Method and device for evaluating cultivated land potential by fusing multi-factor and multi-source remote sensing

Abstract

This invention discloses a method and apparatus for evaluating arable land potential by integrating multi-factor and multi-source remote sensing. The method includes identifying potential patches in a target area based on constraint factors, acquiring remote sensing images of the potential patches and extracting multi-dimensional features, outputting a land cover classification map and obtaining the average cost per acre for land reclamation based on an improved XGBoost model, and obtaining the final evaluation result of the potential patches based on a basic evaluation system comprising three dimensions and seven indicators, as well as a modified evaluation system integrating arable land, ecological, and historical factors. This method can then be applied to arable land protection and utilization practices. This invention constructs a multi-dimensional integrated arable land potential identification and evaluation system, addressing the problems of lack of planning and policy adaptability and economic cost considerations, and achieving automated identification and evaluation of arable land potential. Based on multi-source, multi-temporal, and multispectral remote sensing images and an improved XGBoost model, it improves the accuracy of land cover identification and the scientific quantification of economic costs, overcoming the bottlenecks of single data source, single-temporal analysis, and low efficiency of manual evaluation.

Description

Technical Field

[0001] This invention relates to the field of farmland potential resource assessment and remote sensing information technology, specifically to a method and apparatus for evaluating farmland potential by integrating multi-factor and multi-source remote sensing. Background Technology

[0002] Farmland protection is a basic national policy of my country. With the upgrading of the "three-in-one" protection requirements and the implementation of the policy that "all types of non-farmland can be used as a source of supplementary farmland", the need for more accurate and comprehensive farmland potential assessment is becoming increasingly urgent.

[0003] The existing technology has four major shortcomings: First, the evaluation scope is limited. Traditional methods mostly focus on unused land and do not cover non-arable land types such as agricultural land and construction land. Second, the evaluation system is one-sided. It lacks the three-dimensional integration of "nature-society-economy" and the three-dimensional correction of "arable land-ecology-history", making it difficult to adapt to planning control and ecological protection requirements. Third, the technical means are singular. Some studies use a single remote sensing data source or manual surveys, and most of them are single-temporal data applications. They do not give full play to the advantages of the temporal characteristics of multi-temporal and multispectral remote sensing, resulting in low accuracy of land feature identification, inaccurate quantification of transformation costs, and failure to effectively combine machine learning algorithms to improve evaluation efficiency. Fourth, the application of results is disconnected. The evaluation results are not closely linked with arable land protection planning and land consolidation projects, and there is a lack of closed-loop management mechanism.

[0004] Traditional single remote sensing technologies struggle to comprehensively capture the multidimensional spectral, phenological, and textural features of land cover, leading to insufficient accuracy in identifying easily confused land cover types such as lotus ponds and deep ponds, or mixed woodlands and sparse grasslands. Furthermore, they cannot refine the identification of different land cover types within the same map patch, failing to address the pain point of "map patches not matching actual uses." High-precision technologies such as hyperspectral and lidar suffer from high data acquisition and processing costs, while the traditional combination of multispectral remote sensing and manual surveys is inefficient, failing to balance the needs of large-scale coverage and refined identification. Existing evaluation systems lack a standardized framework encompassing the entire process of "data acquisition - feature extraction - classification and identification - potential evaluation - practical application," making it difficult to meet the practical needs of dynamic monitoring and large-scale promotion of farmland protection. Existing related patents often rely on a single data source, failing to incorporate phenological features based on time-series data mining and ignoring the complementary effects of texture and phenology, resulting in limited accuracy in land cover identification and difficulty in solving the "same object, different spectra" problem. High-precision solutions are too costly, while low-cost solutions lack sufficient accuracy, and the evaluation results are disconnected from farmland restoration operations. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method and apparatus for evaluating arable land potential by integrating multi-factor and multi-source remote sensing. It constructs a multi-dimensional integrated system for identifying and evaluating arable land potential, solving the problems of lack of planning and policy adaptability and economic cost considerations in existing methods. It achieves automated identification and evaluation of arable land potential for all non-arable land types across the entire region. Based on multi-source, multi-temporal, and multispectral remote sensing imagery and an improved XGBoost model, it enhances the accuracy of land cover identification and the scientific rigor of economic cost quantification, overcoming the bottlenecks of low efficiency caused by single data sources, single-temporal analysis, and manual evaluation. It establishes a closed loop of "identification-evaluation-application," resolving the problem of poor integration between potential evaluation results and arable land protection and utilization practices.

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0007] According to a first aspect of this application, a method for evaluating arable land potential that integrates multi-factor and multi-source remote sensing is provided, comprising: Identify several potential patches in the target region based on limiting factors; Acquire remote sensing images of the potential patches; Multidimensional features are extracted based on the remote sensing images, including spectral index features, phenological features, and texture features; The multidimensional features are input into the improved XGBoost model for classification and prediction, and the land cover classification results corresponding to the potential plots are output. Based on the land cover classification results, the comprehensive per-acre transformation cost corresponding to the potential plots is obtained. Guided by the long-term stable use of the transformed farmland, a basic evaluation system was constructed, which includes three dimensions: natural, social, and economic, and seven indicators: topographic slope, soil texture, plot regularity, plot contiguousness, distance to irrigation water source, farmland accessibility, and average transformation cost per mu. The basic evaluation results of potential plots were obtained. A modified evaluation system based on arable land, ecology, and historical factors is constructed to modify the basic evaluation results, and the modified score is used as the final evaluation result of the potential plot.

[0008] In some embodiments of this application, based on the foregoing scheme, the identification of several potential patches in the target region based on constraint factors includes: Using land change survey map patches as units, all non-arable land types are extracted as the initial baseline data; Using the spatial set difference algorithm, constraint factors are automatically and progressively removed from all non-arable land types, followed by geometric topology checks and fine patch cleaning to obtain several potential patches. The calculation formula is as follows:

[0009] Among them, the set of non-arable land baseline data for the entire region is as follows: The four categories of limiting factor sets are as follows: , As a planning constraint factor, As an ecological limiting factor, Terrain-limiting factor, To be a limiting factor in construction.

[0010] In some embodiments of this application, based on the foregoing scheme, the remote sensing image includes multispectral remote sensing image, multi-temporal multispectral remote sensing image, and UAV high-resolution multispectral remote sensing image, and the acquisition of the remote sensing image of the potential patch includes: Multispectral remote sensing images of the potential patches were obtained using data from the ZY-3 satellite. Multi-temporal, multispectral remote sensing images of the potential patches were acquired using data from the Sentinel-2 satellite. High-resolution multispectral remote sensing images of potential patches representing typical areas were acquired using drones.

[0011] In some embodiments of this application, based on the foregoing scheme, the extraction of multidimensional features from the remote sensing image, wherein the multidimensional features include spectral index features, phenological features, and texture features, includes: The spectral index features include the normalized vegetation index, normalized water index, improved normalized water index, and bare soil index corresponding to each potential patch. The phenological characteristics include the annual mean and standard deviation of the normalized vegetation index, the annual mean and standard deviation of the improved normalized water index, and the 12-period index values ​​for each potential patch. The texture features include contrast, homogeneity, entropy, and correlation for each potential patch.

[0012] In some embodiments of this application, based on the foregoing scheme, the training process of the improved XGBoost model includes a sample construction and feature extraction stage, a data cleaning and preprocessing stage, and a model training and optimization stage. In the sample construction and feature extraction stage, a multi-dimensional feature matrix of each training region is constructed using pixels as the basic unit. Core spectral indices, phenological features and texture features are comprehensively introduced to generate the corresponding vector ROI sample for each training region. In the data cleaning and preprocessing stage, image pixel values ​​and category labels are extracted based on vector ROI samples, and invalid pixels are removed by an automatic noise detection algorithm to obtain several training samples. During the model training and optimization phase, a stratified sampling strategy is adopted to divide a number of training samples into a training set and a validation set for training. At the same time, a class weight adjustment mechanism is introduced to give higher penalty weights to the classes with fewer samples in the loss function until the objective function converges or the preset maximum number of iterations is reached, thus obtaining the trained improved XGBoost model.

[0013] In some embodiments of this application, based on the aforementioned scheme, the extraction of multidimensional features from remote sensing images includes UAV high-resolution multispectral remote sensing images. If the UAV high-resolution multispectral remote sensing images are partially missing, a data validity grouping strategy is introduced. This involves dynamic grouping based on the available band combination status of pixels, and separate training for each group. Specifically, this includes: The training samples are divided into several subsets based on the pixels of high-resolution multispectral remote sensing images from UAVs; Detect whether each band of each pixel is effective; If each band of the pixel is valid, a first dedicated XGBoost sub-model is established and trained so that the established first dedicated XGBoost sub-model can make full use of the texture features extracted from the high-resolution multispectral remote sensing image of the UAV. If the pixel has locally missing bands, then label mapping is performed on the locally missing bands, and a second dedicated XGBoost sub-model is established. The global category ID is converted into a local continuous ID to participate in the training of the second dedicated XGBoost sub-model. After prediction, it is restored to a unified category code to avoid the category anomaly problem in multi-class training. If the pixel has a complex band missing pattern, a global fallback model based on XGBoost is constructed, and the ability of the global fallback model based on XGBoost to adaptively split the missing values ​​is used to supplement the prediction. During the training process of the improved XGBoost model, the initial training sample set is divided into a training set and a validation set. The improved XGBoost model is trained using the training set, and key hyperparameters, including the number of trees, maximum depth, learning rate, and training method, are jointly optimized using the validation set. This ensures the model's expressive power while suppressing overfitting and improving the training efficiency of large-scale data. The land cover classification results include regular vegetation land, perennial green and messy vegetation land, non-perennial green and messy vegetation land, winter dry ponds, aquatic crop ponds, scattered stable water bodies, contiguous stable water bodies of rivers and lakes, bare soil, buildings, and hardened ground.

[0014] In some embodiments of this application, based on the foregoing scheme, obtaining the comprehensive per-acre transformation cost corresponding to the potential patch based on the land feature classification results includes: Assign a corresponding average price per acre to each type of land feature; Extract a raster image of the potential patch, the raster image comprising a plurality of graticles; If the potential plot is a single land use plot where all grid cells belong to the same land use type, the average unit price per mu corresponding to the land use type shall be used as the comprehensive average renovation cost per mu of the potential plot. If the potential plot is a mixed land use plot where all grids are mixed with different land types, count the number of grids for each type of land feature within the potential plot. Based on the area of ​​a single grid and the number of grids, obtain the actual land area occupied by each type of land feature. Using the actual land area occupied by each type of land feature as a weight, and combining it with the corresponding average price per mu, calculate the weighted sum to obtain the comprehensive average renovation cost per mu of the potential plot.

[0015] In some embodiments of this application, based on the aforementioned scheme, the basic evaluation system, guided by the long-term stable use of the transformed arable land, is constructed, encompassing three dimensions (natural, social, and economic) and seven indicators (topography slope, soil texture, plot regularity, plot contiguousness, irrigation water source distance, arable land accessibility, and per-acre transformation cost). This system yields basic evaluation results for potential land parcels, including: The weights of each indicator in the basic evaluation system are determined by expert scoring. The weighted sum of the various indicators is used as the basic evaluation result of the potential plot.

[0016] In some embodiments of this application, based on the foregoing scheme, the step of constructing a modified evaluation system based on arable land, ecology, and historical factors to modify the basic evaluation results, and using the modified score as the final evaluation result of the potential plot, includes: Using a product correction model and GIS spatial overlay technology, for the basic evaluation results, if the potential plot is historical farmland or located in a planned agricultural industrial zone, a positive incentive coefficient is set; if the potential plot is an ecological protection control zone or a historical protection and management zone, a negative penalty coefficient is set respectively. The positive incentive coefficient or the negative penalty coefficient is multiplied by the basic evaluation results to output the final evaluation results.

[0017] According to a second aspect of this application, a farmland potential assessment device integrating multi-factor and multi-source remote sensing is provided, comprising: The potential patch identification module is used to identify several potential patches in a target area based on constraint factors; The data acquisition module is used to acquire remote sensing images of the potential patches; The feature extraction module is used to extract multidimensional features based on the remote sensing image, including spectral index features, phenological features, and texture features. The land feature classification and cost accounting module is used to input the multidimensional features into the improved XGBoost model, perform classification prediction, output the land feature classification results corresponding to the potential plots, and obtain the comprehensive per-acre transformation cost corresponding to the potential plots based on the land feature classification results. The basic evaluation module is used to construct a basic evaluation system with three dimensions of natural, social and economic factors and seven indicators, including topographic slope, soil texture, plot regularity, plot contiguousness, distance to irrigation water source, farmland accessibility and average cost per mu, in order to obtain the basic evaluation results of potential plots. The final evaluation module is used to construct a modified evaluation system based on arable land, ecology, and historical factors to modify the basic evaluation results, and the modified score is used as the final evaluation result of the potential plot.

[0018] According to a third aspect of this application, a computer-readable storage medium is provided that stores a computer program thereon, the computer program including executable instructions that, when executed by a processor, implement the method described above.

[0019] According to a fourth aspect of this application, an electronic device is provided, comprising: One or more processors; A memory for storing executable instructions of the processor, which, when executed by the one or more processors, cause the one or more processors to implement the method described above.

[0020] The beneficial effects of this application are as follows: The method and apparatus for evaluating arable land potential that integrates multi-factor and multi-source remote sensing provided in this application include: 1) Standardizing the identification of theoretical potential patches within the current region based on limiting factors such as planning, ecology, topography, and construction; 2) Integrating data from the ZY-3 satellite, Sentinel-2 satellite, and UAVs using multi-source remote sensing multispectral image analysis to construct a spatiotemporal data cube with "multi-temporal-multi-scale-multispectral" characteristics; and using the XGBoost (Extreme Gradient Boosting) ensemble learning algorithm to classify actual land use status and estimate per-acre transformation costs based on multi-dimensional features such as spectral, phenological, and textural characteristics of land features; 3) Constructing a comprehensive potential evaluation model that includes "basic evaluation + corrective evaluation," building a basic evaluation model from three dimensions: natural, social, and economic, and forming a corrective evaluation model by combining planning and policy factors such as arable land, ecology, and history; and 4) Establishing a technical framework for arable land restoration and potential identification and evaluation that includes "theoretical potential identification - current land type identification - comprehensive potential evaluation - intelligent decision support."

[0021] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and do not limit this application. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and are intended to explain the invention, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for evaluating arable land potential that integrates multi-factor and multi-source remote sensing according to the present invention. Figure 2 This is a schematic diagram of the satellite multispectral remote sensing imagery of the present invention; Figure 3 This is a schematic diagram of the multi-temporal multispectral remote sensing imagery of the present invention; Figure 4 This is a schematic diagram of the high-resolution multispectral remote sensing image of the UAV of the present invention; Figure 5 This is a schematic diagram illustrating the spectral index characteristics of the present invention; Figure 6 This is a schematic diagram illustrating the phenological characteristics of the present invention; Figure 7 This is a schematic diagram of the texture features of the present invention; Figure 8 This is a schematic diagram of the land cover classification results in a specific embodiment of the present invention; Figure 9(a) is a schematic diagram of the terrain slope scoring of the basic evaluation system in a specific embodiment of the present invention; Figure 9(b) is a schematic diagram of the soil texture scoring of the basic evaluation system in a specific embodiment of the present invention; Figure 9(c) is a schematic diagram of the plot regularity scoring of the basic evaluation system in a specific embodiment of the present invention; Figure 9(d) is a schematic diagram of the land contiguousness scoring of the basic evaluation system in a specific embodiment of the present invention; Figure 9(e) is a schematic diagram of the irrigation water source distance scoring of the basic evaluation system in a specific embodiment of the present invention; Figure 9(f) is a schematic diagram of the farmland accessibility scoring system in a specific embodiment of the present invention; Figure 9(g) is a schematic diagram of the per-acre renovation cost scoring of the basic evaluation system in a specific embodiment of the present invention; Figure 10 This is a schematic diagram of the final evaluation result in a specific embodiment of the present invention; Figure 11 This is a schematic diagram of a farmland potential evaluation device that integrates multi-factor and multi-source remote sensing according to the present invention. Figure 12 This is a schematic diagram of an electronic device. Detailed Implementation

[0023] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0024] It should be understood that the terms "comprising" and other similar expressions in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, apparatus, or device that includes a series of steps or units and is not limited to the listed steps or units. Furthermore, "first" and "second" are used to distinguish different objects and are not intended to describe a specific order.

[0025] According to the first aspect of this application, Figure 1 As shown in the figure, this embodiment provides a method for evaluating arable land potential that integrates multi-factor and multi-source remote sensing, including: Step S101: Identify several potential patches in the target region based on limiting factors.

[0026] In some embodiments of this example, the method for identifying several potential patches in a target region based on constraint factors is as follows: Basic baseline data: Using land change survey patches within the target area as units, extract all non-arable land types as the initial baseline data.

[0027] Restrictive Screening: A four-level elimination system is constructed, encompassing "planning control - ecological protection - topographic conditions - construction implementation," clearly defining 12 restrictive factors and elimination criteria, such as urban development boundaries and slopes exceeding 15°. Based on GIS spatial overlay analysis technology and the "negative list" elimination logic, a four-level elimination system is constructed, specifically as follows: First, multi-source spatial data are unified to the CGCS2000 coordinate system. Using the entire non-arable land parcels from the land change survey as the calculation base, the spatial set difference algorithm is used to automatically and progressively remove 12 types of limiting factors, including urban development boundaries, ecological protection red lines, slopes above 15°, and historical and cultural protection areas. After geometric topology checks and fine-grained parcel cleaning, potential parcels that meet the rigid constraints of "multi-plan integration" are accurately output, achieving the scientific locking of the spatial range of potential resources. The calculation formula is as follows:

[0028] Among them, the set of non-arable land baseline data for the entire region is as follows: The four categories of limiting factor sets are as follows: , As a planning constraint factor, As an ecological limiting factor, Terrain-limiting factor, To be a limiting factor in construction.

[0029] In this embodiment, planning restriction factors refer to urban construction areas and ecological protection areas that are clearly defined in various planning documents that have already come into effect.

[0030] Ecological limiting factors refer to the limiting areas determined by the sensitivity, vulnerability, or importance of the natural ecosystem itself.

[0031] Topographic limiting factors refer to the physical limitations imposed by landforms (slope, elevation, landforms, etc.) on arable land development.

[0032] Construction limiting factors refer to the restrictions imposed on project approval and actual construction.

[0033] Data standardization: All vector data (planning map, red line map, current status map) are projected into the CGCS2000 coordinate system.

[0034] The four-level elimination process is as follows: Level 1: Planning and control elimination, Level 2: Ecological protection elimination, Level 3: Topographic conditions elimination, and Level 4: Construction and implementation elimination.

[0035] Thus, by following the fully automated process of "base data input - initial screening of land types - planning elimination - project elimination - output of theoretical potential", the potential plots can be accurately located.

[0036] Step S102: Obtain remote sensing images of the potential patches, such as... Figures 2-4 As shown, the remote sensing images include multispectral remote sensing images, multi-temporal multispectral remote sensing images, and UAV high-resolution multispectral remote sensing images.

[0037] In this embodiment, the multispectral remote sensing images were acquired by the ZY-3 satellite (summer and winter seasons, panchromatic resolution 2.1 meters, multispectral resolution 5.8 meters), and the multitemporal multispectral remote sensing images were acquired by the Sentinel-2 satellite (covering 12 periods throughout the year, spatial resolution 10 or 20 meters).

[0038] In this embodiment, high-resolution multispectral remote sensing images of potential patches characterized as typical areas are acquired by UAVs. Typical areas are determined based on factors such as current terrain features, current land use patches, and farmland protection planning zones. Typical areas must cover at least 30% of the total number of potential patches.

[0039] In this way, multispectral remote sensing images, multitemporal multispectral remote sensing images, and UAV high-resolution multispectral remote sensing images are used to construct a three-dimensional data matrix of "sky-air-ground", making full use of the advantages of low cost and wide coverage of multispectral data.

[0040] Step S103: Extract multidimensional features based on the remote sensing image, including spectral index features, phenological features and texture features.

[0041] In some implementations of this embodiment, based on the spectral index characteristics of spectral indices including NDVI (Normalized Difference Vegetation Index), NDWI (Normalized Difference Water Index), MNDWI (Modified Normalized Difference Water Index), and BSI (Bare Soil Index), combined with phenological characteristics (annual mean and standard deviation of normalized vegetation index, annual mean and standard deviation of modified normalized water index, and 12-period index values) and GLCM texture characteristics (four types of indicators: contrast, homogeneity, entropy, and correlation), a multi-dimensional feature set is formed to mine the temporal variation patterns of multi-temporal data.

[0042] In this embodiment, multispectral remote sensing images are acquired using the ZY-3 and Sentinel-2 satellites, and then the normalized vegetation index, normalized water index, improved normalized water index, and bare soil index corresponding to each potential patch are obtained.

[0043] Normalized Difference Vegetation Index (NDVI) is used to distinguish between vegetation and non-vegetation. The calculation formula is as follows:

[0044] in, Normalized Difference Vegetation Index (NDVI) It is in the near-infrared band. It is in the red light band.

[0045] The normalized water index is used to extract water body information, and the calculation formula is as follows:

[0046] in, The normalized water index, It is in the green light band.

[0047] The improved normalized water index is used to eliminate building disturbances and accurately identify ponds and pits. The calculation formula is as follows:

[0048] in, To improve the normalized water quality index, It is in the shortwave infrared band.

[0049] The bare soil index is used to distinguish between bare land and construction land. The calculation formula is as follows:

[0050] in, The bare soil index, It is in the blue light band.

[0051] In some implementations of this embodiment, data from the Sentinel-2A / B dual satellites are used, covering all 12 phases of the year, to ensure complete capture of seasonal changes in land cover (germination, growth, dormancy / maturity) throughout the year, supplemented by winter and summer data from the ZY-3 satellite to provide medium-to-high resolution phenological details.

[0052] In this embodiment, multi-temporal multispectral remote sensing images are preprocessed to obtain the aforementioned spectral index features. Statistical analysis is performed on the 12-period index images to extract core phenological features. The phenological features include the annual mean and standard deviation of the normalized vegetation index, the annual mean and standard deviation of the improved normalized water index, and the 12-period index values ​​corresponding to each potential patch. The annual mean is used to reflect the average state of ground cover throughout the year, and the annual variance is used to reflect the degree of phenological fluctuation. For a single potential patch, its 12-period index values ​​are extracted to construct a time-index value time series curve.

[0053] Phenological characteristics are mainly used to distinguish between perennial green vegetation, winter dryland, aquatic crop ponds, and stable water bodies.

[0054] Non-annual vegetation cover: NDVI shows a single peak (spring rise - summer peak - autumn fall - winter low), with large variance and large variation.

[0055] Aquatic crop ponds (such as lotus ponds): NDVI is high in summer (when lotus leaves cover the pond) and NDWI is high in autumn (when the water surface is exposed during the lotus formation period), showing a bimodal or opposite phase.

[0056] Stable water bodies: NDWI is high and stable throughout the year with small variance.

[0057] Bare soil / buildings: Low index value, small fluctuation throughout the year, and variance close to 0.

[0058] In this embodiment, texture features are extracted using the ZY-3 satellite. The texture features include the contrast, homogeneity, entropy, and correlation of each potential patch.

[0059] The clearer the texture of a feature (the sharper its edges), the higher its contrast value (e.g., in a built-up area). The formula for calculating contrast is:

[0060] The more uniform the texture of land features, the higher the homogeneity value (e.g., flat farmland, water surface). The formula for calculating homogeneity is:

[0061] The more disordered the terrain features, the higher the entropy (e.g., in a messy woodland). The formula for calculating entropy is:

[0062] The formula for calculating correlation is:

[0063] in, These are the row and column indices of the matrix (representing gray levels). The normalized gray-level co-occurrence matrix element values ​​are The probability of occurrence They are respectively In the line, number Mean along the column direction , , They are respectively In the line, number Standard deviation in column direction , This reflects the directional extension of the texture (such as field ridges and roads).

[0064] Thus, based on the complementary advantages of ZY-3's high spatial resolution and Sentinel-2's high temporal resolution, this embodiment constructs a three-in-one high-dimensional feature extraction system of "spectrum-temporal-texture" to provide high-discrimination input for machine learning models. First, core spectral indices were calculated based on multi-source images resampled to 2m resolution. Vegetation was identified using the Normalized Difference Vegetation Index (NDVI>0.2), and the Improved Normalized Difference Water Index (MNDWI) was calculated using shortwave infrared bands to suppress building noise and accurately extract water bodies. The Bare Soil Index (BSI) was used to highlight bare soil and construction land. Second, "temporal statistical dimensionality reduction" was performed on 12 time-series data from the Sentinel-2 satellite throughout the year. The annual mean and standard deviation of the Normalized Difference Vegetation Index (NDVI), the Water Index (NDWI), and the Improved Normalized Difference Water Index (MNDWI) were calculated. By exploring the "high mean-low variance" characteristics of stable water bodies, the "high variance-winter abrupt change" characteristics of seasonal dry ponds, and the "anti-phase alternation" pattern of aquatic crops, the problem of distinguishing between perennial vegetation and seasonal cultivation from single-period images was solved. Finally, a 7-dimensional model was constructed using ZY-3 multispectral remote sensing images. The 7-window Gray-Level Co-occurrence Matrix (GLCM) quantifies and extracts four types of indicators: contrast (edge ​​sharpness), homogeneity (uniformity), entropy (disorder), and correlation (linear direction). This effectively solves the problem of confusion between "same spectrum, different features" in regular and disordered vegetation. Finally, spectral index features, phenological features, and texture features are superimposed to form a full-element feature matrix. For example... Figure 5 As shown, this is a schematic diagram of the spectral index characteristics of the present invention. Figure 6 As shown, this is a schematic diagram of the phenological characteristics of the present invention. Figure 7 The diagram shown is a schematic representation of the texture features of this invention.

[0065] Step S104: Input the multidimensional features into the improved XGBoost model for classification prediction, output the land cover classification results corresponding to the potential plots, and obtain the comprehensive per-acre transformation cost corresponding to the potential plots based on the land cover classification results.

[0066] In some implementations of this embodiment, XGBoost is an ensemble learning model based on Gradient Boosting Decision Tree (GBDT). By progressively stacking multiple Classification and Regression Trees (CART) weak learners, it continuously minimizes classification errors, thereby achieving high-precision representation of complex nonlinear ground features.

[0067] At the model mechanism level, XGBoost completes training by minimizing an objective function with a regularization term:

[0068] in, This is a multi-class log loss function used to measure the difference between the predicted result and the true label. For the first The true label of each sample For the first The predicted output for each sample, The total number of samples; This is a regularization term used to constrain the complexity of the tree structure to prevent overfitting. For the first A tree, The number of decision trees is specified. The algorithm approximates the loss function using a second-order Taylor expansion, and improves the model's convergence speed and generalization ability by simultaneously utilizing first-order and second-order gradients to accurately guide the node splitting direction. For multi-class problems, `objective="multi:softprob"` is set, and the Softmax function is used to output the probability vectors of samples belonging to each class, with the class corresponding to the highest probability being used as the final prediction result.

[0069] In terms of specific implementation, the intelligent classification process is divided into three stages: "feature engineering—dynamic grouping training—global route inference." First, in the feature engineering stage, a multi-dimensional feature matrix is ​​constructed using pixels as the basic unit, comprehensively incorporating core spectral indices, phenological features, and texture features based on the gray-level co-occurrence matrix (GLCM) to enhance the ability to distinguish easily confused land types. Subsequently, image pixel values ​​and category labels are extracted based on vector ROI (Region of Interest) samples, and invalid pixels are removed using an automatic noise detection algorithm to ensure the quality of training samples. After the sample construction is completed, a stratified sampling strategy is adopted to divide the training set and validation set in an 8:2 ratio. Simultaneously, a class weight adjustment mechanism is introduced, assigning higher penalty weights to classes with fewer samples in the loss function to alleviate the bias caused by uneven sample distribution on model learning.

[0070] In some embodiments of this example, high-resolution multispectral remote sensing images (0.1m level) from UAVs are supplemented. The high-resolution multispectral remote sensing images from UAVs are represented as two-dimensional matrix data and used to verify phenological details in typical areas (such as lotus ponds and concentrated farmland areas) to ensure that the temporal characteristics are consistent with the actual growth rhythm of ground objects. The high-resolution multispectral remote sensing images from UAVs acquire normalized vegetation index, normalized water index, and bare soil index, which are mainly used for sample selection and judgment, and training of the improved XGBoost model.

[0071] To address the issue of missing spectral channels in some pixels due to drone imagery covering only local areas, an availability key strategy is introduced. Instead of directly training a single model, dynamic grouping is performed based on the available band combinations of pixels. Specifically, for each pixel, the availability of each band is checked, generating a corresponding availability key. Based on this key, the training sample set is divided into multiple subsets, each corresponding to a specific sensor data combination state (e.g., "drone + satellite full elements" or "satellite bands only"). For grouped areas with sufficient sample size, dedicated XGBoost sub-models are trained to fully utilize the high-resolution texture information provided by drone imagery. Simultaneously, label mapping is performed on potentially missing categories in local areas, converting global category IDs into local continuous IDs for training. After prediction, the IDs are restored to a unified category encoding, avoiding category anomalies in multi-class training. For scenarios with insufficient sample size or complex combinations, an additional global fallback model is constructed, utilizing XGBoost's adaptive splitting capability for missing values ​​to supplement predictions.

[0072] During model training, key hyperparameters are jointly optimized using a validation set, including the number of trees (n_estimators=400), maximum depth (max_depth=8), learning rate (learning_rate=0.05), and training method (tree_method="hist"), to suppress overfitting and improve the efficiency of training large-scale data while ensuring the model's expressive power.

[0073] In the global inference stage, the image data is segmented and dynamic routing inference is performed pixel-by-pixel. First, the validity of each pixel to be predicted is calculated, and a corresponding dedicated sub-model is automatically matched according to its band validity status; if no matching model exists, the global fallback model is called to complete the prediction. The local prediction results output by the sub-model are then restored to a unified land cover category code through an inverse mapping mechanism. Finally, the results are stitched together to generate a global 2m resolution land cover classification map and a corresponding uncertainty probability map, realizing adaptive and high-precision intelligent land cover classification under multi-source remote sensing conditions.

[0074] In one specific embodiment, such as Figure 8 As shown, the improved XGBoost model is based on the Gradient Boosting Decision Tree (GBDT) principle. It minimizes classification error by integrating multiple CART weak learners and using a "greedy strategy + gradient descent" to optimize the loss function. The specific implementation process first constructs a multi-dimensional feature matrix containing spectral index features, phenological time series features, and texture features. 373 typical samples are selected and divided into training and validation sets in an 8:2 ratio, and a class weight adjustment strategy is used to balance the sample distribution. To address the issue of inconsistent spatial coverage from multiple data sources (UAVs and satellites), a "data validity grouping (Availability Key)" strategy is introduced. Dedicated XGBoost sub-models are trained based on the effective band combinations of pixels to achieve adaptive processing of scenes with missing bands. Finally, key parameters are fine-tuned using the validation set to complete pixel-level prediction and remapping of the entire image, outputting a high-precision land cover classification map. The land cover classification map completes land cover classification (3 major categories and 9 subcategories, including terrestrial agricultural land and water-related agricultural land) at a 2m × 2m resolution, with a classification accuracy of over 90%.

[0075] In this embodiment, the sample source consists of 373 groups of land cover blocks of interest, evenly distributed across a certain region. These blocks cover the same study area and the same phase window, fully encompassing nine core land cover categories. Spatially, they cover five typical scenarios, including concentrated farmland areas and lake-field remediation areas. The data preprocessing and feature extraction processes are completely unified, eliminating differences in dimensions and coordinate systems. The entire process strictly adheres to industry standards such as the "Remote Sensing Image Interpretation Specification TD / T1015–2023". The data is randomly stratified at an 8:2 ratio to ensure that the proportion of land cover categories in the training and validation sets is consistent with the overall sample distribution. Classification and identification are performed on three major categories and nine subcategories of core land cover: well-vegetated land, aquatic crop ponds, and bare soil. The validation set is completely independent and does not participate in model training or parameter tuning. The samples include three dimensions of features: spectral (NDVI, NDWI, etc.), phenological (annual mean / variance of indices), and texture (GLCM four-category indicators), forming a standardized sample-feature matrix. For scenarios with imbalanced land cover samples, a class weight adjustment strategy is adopted to balance the contribution of each category to the model loss and avoid model bias.

[0076] Based on the engineering volume of "comprehensive per-acre transformation cost", three major categories and nine subcategories of training results are defined. The land feature classification results include terrestrial agricultural land, water-related agricultural land, and terrestrial non-agricultural land. The terrestrial agricultural land includes land with regular vegetation, land with perennial green and messy vegetation, and land with non-perennial green and messy vegetation. The water-related agricultural land includes winter dry ponds, aquatic crop ponds, scattered stable water bodies, and contiguous stable water bodies of rivers and lakes. The terrestrial non-agricultural land includes bare soil, buildings, and hardened ground.

[0077] Tests showed that the model achieved over 90% precision, recall, and F1 score for single-class land cover, with an overall classification accuracy of 90%. Compared to traditional random forest classification methods, this scheme improved the accuracy of identifying easily confused aquatic crop ponds and deep ponds, as well as perennial and non-perennial vegetation cover by over 15%, and increased image processing efficiency by 10 times. It effectively solved the problem of identifying similar land cover with different spectra and dissimilar land cover with similar spectra in densely water-networked areas of Wuhan, providing accurate land cover data support for the assessment of farmland restoration potential.

[0078] In some embodiments of this example, obtaining the average cost per mu (unit of land area) for the potential plot based on the land feature classification results specifically involves assigning an average price per mu to each land feature and establishing a raster map of the potential plot, the raster map comprising several grids. The nine sub-categories of raster classification results are unified with the farmland restoration potential vector plots to the same plane coordinate system, maintaining a 2m×2m raster resolution. Based on this, precise cost anchoring at the raster grid level is achieved. First, the area conversion of a single 2m×2m raster is completed, clarifying that the area of ​​a single grid is approximately 0.006 mu. Then, based on the pre-established land type-cost mapping table, a GIS raster calculation tool is used to assign an average cost per mu value corresponding to the nine sub-categories of land features to each raster, generating an average cost per mu raster map with the same range and resolution as the classification raster. This ensures that each pixel has a clear unit area cost attribute, fully leveraging the technical advantages of pixel-level refined land feature identification in the project, and accurately capturing cost differences between different land types within the same potential plot. Subsequently, comprehensive cost-per-acre aggregation and accounting were carried out at the potential plot level. Potential plots within the target area were used as the legally mandated basic accounting units. Differentiated accounting methods were adopted for potential plots composed of different land types. For potential plots with all rasters belonging to the same land type, the average unit price per acre corresponding to that land type was directly used as the comprehensive cost-per-acre transformation cost of the potential plot, and the comprehensive cost-per-acre transformation cost of the potential plot was calculated simultaneously. For mixed land type plots, which are common in the project and contain multiple heterogeneous land types, the area-weighted average method was used to calculate the comprehensive cost-per-acre. This involved first counting the number of rasters for each of the nine sub-types of land features within the potential plot, converting them into actual land area, and then using the area of ​​each land type as a weight, combined with the corresponding average unit price per acre, to calculate the comprehensive cost-per-acre transformation cost of the potential plot. Simultaneously, each potential plot was assigned three core attributes: comprehensive cost-per-acre transformation cost, the area ratio of each land type, and the total transformation cost of the plot, generating a vector database of potential plots with cost attributes, matching the project's GIS format database requirements.

[0079] After completing the basic cost accounting, the results will be deeply integrated with the farmland potential suitability evaluation system. The calculated comprehensive per-acre transformation cost will be seamlessly integrated into the project's three-dimensional suitability evaluation system, which combines "natural ecology + social production + economic cost." The accounting results will clearly define only construction costs, excluding additional expenses such as demolition costs, major soil remediation costs, and project management fees. This clear definition of cost accounting boundaries ensures the compliance and practicality of the results. The final accounting results can directly support the farmland restoration potential evaluation research needs of the main project.

[0080] Step S105: The basic evaluation system includes topographic slope, soil texture, plot regularity, plot contiguousness, distance to irrigation water source, ease of cultivation, and comprehensive per-acre transformation cost. Following the "three-in-one" protection concept of arable land quantity, quality, and ecology, a basic evaluation system is constructed with the long-term stable use of transformed arable land as its guiding principle, resulting in the basic evaluation results of the classified potential plots. For example... Figure 8 The diagram shown is a schematic representation of the basic evaluation system of this invention.

[0081] In some implementations of this embodiment, as shown in Figures 9(a)-9(g), a three-dimensional indicator system of "nature-society-economy" is constructed, which includes seven core indicators: topographic slope, soil texture, plot regularity, plot contiguity, distance to irrigation water source, accessibility of cultivation, and comprehensive reclamation cost per acre. The weights are determined by expert scoring (natural ecology 0.391, social production 0.326, economic cost 0.283). The seven core indicators include topographic slope, soil texture, plot contiguity, distance to irrigation water source, plot regularity, accessibility of cultivation, and comprehensive reclamation cost per acre.

[0082] In this embodiment, the threshold for the comprehensive per-mu (unit of land area) renovation cost is determined based on the following criteria: the per-mu renovation cost is graded according to the natural discontinuity method, combined with actual project experience. The cost grading reflects the range of the project's investment affordability; costs below 15,000 yuan / mu are considered low-cost, high-quality projects, while costs above 25,000 yuan / mu are less economical and require careful consideration. The final thresholds are: less than 15,000 yuan / mu (100 points), 15,000-20,000 yuan / mu (80 points), 20,000-22,000 yuan / mu (60 points), 22,000-25,000 yuan / mu (40 points), and greater than 25,000 yuan / mu (20 points).

[0083] In this embodiment, the terrain slope is a quantitative indicator that measures the degree of tilt of potential map patches, and the method for obtaining it is as follows: Acquire digital elevation model data with a spatial resolution of 8 meters; Using the slope analysis tool of the Geographic Information System (ArcGIS), the slope value (degrees) of each raster cell is calculated based on the Digital Elevation Model (DEM). The slope raster map is overlaid with the vector boundary of the potential patch. Through the regional statistical function, the average slope value within the polygon range of each potential patch is calculated as the representative slope value of the potential patch. The terrain slope score is determined based on the representative slope value of the potential patch.

[0084] The thresholds are mainly based on the classification requirements for terrain slope in the "Grading of Cultivated Land Quality" (GB / T33469-2016), and are set in combination with the suitability for agricultural mechanization operations; ≤2° is plains highly suitable for mechanized farming (score 100); 2°-6° is gentle slope suitable for farming (score 80); 6°-15° is medium slope requiring soil and water conservation measures (score 60); >15° is steep slope unsuitable for cultivation (score 40). In one specific embodiment, a potential patch is calculated to have an average slope of 5.2° for the pixels within its range, which falls within the "2°-6°" range, so the terrain slope score is 80 points.

[0085] In this embodiment, soil texture refers to the combination of mineral particles of different diameters in the soil, which can generally be divided into types such as sandy soil, loam, and clay. The method for obtaining this texture is as follows: Obtain the 2021 farmland fertility evaluation data results from the Municipal Bureau of Agriculture and Rural Affairs; The "topsoil texture" field from the arable land fertility assessment data is overlaid with the potential land parcels to be evaluated. If a potential land parcel contains multiple texture types, the dominant texture is determined by its area proportion. If no readily available data is available, the k-nearest neighbor (KNN) algorithm is used to estimate the soil texture of the arable land potential resource parcels, and then the soil types within the potential land parcels are statistically analyzed. The soil texture score is determined based on the soil types within the potential land parcels.

[0086] The thresholds were determined based on the agrophysical properties (water and fertilizer retention, tilth, and aeration) of different soil textures. Medium loam has the most balanced properties and is the ideal soil for cultivation; sandy soil and clay soil have significant shortcomings in water retention and aeration, respectively. The thresholds were determined based on soil texture characteristics as follows: medium loam (100 points), sandy loam, light loam (80 points), heavy loam (60 points), clay (40 points), and sandy soil (20 points).

[0087] In one specific embodiment, the soil survey attribute of a potential plot is displayed as "light loam". According to the standard, this indicator directly scores 80 points.

[0088] In this embodiment, the contiguousness of land parcels reflects the degree of contiguousness between potential land parcels and existing cultivated land, permanent basic farmland, etc., and is obtained by the following method: Based on the 2024 land change survey, existing cultivated land of more than 5 mu (0101 paddy fields, 0102 irrigated land, 0103 dry land) was selected as existing cultivated land. Using GIS spatial analysis tools, the minimum distance between the boundary of each potential plot to be evaluated and the boundaries of all surrounding existing arable land potential plots is calculated. The minimum distance reflects the degree of spatial isolation between the potential plot and existing arable land resources, and the contiguousness score of the plot is determined based on the minimum distance.

[0089] The thresholds aim to encourage contiguous development to achieve economies of scale and infrastructure sharing. The closer to existing farmland, the easier it is to integrate into existing irrigation and drainage systems and management areas, reducing subsequent operating costs. Generally, areas within 80 meters are considered easily integrated contiguous areas. The final thresholds were determined as follows: less than 20 meters (100 points), 20-40 meters (80 points), 40-80 meters (60 points), 80-160 meters (40 points), and greater than 160 meters (20 points).

[0090] In one specific embodiment, the minimum distance between the boundary of a potential plot to be evaluated and the boundary of the nearest existing arable land potential plot is 35 meters, which falls within the "20-40 meters" range, so the score is 80 points.

[0091] In this embodiment, the distance to the irrigation water source reflects the quality of agricultural irrigation water intake conditions, and the method for obtaining this information is as follows: Based on the 2024 land change survey, the water areas and water conservancy facilities (1101 river water surface, 1102 lake water surface, 1103 reservoir water surface, 1104 pond water surface, 1107 ditch) were selected and the potential cultivated land plots were removed as the current water source areas. Using GIS spatial analysis tools (nearest neighbor analysis), the Euclidean distance (straight-line distance) from the boundary of each potential patch to be evaluated to the nearest effective water source (area feature) is calculated, and the score of irrigation water source distance is determined based on the Euclidean distance from the boundary of each potential patch to be evaluated to the nearest effective water source.

[0092] The thresholds are mainly based on empirical values ​​of construction costs and irrigation reliability corresponding to different water intake distances in farmland irrigation design. The closer the distance, the higher the convenience and reliability of water intake, and the lower the irrigation cost. The final thresholds are determined as follows: less than 100 meters (100 points), 100-200 meters (80 points), 200-300 meters (60 points), 300-400 meters (40 points), and greater than 400 meters (20 points).

[0093] In one specific embodiment, it was calculated that the straight-line distance from the boundary of a potential patch to be evaluated to the nearest irrigation canal was 150 meters, falling into the "100-200 meters" range, so the score was 80 points.

[0094] In this embodiment, the regularity of the plot reflects the regularity of the shape of the potential plot. Farmland with a relatively regular shape is more conducive to mechanized farming.

[0095] The vector boundary data (area features) of the potential patches to be evaluated must be consistent with the coordinate system and area calculation unit.

[0096] The shape index is used for measurement, and the calculation formula is as follows:

[0097] in, The index of arable land shape. The perimeter (in meters) of the cultivated area. This represents the area of ​​a cultivated land parcel (in square meters). The minimum value for the cultivated land shape index is 1 (representing a perfect circle or square). The higher the cultivated land shape index value, the more irregular and complex the shape. The score for the regularity of the land parcel is determined based on the shape index.

[0098] Threshold determination criteria: Thresholds refer to the general standards for field planning in land consolidation projects. Potential plots with a shape index close to 1 (1.0-1.1) are extremely regular and conducive to cultivation management (score 100); potential plots with a shape index close to 1.1 (1.1-1.2) are relatively regular (score 80); potential plots with a shape index close to 1.2 (1.2-1.3) are relatively irregular (score 60); potential plots with a shape index close to 1.3 (1.3-1.4) are relatively fragmented (score 40); when the index is greater than 1.4, the potential plots are fragmented, which seriously restricts mechanical efficiency and field management (score 20).

[0099] In one specific embodiment, a square potential plot with an area of ​​1 hectare (10,000 square meters) and a perimeter of 400 meters has a shape index of 1.0, which falls within the "1.0-1.1" range, and therefore scores 100 points.

[0100] In this embodiment, the ease of cultivation reflects the ease of cultivating a field in agricultural production.

[0101] Based on the 2024 land use change survey, urban and rural road land (1004 urban and rural roads and 1006 rural roads) were selected as the current road network.

[0102] Using a spatial analysis method similar to that used for "irrigation conditions," the Euclidean distance from each potential plot to the nearest accessible road was calculated. This Euclidean distance represents the shortest straight-line distance from the potential plot to the road network, signifying the ease of machinery access to the field. A score for ease of cultivation was determined based on this Euclidean distance.

[0103] Threshold determination criteria: The thresholds comprehensively consider the operational efficiency and fuel economy of agricultural machinery. Excessive distance from roads significantly increases the machinery's idle time and costs, affecting tillage efficiency. Generally, a tillage radius within 500 meters is considered to have good economic efficiency. The final thresholds are determined as follows: less than 200 meters (100 points), 200-500 meters (80 points), 500-1000 meters (60 points), 1000-2000 meters (40 points), and greater than 2000 meters (20 points).

[0104] In one specific embodiment, the straight-line distance from the boundary of a potential patch to be evaluated to the nearest field road is 300 meters, falling into the "200-500 meters" range, so the score is 80 points.

[0105] Step S106: Construct a modified evaluation system based on arable land, ecology, and historical factors to modify the basic evaluation results, and use the modified score as the final evaluation result of the potential plot.

[0106] In some implementations of this embodiment, the evaluation system is modified by establishing a three-dimensional correction factor of "farmland protection - ecological protection - historical protection" and setting positive (historical farmland range, etc.) and negative (wildlife habitat, etc.) correction coefficients to achieve policy adaptation of evaluation results. Based on the basic evaluation score (weighted from 7 core indicators), a three-dimensional dynamic correction system of "farmland protection - ecological protection - historical protection" was constructed. Using a product correction model and GIS spatial overlay technology, a positive incentive coefficient of 1.2 to 1.4 was set for historical farmland or farmland located in planned agricultural industrial zones to prioritize support for farmland restoration. Negative penalty coefficients of 0.5 and 0 to 0.2 were set for ecological protection control zones and historical protection management zones, respectively, to strengthen rigid constraints. Finally, the basic score was corrected through automated calculation and a grade reclassification was performed. The positive incentive coefficient or the negative penalty coefficient was multiplied by the basic evaluation result to output the final evaluation result. The final evaluation result is a graded evaluation result of farmland restoration potential that takes into account physical feasibility, ecological security and policy adaptability.

[0107] Thus, using the "single-factor quantification + weighted summation + correction coefficient adjustment" model, a comprehensive score of 0-100 is output, generally divided into three potential levels: "good-average-poor".

[0108] In some implementations of this embodiment, a visualization platform can be used to integrate multi-source data and evaluation results, enabling multi-scale statistics, multi-dimensional queries, and multi-condition retrieval functions. In some implementations of this embodiment, the results can be transformed: a connection mechanism with farmland protection planning and land consolidation projects can be established, and priority ranking and implementation suggestions can be provided.

[0109] In summary, this embodiment breaks through the limitations of traditional evaluation scope, realizing the full-area potential identification of various non-arable land types, which aligns with the "full coverage" policy requirements. It constructs a multi-dimensional integrated evaluation system, solving the problems of existing methods lacking policy adaptability and ecological constraint considerations. Furthermore, based on multispectral remote sensing imagery and combined with machine learning algorithms, it improves the accuracy of land cover identification and the scientific nature of cost quantification, breaking through the bottlenecks of single data source, single-temporal analysis, and low efficiency of manual evaluation. It also establishes a closed-loop mechanism of "identification-evaluation-application" to solve the problem of poor connection between evaluation results and actual work.

[0110] In this embodiment, a target area (521.79 square kilometers) was selected as the empirical area, and more than 30 basic data were collected, including the 2024 land change survey and the results of the delineation of the "three zones and three lines". 1. Data preprocessing: The coordinate system is unified to CGCS2000, the vector data format is Shapefile, the raster data format is GeoTIFF, and outliers and duplicate data are removed.

[0111] 2. Implementation of Comprehensive Potential Identification Initial baseline: Extract a number of non-arable land parcels of a certain city; Level 4 elimination: Overlaying restrictive factors such as urban development boundaries (eliminating 323,700 mu), ecological protection red lines (eliminating 90,200 mu), and slopes with a gradient of more than 15° (eliminating 21,800 mu), the final output is a theoretical potential plot of 59,400 mu.

[0112] 3. Implementation of multi-temporal and multispectral remote sensing interpretation Data acquisition: acquire multispectral data from the ZY-3 satellite from February to July 2025; acquire 12 phases of multispectral data from Sentinel-2 from November 2024 to October 2025; complete a precision flight of the UAV in October 2025 (covering a typical area of ​​50 square kilometers). 4. Feature extraction: The annual mean and variance of spectral indices are calculated using the GEE platform, GLCM texture features are extracted based on ArcGIS, and phenological variation patterns are mined by combining multi-temporal data. 5. Model Training: Select 373 interest samples, divide them into training and validation sets in an 8:2 ratio, set the learning rate to 0.001 and the number of iterations to 50, and complete the training and land cover classification of the XGBoost machine learning model.

[0113] 6. Comprehensive evaluation implementation Single-factor quantification: ArcGIS spatial analysis tools are used to calculate scores for indicators such as terrain slope and distance to irrigation water sources, and the average cost per acre for renovation is calculated based on remote sensing classification results; Basic evaluation: The basic evaluation score is obtained by summing the weights according to their respective weights, and the rationality of the weight allocation is optimized through machine learning algorithms; Correction and adjustment: Factors such as historical cultivated land area (correction coefficient 1.2-1.4) and Baer's Pochard habitat (correction coefficient 0.5) are superimposed to calculate the comprehensive score and classify the potential level.

[0114] 7. Verification of Results Application Visualization platform construction: Develop an information platform that includes 28 data layers and 15 query conditions; Empirical verification: The evaluation results are 92% consistent with the field survey results, with 18,500 mu of good potential plots (56-100 points), which are precisely matched with the regional farmland protection planning needs.

[0115] According to the second aspect of this application, such as Figure 11 As shown in the figure, this embodiment provides a farmland potential evaluation device that integrates multi-factor and multi-source remote sensing, comprising: The potential patch identification module is used to identify several potential patches in a target area based on constraint factors; The data acquisition module is used to acquire remote sensing images of the potential patches, including multispectral remote sensing images, multitemporal multispectral remote sensing images, and UAV high-resolution multispectral remote sensing images. The feature extraction module is used to extract multidimensional features based on the remote sensing image, including spectral index features, phenological features, and texture features. The land feature classification and cost accounting module is used to input the multidimensional features into the improved XGBoost model, perform classification prediction, output the land feature classification results corresponding to the potential plots, and obtain the comprehensive per-acre transformation cost corresponding to the potential plots based on the land feature classification results. The basic evaluation module is used to construct a basic evaluation system with three dimensions of natural, social and economic factors and seven indicators, including topographic slope, soil texture, plot regularity, plot contiguousness, distance to irrigation water source, farmland accessibility and average cost per mu, in order to obtain the basic evaluation results of potential plots. The final evaluation module is used to construct a modified evaluation system based on arable land, ecology, and historical factors to modify the basic evaluation results, and the modified score is used as the final evaluation result of the potential plot.

[0116] Specifically, this embodiment corresponds one-to-one with the above method embodiments. The functions of each module have been described in detail in the corresponding method embodiments, so they will not be repeated here.

[0117] According to a third aspect of this application, this embodiment provides a computer-readable storage medium having a computer program stored thereon, the computer program including executable instructions that, when executed by a processor, implement the method described above.

[0118] The present invention can implement all or part of the processes in the above methods, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0119] According to the fourth aspect of this application, such as Figure 12 As shown, an electronic device is provided, comprising: One or more processors; Memory is used to store executable instructions for the processor, which, when executed by one or more processors, cause one or more processors to implement the methods described above.

[0120] Electronic devices are manifested in the form of general-purpose computing devices. Components of an electronic device may include, but are not limited to: at least one processor, at least one memory, and a bus connecting different device components (including memory and processor).

[0121] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the computer device, connecting all parts of the computer device through various interfaces and lines.

[0122] Memory can be used to store computer programs and / or modules. The processor implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can primarily include a program storage area and a data storage area. The program storage area stores application programs required for operating the device and at least one function (e.g., sound playback, image playback, etc.); the data storage area stores data created based on the use of the mobile phone (e.g., audio data, video data, etc.). Furthermore, memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, SmartMedia Cards (SMC), Secure Digital (SD) cards, Flash Cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0123] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, servers, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and memory) containing computer-usable program code.

[0124] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), servers, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0125] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0126] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0127] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0128] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0129] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for evaluating arable land potential that integrates multi-factor and multi-source remote sensing, characterized in that, include: Identify several potential patches in the target region based on limiting factors; Acquire remote sensing images of the potential patches; Multidimensional features are extracted based on the remote sensing images, including spectral index features, phenological features, and texture features; The multidimensional features are input into the improved XGBoost model for classification and prediction, and the land cover classification results corresponding to the potential plots are output. Based on the land cover classification results, the comprehensive per-acre transformation cost corresponding to the potential plots is obtained. Guided by the long-term stable use of the transformed farmland, a basic evaluation system was constructed, which includes three dimensions: natural, social, and economic, and seven indicators: topographic slope, soil texture, plot regularity, plot contiguousness, distance to irrigation water source, farmland accessibility, and average transformation cost per mu. The basic evaluation results of potential plots were obtained. A modified evaluation system based on arable land, ecology, and historical factors is constructed to modify the basic evaluation results, and the modified score is used as the final evaluation result of the potential plot.

2. The method according to claim 1, characterized in that, The method for identifying several potential patches in the target region based on constraint factors includes: Using land change survey map patches as units, all non-arable land types are extracted as the initial baseline data; Using the spatial set difference algorithm, constraint factors are automatically and progressively removed from all non-arable land types, followed by geometric topology checks and fine patch cleaning to obtain several potential patches. The calculation formula is as follows: Among them, the set of non-arable land baseline data for the entire region is as follows: The four categories of limiting factor sets are as follows: , As a planning constraint factor, As an ecological limiting factor, Terrain-limiting factor, To be a limiting factor in construction.

3. The method according to claim 1, characterized in that, The remote sensing images include multispectral remote sensing images, multi-temporal multispectral remote sensing images, and UAV high-resolution multispectral remote sensing images. The remote sensing images used to acquire the potential patches include: Multispectral remote sensing images of the potential patches were obtained using data from the ZY-3 satellite. Multi-temporal, multispectral remote sensing images of the potential patches were acquired using data from the Sentinel-2 satellite. High-resolution multispectral remote sensing images of potential patches representing typical areas were acquired using drones.

4. The method according to claim 1, characterized in that, The extraction of multidimensional features based on the remote sensing image, including spectral index features, phenological features, and texture features, includes: The spectral index features include the normalized vegetation index, normalized water index, improved normalized water index, and bare soil index corresponding to each potential patch. The phenological characteristics include the annual mean and standard deviation of the normalized vegetation index, the annual mean and standard deviation of the improved normalized water index, and the 12-period index values ​​for each potential patch. The texture features include contrast, homogeneity, entropy, and correlation for each potential patch.

5. The method according to claim 1, characterized in that: The training process of the improved XGBoost model includes a sample construction and feature extraction stage, a data cleaning and preprocessing stage, and a model training and optimization stage. In the sample construction and feature extraction stage, a multi-dimensional feature matrix of each training region is constructed using pixels as the basic unit. Core spectral indices, phenological features and texture features are comprehensively introduced to generate the corresponding vector ROI sample for each training region. In the data cleaning and preprocessing stage, image pixel values ​​and category labels are extracted based on vector ROI samples, and invalid pixels are removed by an automatic noise detection algorithm to obtain several training samples. During the model training and optimization phase, a stratified sampling strategy is adopted to divide a number of training samples into a training set and a validation set for training. At the same time, a class weight adjustment mechanism is introduced to give higher penalty weights to the classes with fewer samples in the loss function until the objective function converges or the preset maximum number of iterations is reached, thus obtaining the trained improved XGBoost model.

6. The method according to claim 1, characterized in that: The extraction of multidimensional features from remote sensing images includes high-resolution multispectral remote sensing images from UAVs. If parts of the high-resolution multispectral remote sensing images from UAVs are missing, a data validity grouping strategy is introduced. This involves dynamic grouping based on the available band combinations of pixels, with separate training for each group. Specifically, this includes: The training samples are divided into several subsets based on the pixels of high-resolution multispectral remote sensing images from UAVs; Detect whether each band of each pixel is effective; If each band of the pixel is valid, a first dedicated XGBoost sub-model is established and trained so that the established first dedicated XGBoost sub-model can make full use of the texture features extracted from the high-resolution multispectral remote sensing image of the UAV. If the pixel has locally missing bands, then label mapping is performed on the locally missing bands, and a second dedicated XGBoost sub-model is established. The global category ID is converted into a local continuous ID to participate in the training of the second dedicated XGBoost sub-model. After prediction, it is restored to a unified category code to avoid the category anomaly problem in multi-class training. If the pixel has a complex band missing pattern, a global fallback model based on XGBoost is constructed, and the ability of the global fallback model based on XGBoost to adaptively split the missing values ​​is used to supplement the prediction. During the training process of the improved XGBoost model, the initial training sample set is divided into a training set and a validation set. The improved XGBoost model is trained using the training set, and key hyperparameters, including the number of trees, maximum depth, learning rate, and training method, are jointly optimized using the validation set. This ensures the model's expressive power while suppressing overfitting and improving the training efficiency of large-scale data. The land cover classification results include regular vegetation land, perennial green and messy vegetation land, non-perennial green and messy vegetation land, winter dry ponds, aquatic crop ponds, scattered stable water bodies, contiguous stable water bodies of rivers and lakes, bare soil, buildings, and hardened ground.

7. The method according to claim 1, characterized in that, The step of obtaining the comprehensive per-acre transformation cost corresponding to the potential patch based on the land cover classification results includes: Assign a corresponding average price per acre to each type of land feature; Extract a raster image of the potential patch, the raster image comprising a plurality of graticles; If the potential plot is a single land use plot where all grid cells belong to the same land use type, the average unit price per mu corresponding to the land use type shall be used as the comprehensive average renovation cost per mu of the potential plot. If the potential plot is a mixed land use plot where all grids are mixed with different land types, count the number of grids for each type of land feature within the potential plot. Based on the area of ​​a single grid and the number of grids, obtain the actual land area occupied by each type of land feature. Using the actual land area occupied by each type of land feature as a weight, and combining it with the corresponding average price per mu, calculate the weighted sum to obtain the comprehensive average renovation cost per mu of the potential plot.

8. The method according to claim 1, characterized in that, The aforementioned basic evaluation system, guided by the long-term stable use of transformed arable land, constructs a framework encompassing three dimensions (natural, social, and economic) and seven indicators (topography slope, soil texture, plot regularity, plot contiguousness, irrigation water source distance, arable land accessibility, and per-acre transformation cost). This framework yields basic evaluation results for potential land parcels, including: The weights of each indicator in the basic evaluation system are determined by expert scoring. The weighted sum of the various indicators is used as the basic evaluation result of the potential plot.

9. The method according to claim 1, characterized in that, The construction of a revised evaluation system based on arable land, ecology, and historical factors to revise the basic evaluation results, and the revised score as the final evaluation result of the potential plot, includes: Using a product correction model and GIS spatial overlay technology, for the basic evaluation results, if the potential plot is historical farmland or located in a planned agricultural industrial zone, a positive incentive coefficient is set; if the potential plot is an ecological protection control zone or a historical protection and management zone, a negative penalty coefficient is set respectively. The positive incentive coefficient or the negative penalty coefficient is multiplied by the basic evaluation results to output the final evaluation results.

10. A device for evaluating arable land potential that integrates multi-factor and multi-source remote sensing, characterized in that, include: The potential patch identification module is used to identify several potential patches in a target area based on constraint factors; The data acquisition module is used to acquire remote sensing images of the potential patches; The feature extraction module is used to extract multidimensional features based on the remote sensing image, including spectral index features, phenological features, and texture features. The land feature classification and cost accounting module is used to input the multidimensional features into the improved XGBoost model, perform classification prediction, output the land feature classification results corresponding to the potential plots, and obtain the comprehensive per-acre transformation cost corresponding to the potential plots based on the land feature classification results. The basic evaluation module is used to construct a basic evaluation system with three dimensions of natural, social and economic factors and seven indicators, including topographic slope, soil texture, plot regularity, plot contiguousness, distance to irrigation water source, farmland accessibility and average cost per mu, in order to obtain the basic evaluation results of potential plots. The final evaluation module is used to construct a modified evaluation system based on arable land, ecology, and historical factors to modify the basic evaluation results, and the modified score is used as the final evaluation result of the potential plot.

Images