A fine-grained optimization method, medium and system for cultivated land protection potential

Abstract

The present invention provides a fine-grained method, medium, and system for optimizing cultivated land conservation potential, belonging to the field of cultivated land data processing technology. The method first acquires multi-source remote sensing, soil, and climate data, extracting a basic feature matrix through multi-dimensional decomposition. A pyramid neural network model is then used to perform multi-scale analysis of soil quality characteristics, and adaptively integrates them through a feature integration module. A contribution optimization equation system is then constructed, comprehensively considering four aspects: benchmark evaluation, spatial correlation, temporal evolution, and weight balance. This allows for accurate evaluation of cultivated land conservation potential and generates an optimized conservation zoning scheme. The pyramid neural network model overcomes the problem of insufficient feature extraction in traditional methods through multi-scale feature extraction and adaptive integration. Furthermore, the contribution optimization equation system, through multi-dimensional information fusion and dynamic optimization, addresses the difficulty in accurately evaluating cultivated land conservation potential in existing technologies.

Description

Technical Field

[0001] The present invention belongs to the technical field of cultivated land data processing, and in particular relates to a fine-grained cultivated land protection potential optimization method, medium and system. Background Art

[0002] Cultivated land protection is a crucial foundation for ensuring national food security, and the assessment of cultivated land conservation potential is a key component in guiding this work. Traditional assessments of cultivated land conservation potential rely primarily on foundational work such as remote sensing image interpretation, soil sampling and analysis, and climate data statistics. These assessments are conducted by establishing an evaluation index system and employing methods such as the Analytic Hierarchy Process (AHP) and the Fuzzy Comprehensive Evaluation Method (Fuzzy Comprehensive Evaluation Method). These methods primarily build evaluation models based on expert experience. Land use information is obtained through visual interpretation or automatic classification of remote sensing images. Combined with soil quality data obtained through field surveys and climate data from meteorological stations, weights for various evaluation indicators are calculated to ultimately produce the cultivated land conservation potential assessment results.

[0003] However, traditional evaluation methods have the following problems: First, the construction of the evaluation index system and the determination of weights mainly rely on expert experience, which is highly subjective and difficult to objectively reflect the differences in cultivated land protection potential in different regions; second, the data acquisition method is single, making it difficult to fully utilize the rich information contained in multi-source remote sensing data, soil monitoring data and climate observation data; third, the evaluation method lacks a systematic analysis of the internal characteristics and external environment of cultivated land patches, and cannot achieve a comprehensive evaluation at multiple scales and dimensions; finally, the evaluation results lack a dynamic updating mechanism, making it difficult to timely reflect the spatiotemporal changes in cultivated land protection potential.

[0004] Against the backdrop of the rapid development of big data and artificial intelligence technologies, how to fully utilize multi-source data to establish an objective and accurate evaluation method for farmland conservation potential and achieve fine-grained evaluation of farmland conservation potential has become a pressing technical challenge. This means that existing technologies struggle to accurately evaluate farmland conservation potential. Summary of the Invention

[0005] In view of this, the present invention provides a fine-grained farmland protection potential optimization method, medium and system, which can solve the technical problem in the existing technology that fine-grained accurate evaluation of farmland protection potential cannot be achieved.

[0006] The present invention is implemented as follows: In a first aspect, the present invention provides a fine-grained method for optimizing cultivated land protection potential, which obtains remote sensing image data, soil monitoring data and climate observation data of a target area as a basic data set; performs multi-dimensional decomposition on the basic data set; extracts cultivated land patches based on a land use type distribution matrix; analyzes the soil quality characteristic matrix using a soil quality characteristic pyramid neural network model; evaluates the cultivated land patch condition index based on a climate suitability matrix; constructs a cultivated land protection potential evaluation model; obtains an initial score using a neural network model training; establishes a cultivated land protection contribution matrix; optimizes and adjusts to obtain an optimized score; divides high, medium and low protection potential areas; performs spatial connectivity analysis, identifies priority protection areas and outputs them.

[0007] The remote sensing image data includes multispectral remote sensing image data, hyperspectral remote sensing image data, radar remote sensing image data, surface cover classification data and normalized vegetation index data.

[0008] The soil monitoring data includes soil organic matter content data, soil bulk density data, soil pH value data, soil nutrient content data, soil water content data, soil texture data and soil erosion degree data.

[0009] The climate observation data include annual average temperature data, annual precipitation data, sunshine hours data, cumulative temperature data, drought index data, precipitation seasonal distribution data and extreme weather event data.

[0010] The basic data set is decomposed into multiple dimensions to obtain a land use type distribution matrix, a soil quality characteristic matrix, and a climate suitability matrix.

[0011] The method comprises extracting cultivated land patches based on the land use type distribution matrix, and calculating the area parameters, boundary length parameters and fragmentation degree parameters of the cultivated land patches.

[0012] The soil quality feature pyramid neural network model includes five pyramid layers that are executed sequentially and a feature synthesis module.

[0013] Among them, the farmland protection contribution matrix is ​​obtained by solving the contribution optimization equation group, which includes a contribution benchmark equation, a spatial correlation equation, a time series evolution equation and a weight balance equation. The contribution benchmark equation is used to calculate the basic contribution of each evaluation indicator, the spatial correlation equation is used to evaluate the spatial coupling effect between evaluation indicators, and the time series evolution equation is used to analyze the temporal variation characteristics of the contribution of evaluation indicators.

[0014] Among them, the pyramid layers include: the first pyramid layer is used to analyze the soil quality characteristics inside the cultivated land patches; the second pyramid layer is used to analyze the combination of soil quality characteristics inside the cultivated land patches and the adjacent areas of the cultivated land patches; the third pyramid layer is used to analyze the combination of soil quality characteristics inside the cultivated land patches, the adjacent areas of the cultivated land patches and the buffer zones around the cultivated land patches; the fourth pyramid layer is used to analyze the combination of soil quality characteristics inside the cultivated land patches, the adjacent areas of the cultivated land patches, the buffer zones around the cultivated land patches and the cultivated land patch watershed units; the fifth pyramid layer is used to analyze the combination of soil quality characteristics inside the cultivated land patches, the adjacent areas of the cultivated land patches, the buffer zones around the cultivated land patches, the cultivated land patch watershed units and the overall region.

[0015] The feature integration module is configured to determine whether the soil quality characteristic index output by the pyramid layer is greater than a soil quality characteristic credibility threshold. If the soil quality characteristic index is greater than the soil quality characteristic credibility threshold, the soil quality characteristic index output by the current pyramid layer is used as the final soil quality characteristic index. If the soil quality characteristic index is less than the soil quality characteristic credibility threshold, the feature integration module proceeds to the next pyramid layer analysis or uses the output of the last pyramid layer as the final soil quality characteristic index.

[0016] The light and heat resource condition index, water condition index, and climate stress factor index of the cultivated land patches are evaluated based on the climate suitability matrix. The cultivated land protection contribution matrix is ​​obtained by solving a contribution optimization equation system, which includes a contribution benchmark equation, a spatial correlation equation, a temporal evolution equation, and a weight balance equation.

[0017] The inputs to the contribution benchmark equation include the normalized values ​​of evaluation indicators obtained from the farmland protection potential evaluation model, the indicator significance coefficients obtained from historical monitoring data, the historical protection effect coefficients obtained from historical protection records, and the regional development constraint coefficients obtained from regional development plans. The inputs to the spatial association equation include the basic contribution matrix of evaluation indicators, the spatial adjacency matrix obtained from the geographic information system, the terrain gradient coefficient obtained from terrain data, the land use intensity index obtained from remote sensing image data, and the landscape connectivity index obtained from landscape pattern analysis. The inputs to the temporal evolution equation include historical farmland protection effectiveness data obtained from the farmland protection monitoring system, the climate change trend coefficient obtained from climate observation data, the socioeconomic development index obtained from the statistical department, and the policy regulation intensity index obtained from policy documents.

[0018] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, are used to execute the above-mentioned fine-grained farmland protection potential optimization method.

[0019] The third aspect of the present invention provides a fine-grained farmland protection potential optimization system, which includes the above-mentioned computer-readable storage medium. The system is any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0020] Compared to existing technologies, the present invention provides a fine-grained method, medium, and system for optimizing farmland conservation potential. This method achieves precise evaluation of farmland conservation potential by constructing a soil quality feature pyramid neural network model and a set of optimization equations for farmland conservation contribution. This method fully utilizes multi-source remote sensing data, soil monitoring data, and climate observation data, establishing an adaptive and dynamic evaluation framework through multi-dimensional decomposition and feature extraction.

[0021] The method of this invention overcomes the problems of traditional technologies. First, by constructing a soil quality feature pyramid neural network model, it achieves automatic extraction of multi-scale features from local to global scales, avoiding the influence of subjective experience. Second, by establishing a set of optimization equations for the contribution of cultivated land protection, it achieves effective fusion and comprehensive evaluation of multi-source data, improving the objectivity and reliability of the evaluation results. Finally, by introducing temporal evolution equations and spatial correlation equations, it achieves dynamic updating and spatial optimization of the evaluation results.

[0022] The present invention solves the technical problem of accurate evaluation of cultivated land protection potential, which is mainly reflected in the following aspects: First, adaptive feature extraction is realized through the pyramid neural network model, which improves the accuracy of feature expression; second, scientific fusion of multi-dimensional information is realized through the contribution optimization equation group, which enhances the reliability of the evaluation results; third, real-time updating of the evaluation results is achieved through the dynamic optimization mechanism, which ensures the timeliness of the evaluation results and solves the problem of difficulty in achieving accurate evaluation of cultivated land protection potential in the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] like Figure 1 FIG. 1 is a flow chart of a fine-grained farmland protection potential optimization method provided by the first aspect of the present invention. The method comprises the following steps:

[0026] S01. Acquire remote sensing image data, soil monitoring data, and climate observation data of the target area as basic data sets, wherein the remote sensing image data includes multispectral remote sensing image data, hyperspectral remote sensing image data, radar remote sensing image data, surface cover classification data, and normalized difference vegetation index data; the soil monitoring data includes soil organic matter content data, soil bulk density data, soil pH value data, soil nutrient content data, soil moisture content data, soil texture data, and soil erosion degree data; and the climate observation data includes annual average temperature data, annual precipitation data, sunshine hours data, cumulative temperature data, drought index data, precipitation seasonal distribution data, and extreme weather event data;

[0027] S02. Perform multi-dimensional decomposition on the basic data set to obtain a land use type distribution matrix, a soil quality characteristic matrix, and a climate suitability matrix;

[0028] S03. Extracting cultivated land patches based on the land use type distribution matrix, and calculating area parameters, boundary length parameters, and fragmentation degree parameters of the cultivated land patches;

[0029] S04. Analyze the soil quality feature matrix using a soil quality feature pyramid neural network model to obtain a soil fertility index, a soil physical structure index, and a soil chemical property index of the cultivated land patch, wherein the soil quality feature pyramid neural network model includes five pyramid layers and a feature synthesis module that are executed sequentially:

[0030] The first pyramid layer is used to analyze the soil quality characteristics inside the cultivated land patch;

[0031] The second pyramid layer is used to analyze the combination of soil quality characteristics inside the cultivated land patch and the area adjacent to the cultivated land patch;

[0032] The third pyramid layer is used to analyze the combination of soil quality characteristics within the cultivated land patch, the adjacent area of ​​the cultivated land patch and the buffer zone around the cultivated land patch;

[0033] The fourth pyramid layer is used to analyze the soil quality characteristic combination inside the cultivated land patch, the adjacent area of ​​the cultivated land patch, the buffer zone around the cultivated land patch and the watershed unit of the cultivated land patch;

[0034] The fifth pyramid layer is used to analyze the combination of soil quality characteristics inside the cultivated land patch, the adjacent area of ​​the cultivated land patch, the buffer zone around the cultivated land patch, the watershed unit of the cultivated land patch and the overall soil quality characteristics of the region;

[0035] The feature integration module is used to determine whether the soil quality characteristic index output by the pyramid layer is greater than the soil quality characteristic credibility threshold. When it is greater than the soil quality characteristic credibility threshold, the soil quality characteristic index output by the current pyramid layer is used as the final soil quality characteristic index. When it is less than the soil quality characteristic credibility threshold, the next layer of pyramid analysis is continued or the output of the last layer of pyramid is used as the final soil quality characteristic index.

[0036] S05. Evaluate the light and heat resource condition index, water condition index, and climate stress factor index of the cultivated land patch according to the climate suitability matrix;

[0037] S06. Constructing a farmland protection potential evaluation model, inputting the area parameter, boundary length parameter, fragmentation degree parameter, soil fertility index, soil physical structure index, soil chemical property index, light and heat resource condition index, water condition index, and climate stress factor index of the farmland patch into the farmland protection potential evaluation model;

[0038] S07. Using a neural network model to train the farmland protection potential evaluation model to obtain an initial farmland protection potential score;

[0039] S08. Establishing a cultivated land protection contribution matrix based on the initial cultivated land protection potential score, wherein the cultivated land protection contribution matrix is ​​obtained by solving a contribution optimization equation group, wherein the contribution optimization equation group includes a contribution benchmark equation, a spatial correlation equation, a temporal evolution equation, and a weight balance equation;

[0040] S09. Optimizing and adjusting the initial score of the cultivated land protection potential according to the cultivated land protection contribution matrix to obtain an optimized score of the cultivated land protection potential;

[0041] S10, dividing the cultivated land protection potential optimization score into high protection potential areas, medium protection potential areas, and low protection potential areas, and generating a cultivated land protection zoning plan matrix;

[0042] S11, performing spatial connectivity analysis on the farmland protection zoning scheme matrix, identifying and outputting priority protection areas;

[0043] The contribution benchmark equation is used to calculate the basic contribution of each evaluation indicator. The input includes the normalized value of the evaluation indicator obtained from the cultivated land protection potential evaluation model, the indicator significance coefficient obtained from historical monitoring data, the historical protection effect coefficient obtained from historical protection records, and the regional development constraint coefficient obtained from the regional development plan. The output is the basic contribution matrix of the evaluation indicator.

[0044] The spatial correlation equation is used to evaluate the spatial coupling effect between evaluation indicators. The input includes the basic contribution matrix of the evaluation indicators, the spatial adjacency matrix obtained from the geographic information system, the terrain gradient coefficient obtained from the terrain data, the land use intensity index obtained from the remote sensing image data, and the landscape connectivity index obtained from the landscape pattern analysis. The output is the spatial correlation contribution matrix.

[0045] The time series evolution equation is used to analyze the temporal variation characteristics of the contribution of the evaluation indicators. The input includes the farmland protection effectiveness data obtained from the farmland protection monitoring system over the years, the climate change trend coefficient obtained from the climate observation data, the socio-economic development index obtained from the statistical department, and the policy regulation intensity index obtained from the policy documents. The output is the time series evolution contribution matrix.

[0046] The weight balance equation is used to comprehensively balance the final contribution of each evaluation indicator. The input includes the basic contribution matrix of the evaluation indicators, the spatial correlation contribution matrix, the temporal evolution contribution matrix, and the regional differentiation correction coefficient obtained from the regional differentiation assessment. The output is the final farmland protection contribution matrix.

[0047] The specific implementation of the above steps is described in detail below. The specific implementation of step S01 involves acquiring and preprocessing remote sensing image data, soil monitoring data, and climate observation data. The specific implementation first involves selecting the data source. Remote sensing image data utilizes multi-source remote sensing data from Landsat 8 and Sentinel-2, among others. The raw images are processed using preprocessing methods such as geometric correction, radiometric correction, and atmospheric correction. Soil monitoring data is obtained through field sampling and indoor analysis. Sampling points are arranged using a stratified sampling method, with 1 to 3 sampling points per square kilometer and a sampling depth of 0 to 20 cm. Climate observation data is derived from field data at meteorological stations and meteorological satellite remote sensing data. Next, data quality control is performed, using the triple standard deviation method to eliminate outliers. Missing data is supplemented using spatiotemporal interpolation using Kriging. Finally, data normalization is performed, using the minimum-maximum normalization method to normalize all data to a range of 0 to 1. The purpose of this step is to provide high-quality basic data support for subsequent analysis.

[0048] The specific implementation method of step S02 is to use the principal component analysis method to reduce the dimension and extract features of the basic data set. First, the basic data set is subjected to correlation analysis, the correlation coefficient matrix between the indicators is calculated, and the redundant indicators with correlation coefficients greater than 0.95 are eliminated. Secondly, the principal component analysis method is used to extract the main eigenvectors, and the principal components with cumulative contribution rates greater than 85% are selected as the feature matrix. Then, based on the load coefficients of each principal component, the original data is reconstructed into a land use type distribution matrix, a soil quality characteristic matrix, and a climate suitability matrix. Finally, the reconstructed feature matrix is ​​quality verified, and the data quality is ensured by calculating the reconstruction error, and the reconstruction error threshold is set to 0.15. The purpose of this step is to reduce the data dimension and extract key feature information.

[0049] The specific implementation method of step S03 is to extract and analyze the characteristics of cultivated land patches based on the land use type distribution matrix. First, the regional growing algorithm is used to extract cultivated land patches, the growth threshold is set to 0.8, and the 8-neighborhood connectivity criterion is used to merge regions. Secondly, the area parameters of each cultivated land patch are calculated using the grid counting method and converted into the actual area according to the spatial resolution. Then, the boundary tracking algorithm is used to calculate the boundary length parameters of the cultivated land patches, and the Freeman chain code method is used to extract the boundary characteristics. Finally, the fragmentation degree parameters are calculated using the landscape index method, including the patch density index, the boundary density index and the fractal dimension index, and the fragmentation index is comprehensively calculated by the weighted average method, and the weight coefficient is determined by the hierarchical analysis method. The purpose of this step is to quantitatively describe the spatial characteristics of cultivated land patches.

[0050] The specific implementation method of step S04 is to construct and apply a pyramid neural network model of soil quality characteristics. First, a pyramid network structure is designed, and a convolutional neural network is used to extract features in each layer. The convolution kernel size increases layer by layer from 3×3 to 11×11, and the pooling layer uses maximum pooling. Secondly, a feature integration module is designed, and an attention mechanism is used to adaptively fuse features at different levels. The attention weight is optimized by a back-propagation algorithm. Then, the soil quality feature credibility threshold is determined and set to 0.85 through a cross-validation method. Finally, the network model is trained, and the network parameters are optimized using the stochastic gradient descent method. The initial value of the learning rate is set to 0.01, and the cosine annealing strategy is used for dynamic adjustment. The number of training rounds is set to 1000. The purpose of this step is to achieve multi-scale extraction and comprehensive analysis of soil quality characteristics.

[0051] The specific implementation method of step S05 is to evaluate the climate condition index based on the climate suitability matrix. First, the light and heat resource condition index is calculated. The fuzzy comprehensive evaluation method is used to fuzzify the indicators such as the annual average temperature, sunshine hours and cumulative temperature, and the membership function adopts the Gaussian function. Secondly, the water condition index is calculated. The drought index method is used to combine the annual precipitation and the seasonal distribution characteristics of precipitation to construct a water balance model. Then, the climate stress factor index is calculated, and the extreme value analysis method is used to evaluate the impact of extreme weather events on cultivated land. Finally, the entropy weight method is used to determine the weight of each indicator and construct a comprehensive evaluation model. The purpose of this step is to evaluate the suitability of climate conditions for cultivated land.

[0052] The specific implementation method of steps S06 and S07 is to construct an evaluation model for the potential for cultivated land protection and train it. First, a deep neural network is used to construct an evaluation model. The network structure includes an input layer, multiple hidden layers, and an output layer. The hidden layer uses a residual connection structure to improve model performance. Secondly, batch normalization technology is used to standardize the input features, and data enhancement technology is used to expand the training samples. Then, the cross entropy loss function is used to measure the model performance, and the Adam optimizer is used for parameter optimization with a learning rate of 0.001. Finally, the early stopping method is used to prevent overfitting, and training is stopped when the accuracy of the validation set no longer improves. The purpose of this step is to establish a prediction model for the potential for cultivated land protection.

[0053] The specific implementation method of steps S08 and S09 is to construct and solve the contribution optimization equation group. First, a contribution benchmark equation is established, and the basic contribution of each indicator is determined by multiple regression analysis. Secondly, a spatial association equation is constructed, and the spatial autocorrelation analysis method is used to evaluate the spatial coupling effect between indicators. Then, a time series evolution equation is established, and the time series analysis method is used to study the dynamic change characteristics of the indicator contribution. Finally, the various contributions are integrated through the weight balance equation, and the genetic algorithm is used to solve the optimization equation group. The population size is set to 100, the evolutionary generations are 500, the crossover probability is 0.8, and the mutation probability is 0.1. The purpose of this step is to optimize the arable land protection potential score.

[0054] The specific implementation method of steps S10 and S11 is to carry out arable land protection zoning and priority area identification. First, the natural breakpoint classification method is used to divide the optimized score of arable land protection potential into three levels: high, medium and low, and the classification threshold is determined by the maximum inter-class variance method. Secondly, the spatial clustering analysis method is used to identify contiguous protection areas. The density clustering algorithm DBSCAN is used, the core point neighborhood radius is set to 500 meters, and the minimum number of points is set to 5. Then the minimum spanning tree algorithm is used to analyze the connectivity between regions and construct a spatial connectivity network. Finally, the graph theory method is used to identify key nodes and determine the priority protection areas. The purpose of this step is to form a scientific and reasonable spatial layout plan for arable land protection.

[0055] The soil quality characteristic pyramid neural network model achieves multi-scale feature analysis from local to global scales through five pyramid levels of varying scales. Each pyramid level expands the scope of analysis from the previous one, from a single patch of cultivated land to the entire region. This hierarchical analysis structure fully captures soil quality characteristic information at different spatial scales, avoiding the limitations of traditional single-scale analysis. Furthermore, the feature integration module achieves adaptive optimization of feature analysis by setting a credibility threshold, improving computational efficiency while ensuring analysis quality.

[0056] The optimized equations for cultivated land conservation contribution constitute a comprehensive analytical system consisting of four interrelated equations. The contribution benchmark equation establishes a basic contribution assessment framework using multiple evaluation indicators; the spatial correlation equation considers the spatial coupling between indicators; the temporal evolution equation incorporates dynamic analysis along the time dimension; and the weighted balance equation achieves a comprehensive balance of multidimensional information. This multi-equation analytical approach comprehensively considers all factors influencing cultivated land conservation potential assessment, providing more accurate and reliable results.

[0057] The soil quality characteristic pyramid neural network model and the cultivated land protection contribution optimization equation system, on the one hand, the pyramid model realizes multi-scale adaptive analysis of soil quality characteristics, improving the accuracy and reliability of feature extraction; on the other hand, the equation system systematically constructs a multi-dimensional and dynamic cultivated land protection potential evaluation framework, providing a more scientific basis for cultivated land protection decision-making. This technical combination significantly improves the refinement of cultivated land protection potential evaluation.

[0058] The overall implementation of this method utilizes a multi-level, multi-dimensional analytical framework, combining data-driven and model-driven approaches to achieve refined evaluation and optimization of cultivated land conservation potential. During the data acquisition and preprocessing phase, multi-source remote sensing data fusion technology and geostatistical analysis methods are employed to ensure the quality and reliability of the underlying data. During the feature extraction and analysis phase, deep learning and machine learning algorithms are employed to automatically extract cultivated land features and conduct multi-scale analysis. During the evaluation model construction and optimization phase, intelligent optimization algorithms and spatial analysis methods are employed to achieve scientific evaluation and spatial optimization of cultivated land conservation potential.

[0059] The convolutional layer output feature calculation of the soil quality feature pyramid neural network model is expressed as follows:

[0060] ;

[0061] Where, For the Layer feature map output, where They correspond to the five pyramid layers respectively; For the Layer feature map input; For the Layer convolution kernel weights; is the bias term; As the activation function, the ReLU function is used; Represents the convolution operation, when hour, Represents the input soil quality characteristic matrix.

[0062] The attention mechanism weight calculation is expressed as follows:

[0063] ;

[0064] ;

[0065] Where, For the The attention weight of each feature, ; Score for attention; is in hidden state; is the query vector; is the weight matrix, where is the hidden state weight matrix, is the query vector weight matrix; is the parameter vector, , is the number of features.

[0066] The contribution benchmark equation is expressed as follows:

[0067] ;

[0068] Where, As basic contribution; For the Normalized value of the evaluation index; is the corresponding weight; is the significance coefficient of the indicator; is the historical protection effect coefficient; is the regional development constraint coefficient; is the regression coefficient; is the random error term; It is the total number of evaluation indicators (including area parameter, boundary length parameter, fragmentation degree parameter, soil fertility index, soil physical structure index, soil chemical property index, light and heat resource condition index, water condition index, and climate stress factor index).

[0069] The spatial correlation equation is expressed as follows:

[0070] ;

[0071] Where, is the spatial correlation contribution; is the spatial weight matrix element, representing the area and region The spatial relationship between Index of the current evaluation unit; is the index of the adjacent evaluation unit, ; is the total number of adjacent evaluation units; is the spatial autocorrelation coefficient; is the terrain gradient coefficient; is the land use intensity index; is the landscape connectivity index; is the weight coefficient; is the random error term.

[0072] The time series evolution equation is expressed as follows:

[0073] ;

[0074] Where, Contribution for the current period; The contribution of the previous period; is the time attenuation coefficient; The protection effectiveness index over the years; is the climate change trend coefficient; is the socio-economic development index; is the weight coefficient; is the random error term.

[0075] The weight balance equation is expressed as follows:

[0076] ;

[0077] Where, The final contribution; is the comprehensive weight coefficient; is the regional differentiation correction coefficient; is the weight of the regional differentiation correction coefficient; is the random error term.

[0078] Parameter acquisition method description:

[0079] 1. The convolutional layer parameters are optimized by the back-propagation algorithm, with an initial learning rate of 0.01 and the Adam optimizer.

[0080] 2. The attention mechanism parameters are optimized by gradient descent, and the initial values ​​are randomly initialized in the range of [-0.1, 0.1];

[0081] 3. In the basic contribution equation, the normalized values ​​of the evaluation indicators are obtained through the minimum and maximum standardization method, the significance coefficient is calculated through principal component analysis, the historical protection effect coefficient is obtained through time series analysis, and the regional development constraint coefficient is determined through the hierarchical analysis method;

[0082] 4. In the spatial association equation, the spatial weight matrix was constructed using the inverse distance weighted method, the autocorrelation coefficient was calculated using the Moran's I index, the terrain gradient coefficient was calculated using the digital elevation model, and the land use intensity index and landscape connectivity index were calculated using the landscape index method;

[0083] 5. In the time series evolution equation, the time decay coefficient ranges from 0.8 to 0.95. The protection effectiveness index over the years is calculated using the cumulative scoring method. The climate change trend coefficient is obtained using the Mann-Kendall test. The socioeconomic development index is calculated using the entropy weight method.

[0084] 6. In the weight balance equation, the comprehensive weight coefficient is determined by combining the analytic hierarchy process and the entropy weight method, and the regional differentiation correction coefficient is obtained by cluster analysis;

[0085] 7. Error terms Normal distribution .

[0086] Explanation of the principle and significance of equation construction:

[0087] 1. The convolution layer uses multi-scale convolution kernels to extract features and introduces nonlinearity through the ReLU activation function to improve the model's expressiveness;

[0088] 2. The attention mechanism adopts an additive attention model and implements weight normalization through the softmax function to enhance the model's attention to important features;

[0089] 3. The contribution benchmark equation takes into account the direct impact of the indicator, historical impact, and regional constraints, and adopts a linear weighted form to facilitate parameter estimation and interpretation;

[0090] 4. The spatial correlation equation introduces spatial lag terms, takes into account spatial autocorrelation, and combines topography, land use and landscape characteristics to reflect spatial heterogeneity;

[0091] 5. The time series evolution equation adopts an autoregressive structure, introduces the time decay effect, and takes into account the influence of external factors such as climate and social economy;

[0092] 6. The weight balance equation realizes the fusion of multi-dimensional information through weighted combination and introduces correction coefficients to deal with regional differences.

[0093] Among them, the derivation process of the convolutional layer feature extraction equation is: first based on the basic form of the traditional convolutional neural network , considering the multi-scale nature of soil characteristics, a hierarchical structure is introduced to form ; In order to enhance the nonlinear expression ability of the model, the ReLU activation function is introduced, and finally This equation is optimized using a backpropagation algorithm, with the mean squared error (MSE) as the loss function. The optimization objective is to minimize the error between the predicted and true values. In this way, the model can adaptively extract soil characteristic information at different scales.

[0094] Among them, the derivation process of the attention mechanism weight calculation equation is: Based on the traditional attention mechanism, the importance score of each feature is first calculated , considering the correlation between features, the query vector is introduced and hidden state , build ;Then the score is converted into weight through the softmax function Parameter optimization is achieved by minimizing the cross-entropy loss function using gradient descent. This design enables the model to automatically identify and emphasize important feature information.

[0095] Among them, the derivation process of the contribution benchmark equation is: starting from the basic relationship of the evaluation indicators , taking into account the influence of historical factors and regional restrictions, the significance coefficient is introduced , historical effect coefficient and constraint coefficient , build The coefficients are obtained through multiple regression analysis, the significance coefficients are determined through principal component analysis, the historical effect coefficients are calculated based on time series analysis, and the constraint coefficients are determined through the analytic hierarchy process. This construction method can fully reflect the comprehensive impact of the evaluation indicators.

[0096] Among them, the derivation process of the spatial correlation equation: Based on the theory of spatial econometrics, the spatial lag term is introduced , considering the influence of topography, land use and landscape characteristics, construct The spatial weight matrix was constructed using the inverse distance weighting method, the autocorrelation coefficient was determined through the spatial autocorrelation test, and the characteristic coefficients were estimated using a spatial regression model. This equation structure can effectively capture spatial dependence and heterogeneity.

[0097] Among them, the derivation process of the time series evolution equation: using the autoregressive model framework , introduce external influencing factors, build The time decay coefficient is determined through time series analysis, and the coefficients of each influencing factor are estimated using a panel data regression model. This equation can reflect the dynamic evolution of contribution and the degree of influence of external factors.

[0098] Among them, the derivation process of the weight balance equation is based on the multi-criteria decision-making theory and adopts the weighted summation form. , considering regional differences, introducing correction terms, we get The weight coefficients were determined by combining the analytic hierarchy process and the entropy weight method, and the correction coefficients were obtained through cluster analysis. This structure enables the effective integration of multi-dimensional information and the reasonable adjustment of regional differences.

[0099] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, are used to execute the above-mentioned fine-grained farmland protection potential optimization method.

[0100] The third aspect of the present invention provides a fine-grained farmland protection potential optimization system, which includes the above-mentioned computer-readable storage medium. The system is any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0101] Specifically, the principles of the present invention are as follows: The technical principles of the present invention are based on deep learning and optimization theory. In terms of feature extraction, a pyramid-structured deep neural network is used to achieve multi-scale feature extraction by gradually expanding the receptive field. Each layer contains convolution kernels of different sizes, which can capture feature information at different scales. By introducing the attention mechanism and residual structure, the model's ability to learn important features is enhanced, while solving the problem of difficult deep network training. The feature integration module achieves effective fusion of features at different levels through an adaptive weight mechanism.

[0102] In terms of evaluation optimization, the constructed contribution optimization equations encompass four aspects: benchmark evaluation, spatial correlation, temporal evolution, and weight balance. The benchmark equation considers the direct impact of evaluation indicators and establishes the relationship between indicators and contribution through multivariate regression analysis. The spatial correlation equation incorporates spatial autocorrelation and heterogeneity analysis, characterizing the spatial dependencies between indicators through a spatial weight matrix. The temporal evolution equation employs an autoregressive structure to analyze the dynamic evolution of contribution. The weight balance equation utilizes multi-criteria decision-making theory to achieve the scientific integration of multidimensional information.

[0103] The solution proposed in this paper is able to accurately assess farmland conservation potential primarily because: first, the pyramid neural network model overcomes the inadequate feature extraction of traditional methods through multi-scale feature extraction and adaptive fusion; second, the contribution optimization equations, through multi-dimensional information fusion and dynamic optimization, address the subjectivity and untimely updates of traditional evaluation results. The entire technical solution forms a comprehensive analysis and evaluation system, with clear logical relationships and mutual support between modules, collectively achieving the goal of accurately assessing farmland conservation potential.

[0104] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0105] The specific implementation of step S01 is data acquisition and preprocessing. First, the data source is selected. Remote sensing image data uses Landsat 8 panchromatic and multispectral band data with spatial resolutions of 15 meters and 30 meters, respectively. Hyperspectral data uses Sentinel 2 data with a spatial resolution of 10 meters. Radar data uses Sentinel 1 data. The remote sensing image is geometrically precisely corrected using the control point method, with no fewer than 20 control points and a registration accuracy better than 0.5 pixels. Radiometric correction uses the radiometric calibration equation: , where is a digital quantized value, is the radiance value, and are gain coefficient and offset respectively. Atmospheric correction uses the radiation transfer equation: , where is the radiation brightness received by the sensor, is the radiation brightness of the target itself, is the atmospheric transmittance, Soil monitoring data was collected using a stratified sampling method, with a sampling density of 1 to 3 points per square kilometer and a sampling depth of 0 to 20 centimeters.

[0106] The specific implementation of step S02 is multi-dimensional data decomposition. First, the basic data set is subjected to correlation analysis and the Pearson correlation coefficient is calculated: , where and are the observed values ​​of two variables, and is the mean. The indicators with correlation coefficient greater than 0.95 are screened and merged. Then the principal component analysis method is used to extract features and calculate the covariance matrix: , where is the sample matrix, is the sample mean. Solve the characteristic equation: , obtain eigenvalues ​​and eigenvectors. Select principal components with cumulative contribution rates greater than 85% to reconstruct the characteristic matrix.

[0107] The specific implementation of step S03 is to extract the features of cultivated land patches. First, the cultivated land patches are extracted using the region growing algorithm. The seed point selection criterion is gray value similarity, and the growth criterion is: , where is the pixel grayscale value, is the regional mean, is the standard deviation, is the growth threshold, which is set to 2.0. The 8-neighborhood connectivity criterion is used to merge regions. The patch area parameters are calculated as follows: , where is the actual area of ​​the pixel. The boundary length is calculated using the Freeman chain code method: , where is the length of the chain code unit. Fragmentation index calculation: , where is the area, is the circumference.

[0108] The specific implementation of step S04 is soil quality feature analysis. The convolution feature extraction of each pyramid layer adopts the above convolution equation: The convolution kernel size increases from 3×3 in the first layer to 11×11 in the fifth layer. In the feature integration module, the attention weight is calculated using the aforementioned equation: and The soil quality feature credibility threshold was determined to be 0.85 through cross-validation. Calculation of the current layer feature credibility: , where is the true value, is the predicted value.

[0109] The specific implementation of step S05 is the climate condition assessment. The light and heat resource condition index adopts the fuzzy comprehensive evaluation method to construct an evaluation matrix: , where For the The evaluation object is The membership function uses the Gaussian function: , where is the center value, is the standard deviation. The water condition index is calculated based on the drought index: , where is the precipitation, is the potential evaporation. The climate stress factor index is calculated by extreme value analysis: , where is the weight coefficient.

[0110] The specific implementation of steps S06 and S07 is to evaluate model construction and training. The deep neural network structure includes the number of input layer nodes as the feature dimension, and the hidden layer adopts a residual structure: , where is the residual map. Batch normalization uses: , where is the batch mean, is the batch variance, Is a smoothing term. Cross entropy loss function: , where is the true label, is the predicted probability.

[0111] The specific implementation of steps S08 and S09 is contribution evaluation and optimization. The contribution benchmark equation, spatial correlation equation, temporal evolution equation, and weight balance equation use the aforementioned formulas. The system of equations is solved using a genetic algorithm, with real number encoding for chromosome encoding, a population size of 100, and 500 generations of evolution. The selection operator uses the roulette wheel method, and the crossover operator uses arithmetic crossover: , where is the cross coefficient, which is set to 0.8. The mutation operator uses Gaussian mutation: , where The fitness function is the inverse of the mean square error: .

[0112] The specific implementation of steps S10 and S11 is to generate and optimize the partitioning scheme. The natural breakpoint classification method is used to determine the partitioning threshold: , where For the level, is the sample value, is the intra-class mean. Spatial clustering uses the DBSCAN algorithm, and the core point determination criteria are: , where for point of The point set in the neighborhood, The value is 500 meters. The value is 5. The minimum spanning tree algorithm is used to analyze regional connectivity: , where is the edge weight.

[0113] Furthermore, the specific implementation of step S04 in this embodiment is to construct and apply a soil quality feature pyramid neural network model. The soil quality feature pyramid neural network model includes five pyramid layers executed sequentially and a feature integration module. Each pyramid layer uses a deep convolutional neural network structure and is constructed according to a multi-scale feature analysis approach from local to global.

[0114] The first pyramid layer is used to extract soil quality characteristics within cultivated land patches. The network structure consists of four convolutional layers and three pooling layers. The first and second convolutional layers use 3×3 convolution kernels with SAME padding and a stride of 1, and have output channels of 64 and 128, respectively. The third and fourth convolutional layers use 5×5 convolution kernels with SAME padding and a stride of 1, and have output channels of 256 and 512, respectively. All pooling layers use 2×2 max pooling with a stride of 2. The first pyramid layer inputs data on soil organic matter content, bulk density, pH, nutrient content, moisture content, texture, and erosion severity within the cultivated land patches, fed into an input layer with a feature map size of 64×64. This layer uses batch normalization and a ReLU activation function to achieve feature normalization and nonlinear mapping.

[0115] The second pyramid layer is used to extract soil quality feature combinations within and adjacent to cultivated land patches. The network structure consists of four convolutional layers and three pooling layers. The first and second convolutional layers use 5×5 convolution kernels with SAME padding and a stride of 1, resulting in output channels of 128 and 256, respectively. The third and fourth convolutional layers use 7×7 convolution kernels with SAME padding and a stride of 1, resulting in output channels of 512 and 1024, respectively. All pooling layers use 2×2 max pooling with a stride of 2. The second pyramid layer's input data includes the output feature maps of the first pyramid layer and soil quality data from the adjacent areas. The feature maps are scaled to 128×128 pixels. This layer incorporates an attention mechanism to highlight key feature information by calculating feature importance weights.

[0116] The third pyramid layer is used to extract soil quality feature combinations within cultivated land patches, adjacent areas, and surrounding buffer zones. The network structure consists of four convolutional layers and three pooling layers. The first and second convolutional layers use 7×7 convolution kernels with SAME padding and a stride of 1, resulting in output channels of 256 and 512, respectively. The third and fourth convolutional layers use 9×9 convolution kernels with SAME padding and a stride of 1, resulting in output channels of 1024 and 2048, respectively. All pooling layers use 2×2 max pooling with a stride of 2. The third pyramid layer's input data includes the output feature maps of the second pyramid layer and soil quality data for the buffer zones. The feature maps are scaled to 256×256. This layer uses a residual connection structure to alleviate the training difficulties of deep networks.

[0117] The fourth pyramid layer extracts soil quality feature combinations within cultivated land patches, adjacent areas, buffer zones around cultivated land patches, and watershed units within cultivated land patches. The network architecture consists of four convolutional layers and three pooling layers. The first and second convolutional layers use 9×9 convolution kernels with SAME padding and a stride of 1, resulting in 512 and 1024 output channels, respectively. The third and fourth convolutional layers use 11×11 convolution kernels with SAME padding and a stride of 1, resulting in 2048 and 4096 output channels, respectively. All pooling layers use 2×2 max pooling with a stride of 2. The fourth pyramid layer's input data includes the output feature maps of the third pyramid layer and soil quality data for the watershed units. The feature maps are scaled to 512×512. This layer incorporates a channel-wise attention mechanism to enhance information interaction between feature channels.

[0118] The fifth pyramid layer is used to extract soil quality features within cultivated land patches, adjacent areas, buffer zones around cultivated land patches, watershed units within cultivated land patches, and the entire region. The network architecture consists of four convolutional layers and three pooling layers. The first and second convolutional layers use 11×11 convolution kernels with SAME padding and a stride of 1, resulting in output channels of 1024 and 2048, respectively. The third and fourth convolutional layers use 13×13 convolution kernels with SAME padding and a stride of 1, resulting in output channels of 4096 and 8192, respectively. All pooling layers use 2×2 max pooling with a stride of 2. The fifth pyramid layer's input data includes the output feature maps of the fourth pyramid layer and regional soil quality data. The feature maps are scaled to 1024×1024. This layer uses a global average pooling layer to extract global information from the features.

[0119] The feature integration module consists of a feature fusion unit and a judgment unit. The feature fusion unit uses an adaptive weighting method to fuse features output from different pyramid layers. The weight coefficients are optimized using a backpropagation algorithm. The judgment unit calculates feature credibility to determine whether the output of the current pyramid layer meets the accuracy requirements. When the feature credibility is greater than the threshold of 0.85, the output of the current layer is used as the final feature. When the feature credibility is less than the threshold, the analysis proceeds to the next layer or the output of the last layer is used as the final feature. Feature credibility is calculated using a cross-validation method, which compares the predicted results with the actual values.

[0120] The innovations of this pyramid neural network model lie in: employing a multi-scale feature extraction strategy to achieve progressive analysis from local to global perspectives; introducing an attention mechanism and residual structure to improve feature extraction accuracy; and implementing adaptive feature selection through a feature integration module to ensure the reliability of the analysis results. The model was trained using a stochastic gradient descent optimizer with an initial learning rate of 0.01, dynamically adjusted using a cosine annealing strategy, and 1000 training rounds. Within each round, network parameters were updated using a backpropagation algorithm using a mini-batch stochastic gradient descent method with a batch size of 32. To prevent overfitting, L2 regularization with a weight decay coefficient of 0.0001 and a dropout layer with a random dropout rate of 0.5 were employed. This model effectively extracts multi-scale soil quality features, providing reliable feature support for the evaluation of cultivated land conservation potential.

[0121] This method achieves a refined assessment of farmland conservation potential through a multi-level, multi-dimensional analytical framework. At the data processing level, multi-source data fusion and feature extraction techniques are employed to ensure the reliability of the underlying data. At the model construction level, deep learning and optimization algorithms are used to automatically extract features and optimize model training. At the evaluation and optimization level, a comprehensive contribution evaluation system is constructed, achieving a comprehensive balance of multi-dimensional information. The innovation of this method lies in the establishment of an adaptive and dynamic evaluation framework, which improves the accuracy and reliability of the assessment of farmland conservation potential.

[0122] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below:

[0123] The study area lies between 35° and 36° north latitude and 115° and 116° east longitude, covering approximately 8,000 square kilometers, of which approximately 4,500 square kilometers are cultivated land. The region boasts a complex terrain consisting of plains, hills, and mountains, with altitudes ranging from 10 to 800 meters above sea level. The climate is warm temperate, semi-humid monsoon, with an average annual temperature of 14.2°C and annual precipitation of 850 mm.

[0124] First, the research team established a multi-source data collection system. Remote sensing data collection includes: panchromatic imagery (15-meter resolution) and multispectral imagery (30-meter resolution) acquired by Landsat 8 on May 1, 2024; multi-temporal multispectral data (10-meter resolution) acquired by Sentinel-2 on April 28 and May 3, 2024; and SAR data acquired by Sentinel-1 on May 2, 2024. A systematic stratified sampling scheme was designed for soil monitoring data collection: 100 standard plots of 1 square kilometer were established within the study area, with 25 sampling points within each standard plot. The sampling depths were divided into two layers: 0-20 cm and 20-40 cm, for a total of 5,000 samples. Climate data were obtained from continuous observations recorded by 25 meteorological stations within the study area from 2019 to 2024. Fifty automatic weather stations were also deployed for enhanced observations.

[0125] The analysis results of the first batch of soil monitoring data are shown in Table 1 below:

[0126] Table 1 Analysis results of the first batch of soil monitoring data

[0127]

[0128] The specific implementation process of the soil quality feature pyramid neural network model is shown in Table 2 below:

[0129] Table 2 Normalization results of the first pyramid layer input data:

[0130]

[0131] Specific parameter settings and training results of each layer of the pyramid network: the first layer uses 64 3×3 convolution kernels, the activation function uses ReLU, the pooling uses 2×2 maximum pooling, and the batch normalization layer parameters are =0.02, =0.015; the second layer uses 128 5×5 convolution kernels, with all other parameters set identically; the third layer uses 256 7×7 convolution kernels; the fourth layer uses 512 9×9 convolution kernels; and the fifth layer uses 1024 11×11 convolution kernels. The feature extraction performance of each layer is evaluated using confidence: the average confidence is 0.832 for the first layer, 0.856 for the second layer, 0.878 for the third layer, 0.891 for the fourth layer, and 0.902 for the fifth layer.

[0132] The attention weight calculation results of the feature integration module: =0.15 (local features), =0.20 (neighboring features), =0.25 (buffer characteristics), =0.22 (watershed characteristics), =0.18 (regional features). The Adam algorithm was used for weight optimization. The initial learning rate was set to 0.01 and decayed to 0.8 times the original value every 200 rounds.

[0133] The spatial distribution characteristics of the final evaluation results are as follows: the high protection potential area is about 1,800 square kilometers (40%), mainly distributed in the plains, with an average soil quality characteristic index of 0.856 and an average climate suitability index of 0.823; the medium protection potential area is about 2,025 square kilometers (45%), mainly distributed in the hilly areas, with an average soil quality characteristic index of 0.685 and an average climate suitability index of 0.742; the low protection potential area is about 675 square kilometers (15%), mainly distributed in the mountainous areas, with an average soil quality characteristic index of 0.523 and an average climate suitability index of 0.612.

[0134] Compared with traditional evaluation methods, the technical solution adopted in this embodiment has significant advantages: First, the traditional method uses expert scoring to determine weights, which is highly subjective and results in large fluctuations in evaluation results. However, this solution automatically extracts features through a pyramid neural network, which is highly objective and produces stable evaluation results. Second, the traditional method can only process a limited number of evaluation indicators, while this solution can simultaneously process multi-source, multi-dimensional, and multi-scale data, making more effective use of information. Third, the traditional method lacks spatial correlation analysis, while this solution fully considers spatial dependence and heterogeneity through multi-level spatial analysis. Finally, the evaluation results of the traditional method are static and difficult to update, while this solution can update the evaluation results in real time through a dynamic optimization mechanism. Compared with the expert evaluation results, the evaluation accuracy of this solution increased by 23.5%, the spatial precision increased by 35.8%, and the time response speed increased by 68.2%.

[0135] This example fully demonstrates the feasibility and superiority of this technical solution in the refined evaluation of farmland conservation potential, providing a scientific basis for decision-making on farmland conservation. Adaptive feature extraction is achieved through a soil quality feature pyramid neural network model, and dynamic optimization of evaluation results is achieved through a set of contribution optimization equations, ultimately solving the technical problem of accurately evaluating farmland conservation potential.

[0136] It should be noted that the variables involved in the present invention are explained in detail as shown in Table 3 below.

[0137] Table 3 Variable Explanation Table

[0138]

[0139] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A fine-grained method for optimizing cultivated land conservation potential, characterized in that: include: Acquire remote sensing image data, soil monitoring data, and climate observation data of the target area as basic data sets; The basic data set is decomposed in multiple dimensions to obtain a land use type distribution matrix, a soil quality characteristic matrix, and a climate suitability matrix; cultivated land patches are extracted based on the land use type distribution matrix, and the area parameters, boundary length parameters, and fragmentation degree parameters of the cultivated land patches are calculated; the soil quality characteristic matrix is ​​analyzed using a soil quality characteristic pyramid neural network model; the light and heat resource condition index, water condition index, and climate stress factor index of the cultivated land patches are evaluated based on the climate suitability matrix; a cultivated land protection potential evaluation model is constructed, and the area parameters, boundary length parameters, fragmentation degree parameters, soil fertility index, soil physical structure index, soil chemical property index, light and heat resource condition index, water condition index, and climate stress factor index of the cultivated land patches are input into the cultivated land protection potential evaluation model; and an initial score for cultivated land protection potential is obtained using a neural network model training; Establishing a cultivated land protection contribution matrix based on the initial score of cultivated land protection potential, wherein the cultivated land protection contribution matrix is ​​obtained by solving a contribution optimization equation group including a contribution benchmark equation, a spatial correlation equation, a temporal evolution equation, and a weight balance equation; optimizing and adjusting to obtain an optimized score; Divide high, medium and low protection potential areas; conduct spatial connectivity analysis, identify priority protection areas and output them; the soil quality characteristic pyramid neural network model includes five pyramid layers and feature synthesis modules that are executed in sequence; the pyramid layers include: the first pyramid layer is used to analyze the soil quality characteristics inside the cultivated land patch; the second pyramid layer is used to analyze the combination of soil quality characteristics inside the cultivated land patch and the adjacent areas of the cultivated land patch; the third pyramid layer is used to analyze the combination of soil quality characteristics inside the cultivated land patch, the adjacent areas of the cultivated land patch and the buffer zone around the cultivated land patch; the fourth pyramid layer is used to analyze the combination of soil quality characteristics inside the cultivated land patch, the adjacent areas of the cultivated land patch, the buffer zone around the cultivated land patch and the cultivated land patch The fifth pyramid layer is used to analyze the soil quality characteristic combination of the interior of the cultivated land patch, the adjacent area of ​​the cultivated land patch, the buffer zone around the cultivated land patch, the cultivated land patch watershed unit and the overall soil quality characteristic combination of the region; the feature synthesis module is used to determine whether the soil quality characteristic index output by the pyramid layer is greater than the soil quality characteristic credibility threshold. When it is greater than the soil quality characteristic credibility threshold, the soil quality characteristic index output by the current pyramid layer is used as the final soil quality characteristic index; when the feature synthesis module is less than the soil quality characteristic credibility threshold, it continues to the next layer of pyramid analysis or uses the output of the last layer of pyramid as the final soil quality characteristic index.

2. The method according to claim 1, characterized in that The remote sensing image data includes multispectral remote sensing image data, hyperspectral remote sensing image data, radar remote sensing image data, surface cover classification data and normalized vegetation index data.

3. The method according to claim 2, characterized in that The soil monitoring data includes soil organic matter content data, soil bulk density data, soil pH value data, soil nutrient content data, soil moisture content data, soil texture data and soil erosion degree data.

4. The method according to claim 3, characterized in that The climate observation data includes annual average temperature data, annual precipitation data, sunshine hours data, cumulative temperature data, drought index data, precipitation seasonal distribution data and extreme weather event data.

5. The method according to claim 4, characterized in that The contribution benchmark equation is used to calculate the basic contribution of each evaluation indicator, the spatial correlation equation is used to evaluate the spatial coupling effect between evaluation indicators, and the time series evolution equation is used to analyze the time variation characteristics of the contribution of evaluation indicators.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the fine-grained farmland protection potential optimization method according to any one of claims 1 to 5.

7. A fine-grained farmland conservation potential optimization system, characterized by: The system comprises the computer-readable storage medium according to claim 6, wherein the system is any one of a computer, a server, and a single-chip microcomputer, the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

Images