressing the core shortcomings of farmland detection in hilly areas: insufficient accuracy, low efficiency, and poor adaptability.

[0220] S5, Closed-Loop Feedback Terrain Optimization

[0221] like Figure 5 As shown, this step is based on the UAV flight parameters (flight path spacing, diagonal cross flight angle), sensor deployment density (ground sensor spacing) and detection threshold dynamically adjusted by S4. By obtaining the detection error and false judgment rate, a closed-loop optimization mechanism is constructed to iteratively adjust the relevant parameters. This solves the defects of existing technology detection strategies that are fixed and cannot adapt to dynamic changes in terrain, and achieves long-term dynamic adaptation between terrain and detection strategy, continuously improving detection accuracy and efficiency.

[0222] S51. Feedback Indicator Acquisition and Calculation

[0223] Two core feedback indicators, detection error and false positive rate, were obtained. All data were obtained through field verification and stratified random sampling statistics to ensure the authenticity and reliability of the indicators, providing an accurate basis for parameter optimization. The calculation logic of the indicators directly corresponds to the detection parameters (pest and disease density, soil moisture) in S4.

[0224] S511, Calculation of Detection Error

[0225] The detection error is obtained by comparing the detection value output by the AI ​​model with the actual value of the ground sampling. It focuses on calculating two key indicators: the relative error of pest and disease density and the relative error of soil moisture, so as to avoid the problem of insufficient adaptability of absolute error due to different detection bases.

[0226] The relative error of pest density is used to assess the accuracy of aphid density detection. Aphids are a key pest type monitored in S4, and the accuracy of their density detection directly affects control decisions. The calculation formula is:

[0227]

[0228] in, This represents the relative error in pest and disease density. Aphid density detected by the AI ​​model (unit: heads / 100 plants). The actual aphid density value (unit: heads / 100 plants) is obtained from on-site sampling. A relative error threshold of 3% for pest density is set. If the aphid density detection error exceeds 3%, it will lead to misjudgment of the timing of prevention and control. If the error is too large, low-density aphids may be misjudged as meeting the standard and the early prevention and control window may be missed, while high-density aphids may be misjudged as exceeding the standard and over-control may be implemented.

[0229] The relative error of soil moisture is used to assess the accuracy of soil moisture detection. Soil moisture is a core indicator for crop growth status assessment in S4, and its accuracy affects irrigation decisions. The calculation formula is as follows:

[0230]

[0231] in, This is due to the relative error in soil moisture. Soil moisture value detected by AI model (unit: %, volumetric water content). This represents the actual soil moisture value obtained from on-site sampling (unit: %, volumetric water content). A relative error in soil moisture exceeding 5% can lead to an irrigation deviation of over 10%, thereby affecting crop growth. For example, in arid areas, irrigation may be delayed due to misjudging that the moisture level is within acceptable limits, while in humid areas, over-irrigation may occur due to misjudging that the moisture level is too low.

[0232] Field sampling follows the principle of "stratified random sampling": the slope levels (low, medium, high) divided by S1 and the farmland morphology types (contiguous areas, small patches) determined by S3 are combined into 6 sampling units, with 30 sampling points set up in each unit. The spacing between sampling points is consistent with the spacing between ground sensors to ensure sample representativeness. Aphid density sampling adopts the "five-point sampling method" for manual counting, and soil moisture sampling adopts a high-precision time domain reflectometer (detection accuracy ±1%). The sampling frequency is synchronized with the detection frequency of S4 (once every 5 days) to avoid errors caused by time differences.

[0233] S512, Calculation of False Positive Rate

[0234] The false positive rate is obtained by statistically analyzing the proportion of abnormal false positive samples. Abnormal false positive samples include two categories: samples that the AI ​​model misclassifies as "excessive pests and diseases" but are normal in actual field testing, and samples that the AI ​​model misclassifies as "normal" but are "excessive pests and diseases" in actual field testing. Both types of samples can lead to management decision-making errors and must be included in the statistics. The formula for calculating the false positive rate is:

[0235]

[0236] in, For the false positive rate, To avoid misjudging the number of samples exceeding the standard, To misjudge the number of normal samples, This represents the total number of valid samples tested across the entire testing area each month (excluding invalid samples due to sampling failures or missing data).

[0237] Setting the false positive rate threshold to 8% means that if the false positive rate exceeds 8%, the risk of management decision-making errors based on the test results will increase from 5% to more than 15%. If the false positive rate exceeds the limit, it may lead to more than 20% waste of pesticides. If the false positive rate is normal, it may lead to more than 30% risk of pest and disease spread. The 8% threshold can ensure the reliability of the test results, while avoiding the increase in optimization costs caused by an overly strict threshold (such as 5%).

[0238] The sample statistics cover all 6 sampling units, with no less than 50 valid samples per unit per month, to ensure that the false positive rate can reflect the detection performance under different terrains and farmland morphologies, and to avoid statistical bias caused by single-scene samples.

[0239] S52, Terrain Parameter Optimization Rules

[0240] Based on detection errors and misjudgment rates, differentiated parameter optimization rules were formulated for six combinations of S1 slope level and S3 farmland morphology type to ensure that adjustment measures are accurately adapted to specific scenarios and avoid the imbalance between accuracy and efficiency caused by "one-size-fits-all" optimization. All adjustments are based on the gradient optimization of the initial parameters of S4.

[0241] Steep slopes and small plots of farmland (slope ≥ 25°, small plots) represent the most challenging scenario for detection in S4. The optimization rules focus on increasing detection density and shadow coverage. When the relative error of soil moisture is > 5%, it indicates that the current 3m ground sensor spacing is too large to capture drastic changes in soil moisture in steep slope areas (e.g., the difference in moisture between the top and bottom of the slope can reach 15%). Reducing the ground sensor spacing from 3m to 2m increases the density of soil moisture sampling points by 125%, reduces the interpolation error to below 4%, and only increases sensor deployment costs by 30%. When the relative error of pest and disease density is > 3% or the misjudgment rate is > 8%, it indicates that the 45° diagonal cross-flying angle of the drone in S4 cannot cover the shadow area of ​​the steep slope. Increasing the angle from 45° to 60° increases the shadow area coverage from 82% to 98%. If the error still exceeds the standard after adjustment, the drone flight path spacing is further reduced from 3m to 2.5m. By increasing the flight density, detailed capture is enhanced, ensuring that the aphid density detection error is reduced to below 3%.

[0242] Gentle slopes and contiguous farmland (slope ≤ 5°, contiguous farmland) represent the scenario with the highest detection efficiency in S4. Optimization rules emphasize "fine-tuning accuracy, prioritizing efficiency." When the relative error of pest and disease density is > 3%, reducing the drone flight path spacing from 10m to 8m increases the detection point density by 56%, reduces the error to below 2.5%, and only increases the detection time per 10,000 mu by 0.5 days, resulting in an efficiency loss of less than 10%. When the false positive rate is > 8%, increasing the drone's oblique cross-flight angle from 0° to 5° supplements the detection details at the edges of contiguous farmland, reducing the false positive rate to below 6% without adding extra flight time. When the relative error of soil moisture is > 5%, reducing the ground sensor spacing from 10m to 8m matches the flight path spacing, ensuring spatial coordination between soil moisture data and aerial data.

[0243] Medium-slope contiguous farmland (5° < slope < 25°, contiguous farmland): When the index exceeds the standard, the spacing between UAV flight paths is reduced from 8m to 6m, the spacing between ground sensors is reduced from 8m to 6m, and the angle of UAV diagonal cross flight is increased from 15° to 20° to adapt to the slight undulation of medium slope. The error can be reduced to within the threshold, and the detection efficiency only decreases by 8%.

[0244] For small plots of farmland with moderate slope (5° < slope < 25°, small plots of farmland): When the indicators exceed the standard, the spacing between UAV flight paths is reduced from 4m to 3.5m, and the spacing between ground sensors is reduced from 4m to 3.5m to balance accuracy and data volume. The angle of UAV diagonal cross flight is increased from 25° to 30° to cover the shaded area of ​​moderate slope, and the error can be reduced to within the threshold.

[0245] All adjustments follow the principle of "gradient fine-tuning," with each parameter adjustment not exceeding 20% ​​to avoid data redundancy (e.g., reducing the spacing directly from 4m to 2m would increase the data volume by 300%) or insufficient accuracy caused by cross-level adjustments, ensuring that each optimization can specifically address the specific problem.

[0246] S53, Parameter Iterative Updates and Long-Term Adaptation

[0247] The system sets a monthly iteration cycle and automatically collects feedback indicator data from the entire region through the detection system. Based on the optimization rules, it updates the "terrain-data weight matrix" (S2 construction) and the "strategy parameter table" (S4 formulation) to achieve long-term dynamic adaptation of detection parameters and avoid the long-term accuracy decline caused by the "one-time parameters" of existing technologies.

[0248] The iterative update adopts a logic of local first and then global: First, for a single out-of-standard combination unit (such as a small plot of farmland on a steep slope), select 10% of the area of ​​the unit for parameter adjustment, continuously monitor for 7 days, and after confirming that the detection error and false judgment rate have dropped to within the threshold, proceed to the second step; Second, synchronize the optimized parameters to combination units with the same characteristics in the entire area (such as all small plots of farmland on steep slopes) to avoid duplicate optimization; Third, update the "terrain-data weight matrix" of S2, such as fine-tuning the weight of UAV data in steep slope areas from 70% to 75%, and fine-tuning the weight of ground data from 30% to 25%, to adapt to the higher accuracy of the optimized UAV parameters and ensure the coordination of data fusion and detection parameters.

[0249] After monthly updates, the adjusted parameters undergo dual verification for both accuracy and efficiency: accuracy verification is performed through on-site sampling comparisons to ensure that error and false positive rates meet standards; efficiency verification is performed by statistically analyzing the testing time per 10,000 mu (approximately 667 hectares) to ensure that it does not exceed 5 days. This verification process avoids efficiency reduction caused by optimized parameters, ensuring a balance between improved accuracy and guaranteed efficiency.

[0250] This closed-loop optimization mechanism has two main aspects: First, it dynamically adapts to minor changes in terrain (such as changes in field morphology caused by seasonal rain erosion) and adjustments in farmland management (such as changes in fertilization methods in organic farmland), continuously correcting parameter deviations and ensuring that the detection system maintains high accuracy over the long term. Second, it replaces manual adjustments with data-driven approaches, reducing subjective errors and making the adaptation of parameters to terrain and farmland attributes more scientific, thus comprehensively solving the problem of insufficient long-term adaptability caused by fixed detection parameters in existing technologies.

[0251] The implementation principle of the AI-based integrated air-ground farmland detection method in this application is as follows: First, digital elevation model data of the area to be detected is acquired, slope, aspect, and altitude features are extracted and classified. Then, adaptive terrain preprocessing is performed on satellite remote sensing data, UAV remote sensing data, and ground sensor data for different terrain features. Subsequently, fusion weights are assigned to the multi-source data according to slope, and linear weighted fusion is used to generate a 1m resolution detection layer. Next, farmland boundaries are extracted using an improved Canny edge detection method based on the detection layer. The morphological type of small / contiguous farmland is determined by combining area, shape complexity, and slope. Terrain-related features, NDVI fluctuations, soil microbial diversity, and pesticide residues are then fused. The system generates organic farming confidence scores to classify farmland types, then dynamically adjusts UAV parameters and sensor density based on terrain and farmland attributes. Finally, it iteratively optimizes parameters through detection errors. This directly addresses the core shortcomings of existing technologies, such as ignoring terrain in multi-source data fusion, insufficient accuracy in farmland attribute identification, and fixed detection strategies. It reduces the multi-source data fusion error from 0.5m to 0.3m, increases the accuracy of pest and disease detection in small plots of steep slope farmland from 93% to over 95%, reduces the misjudgment rate of aphids in sunny organic farmland from 12% to below 8%, improves detection efficiency by 30% compared to existing technologies, and reduces the missed detection rate in small plots of steep slope farmland from 12% to below 3%. This comprehensively solves the problems of insufficient detection accuracy, low efficiency, and poor adaptability in hilly areas.

[0252] This application also discloses an integrated sky-ground farmland detection system based on artificial intelligence. For example... Figure 6 As shown, an AI-based integrated air-ground farmland detection system includes a terrain preprocessing module, a data fusion module, a farmland identification module, a dynamic switching module, and a feedback optimization module. It deeply embeds slope, aspect, and elevation terrain features into the entire detection process, enabling terrain-adaptive and accurate detection, and improving detection accuracy and efficiency.

[0253] The terrain preprocessing module acquires digital elevation model (DEM) data for the area to be inspected. Data acquisition employs two methods: large-scale areas utilize SRTM 30m DEM data, while localized complex terrain is assessed using oblique photogrammetry by a multi-rotor UAV equipped with an RGB camera and an IMU (Inertial Measurement Unit). This module extracts three core features: slope, aspect, and elevation. Slope is calculated using the cosine method and classified into three levels: gentle slopes ≤5°, sloping slopes 5°-25°, and steep slopes ≥25°. Aspect is calculated using the azimuth method and classified into two categories: sunny slopes 0°-180° and shady slopes 180°-360°. Elevation is classified into three levels: low hills ≤500m, medium hills 500m-1000m, and high hills ≥1000m. Finally, a 1m resolution terrain feature layer is generated, providing a basis for terrain decision-making in subsequent modules.

[0254] The data fusion module preprocesses and fuses multi-source remote sensing data based on terrain features. Satellite remote sensing data is processed according to slope level: gentle slopes are upscaled from 10m to 0.5m using bicubic interpolation; medium slopes are corrected for geometric distortion using local linear transformation super-resolution; steep slopes are discarded and replaced with UAV data. Atmospheric correction methods are adapted according to altitude level: dark target method for low hills, 6S model for medium hills, and MODTRAN model for high hills. UAV remote sensing data resolution is adjusted according to slope: steep slopes maintain the original 0.1m resolution; gentle and medium slopes are downsampled using mean to generate a 1m resolution layer, and illumination compensation is performed according to slope aspect. Ground sensor data is interpolated according to slope: inverse distance weighted interpolation is used for steep slopes, and ordinary kriging interpolation is used for gentle and medium slopes, all generating 1m resolution raster. The preprocessed data are assigned fusion weights according to slope, and a unified 1m resolution detection layer is generated through linear weighted fusion to reduce fusion errors.

[0255] The farmland identification module identifies farmland attributes based on detection layers and terrain features. An improved Canny edge detection algorithm is used to extract farmland boundaries, with the edge detection threshold adaptively adjusted according to slope. For high-slope, shady areas, the threshold is further increased to 0.8 to suppress false edges caused by terrain shadows. Farmland area, perimeter, and shape complexity are extracted. Morphological types are determined by a joint rule of area, slope, and shape complexity: farmland meeting any of the following conditions is considered small plots; otherwise, it is considered contiguous farmland. A logistic regression model is used to classify farmland types: integrating terrain correlation features, NDVI time-series fluctuation features, soil microbial Shannon index, and pesticide residue levels, outputting an organic farming confidence score. A confidence score ≥ 0.85 is used to classify farmland as organic; otherwise, it is considered conventional farmland.

[0256] The dynamic switching module adjusts detection parameters based on terrain features and farmland attributes. UAV and sensor parameters are set according to farmland morphology and slope: in contiguous farmland, the spacing between the UAV flight path and the ground sensor decreases as the slope increases, while the angle of oblique cross-flight increases with slope; smaller plots of farmland use smaller spacing and larger cross-flight angles than contiguous farmland with the same slope. Detection thresholds are set according to slope aspect and farmland type: pest and disease thresholds for sunny-slope organic farmland ≤3 pests / 100 plants, for shady-slope organic farmland ≤5 pests / 100 plants, for conventional farmland ≤15 pests / 100 plants on sunny slopes and ≤20 pests / 100 plants on shady slopes; simultaneously, the detection frequency is adjusted: soil microorganisms in organic farmland are detected every 2 weeks on sunny slopes and monthly on shady slopes, and once quarterly for conventional farmland.

[0257] The feedback optimization module iteratively optimizes parameters through a closed-loop mechanism. Detection errors and misclassification rates are obtained through stratified random sampling: detection errors include relative errors in pest and disease density and relative errors in soil moisture; the misclassification rate is calculated by statistically analyzing the proportion of abnormal samples. Parameters are adjusted for different combinations of slope and farmland morphology: when the soil moisture error in small plots of farmland on steep slopes is >5%, the spacing between ground sensors is reduced from 3m to 2m; when the misclassification rate for aphids is >8%, the drone's intersection angle is increased from 45° to 60°. The "Terrain-Data Weight Matrix" and "Strategy Parameter Table" are automatically updated monthly to ensure that parameters adapt to long-term terrain changes, improving detection efficiency by 30% compared to existing technologies and reducing the detection time per 10,000 mu (approximately 667 hectares) from 7 days to 5 days.

[0258] This system forms a closed-loop detection system through the collaboration of five modules, which solves the core defects of existing technologies, improves the detection accuracy by 10%-15%, reduces the missed detection rate of small plots of farmland on steep slopes to below 3%, and reduces the misjudgment rate of organic farmland to below 8%, providing reliable system support for precision agriculture in hilly and mountainous areas.

[0259] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A sky-ground integrated farmland detection method based on artificial intelligence, characterized in that, Includes the following steps: Obtain digital elevation model data of the area to be detected, extract slope features, aspect features and altitude features based on the digital elevation model data, and classify slope grade, aspect category and altitude level; Acquire multi-source remote sensing data, including satellite remote sensing data, UAV remote sensing data, and ground sensor data, and perform terrain-adaptive preprocessing on the multi-source remote sensing data based on the slope grade, slope aspect category, and altitude level; The preprocessed multi-source remote sensing data is subjected to terrain-weighted fusion to generate a detection layer with uniform resolution. The terrain-weighted fusion assigns fusion weights to remote sensing data from different sources according to the slope level. Based on the detection layer, an edge detection algorithm is used to extract farmland boundary features, wherein the edge detection threshold of the edge detection algorithm is adaptively adjusted according to the slope level; Based on the extracted farmland boundary features, the farmland area and shape complexity features are calculated, and the farmland morphology type is determined in combination with the slope level. Farmland types are classified by combining the slope aspect category, slope grade, and crop growth characteristics obtained from multi-source remote sensing data; Based on the slope grade, slope aspect category, farmland morphology type, and farmland type, the drone flight parameters and sensor deployment density are dynamically adjusted.

2. The method according to claim 1, characterized in that, The terrain adaptive preprocessing includes: For different slope grades, corresponding spatial geometric correction, resolution adjustment and raster data interpolation methods are used for satellite remote sensing data, UAV remote sensing data and ground sensor data, respectively. Appropriate illumination compensation methods are adopted for different slope aspect categories; Appropriate atmospheric correction methods are adopted for different altitude levels.

3. The method according to claim 2, characterized in that, The corresponding spatial geometric correction methods for the satellite remote sensing data include: using bicubic interpolation to improve resolution at low slope levels, using super-resolution methods to handle geometric distortion at medium slope levels, and discarding satellite remote sensing data and using only UAV remote sensing data and ground sensor data at high slope levels. The corresponding resolution adjustment methods for the UAV remote sensing data include: maintaining the original high resolution at high slope levels, and using downsampling methods to generate raster layers with uniform resolution at low and medium slope levels. The corresponding grid data interpolation methods used for the ground sensor data include: using inverse distance weighted interpolation for high slope grades, and using ordinary Kriging interpolation for low and medium slope grades.

4. The method according to claim 3, characterized in that, The terrain-weighted fusion method assigns fusion weights to remote sensing data from different sources based on slope grade, including: At the low slope level, the fusion weight of UAV remote sensing data is less than that of satellite remote sensing data but greater than that of ground sensor data; At the medium slope level, the fusion weight of satellite remote sensing data is less than that of UAV remote sensing data but greater than that of ground sensor data. At the high slope level, the fusion weight of UAV remote sensing data is greater than that of ground sensor data, while the fusion weight of satellite remote sensing data is 0.

5. The method according to claim 1, characterized in that, The extraction of farmland boundary features using an edge detection algorithm includes: An edge detection threshold is set based on the slope level, and the edge detection threshold is positively correlated with the slope level. When the slope grade is a high slope grade, the terrain shadow area is further identified according to the slope aspect category; Within the identified terrain shadow area, the edge detection threshold is increased.

6. The method according to claim 5, characterized in that, The farmland morphology type is determined based on a joint assessment of the farmland area, shape complexity, and slope grade, specifically including: If any of the following conditions are met, the land is classified as a small plot of farmland: The area of ​​the farmland is less than the first area threshold; The farmland area is between a first area threshold and a second area threshold, and the slope grade is not lower than the low slope grade, while the shape complexity is higher than the first complexity threshold. The farmland area is greater than the second area threshold, the slope grade is a high slope grade, and the shape complexity is higher than the second complexity threshold. If any of the above conditions are not met, it is determined to be contiguous farmland; Wherein, the second area threshold is greater than the first area threshold, and the first complexity threshold is higher than the second complexity threshold.

7. The method according to claim 6, characterized in that, The dynamic adjustment of UAV flight parameters and sensor deployment density includes: Based on the farmland morphology and slope grade, different drone flight path spacing, drone diagonal cross flight angles, and ground sensor spacing are set. When the farmland morphology is contiguous farmland, the spacing between the UAV flight paths and the spacing between the ground sensors decrease as the slope grade increases, while the angle of the UAV's diagonal cross flight increases as the slope grade increases. When the farmland morphology is small plots of farmland, a smaller drone flight path spacing and ground sensor spacing, as well as a larger drone oblique cross flight angle, are used compared to contiguous farmland of the same slope grade.

8. The method according to claim 7, characterized in that, It also includes the following steps: The detection error and the false positive rate are obtained. The detection error is obtained by comparing the detected value with the actual measured value. The false positive rate is obtained by statistically analyzing the proportion of abnormal false positive samples. Based on the detection error and false judgment rate, the flight parameters of the UAV and the sensor deployment density are iteratively optimized. When the detection error exceeds the corresponding error threshold under a specific combination of slope grade and farmland morphology type, or when the misjudgment rate exceeds the corresponding misjudgment rate threshold, at least one of the following should be adjusted: the UAV flight path spacing, the UAV oblique cross flight angle, and the ground sensor spacing.

9. The method according to claim 1, characterized in that, The classification of farmland types includes: Construct terrain-related features based on the slope aspect category and the slope grade; Based on the detection layer, extract the time series variation features of vegetation index; Soil microbial diversity characteristics were obtained based on ground sensor data. Characteristics of pesticide residues were obtained based on chemical analysis; The terrain correlation features are fused with the vegetation index time series variation features, soil microbial diversity features, and pesticide residue features to generate an organic farming confidence score. Based on the slope aspect category, a corresponding confidence threshold is selected. When the confidence level of organic farming reaches the selected confidence threshold, it is determined to be organic farmland; otherwise, it is determined to be conventional farmland.

10. An integrated air-ground farmland monitoring system based on artificial intelligence, characterized in that, include: The terrain preprocessing module is used to acquire digital elevation model data of the area to be detected, and extract slope features, aspect features and altitude features based on the digital elevation model data, and classify slope grade, aspect category and altitude level. The data fusion module is used to perform terrain-adaptive preprocessing on multi-source remote sensing data based on the slope grade, aspect category and altitude level, and to perform terrain-weighted fusion on the preprocessed multi-source remote sensing data to generate a detection layer with uniform resolution. The terrain-weighted fusion assigns fusion weights to remote sensing data from different sources according to the slope grade. The farmland identification module is used to extract farmland boundary features based on the detection layer using an edge detection algorithm. The edge detection threshold of the edge detection algorithm is adaptively adjusted according to the slope level. The module calculates farmland area and shape complexity features, and determines farmland morphology type based on the slope level. It also classifies farmland type based on slope aspect, slope level, and crop growth characteristics. The dynamic switching module is used to dynamically adjust the UAV flight parameters and sensor deployment density according to the slope level, slope aspect category, farmland morphology type and farmland type. The feedback optimization module is used to obtain the detection error and the false judgment rate, and to iteratively optimize the UAV flight parameters and sensor deployment density based on the detection error and the false judgment rate.

Images