Convolutional neural network-based arable land reserve resource investigation and evaluation method and system

Abstract

The application discloses a cultivated land reserve resource investigation and evaluation method and system based on a convolutional neural network, and particularly relates to the technical field of remote sensing image processing and cultivated land resource evaluation, and is used for solving the problem that the evaluation result is inaccurate due to the amplification of the error of the previous stage segmentation in the cascade convolutional neural network being transmitted to the next stage; the boundary fuzzy area is accurately positioned by adopting a double-threshold judgment criterion of the fractal dimension change rate and the optimal transmission cost distance and by identifying the non-cultivated land transition zone patches in the remote sensing image; secondly, the feature activation trajectory of the transition zone patch in the convolutional network is extracted, and an error propagation weight coefficient calculation model is established; then, the weighted feature offset is generated by combining the boundary channel response value and the weight coefficient, and an error compensation mask is constructed; finally, the evaluation indexes of the terrain slope and the contiguity are corrected through the pixel-level fusion of the mask and the feature map of the next stage, the error transmission effect of the cascade architecture is significantly inhibited, and the reliability of the cultivated land reserve resource grading result is improved.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and arable land resource evaluation, and in particular to a method and system for surveying and evaluating arable land reserve resources based on convolutional neural networks. Background Technology

[0002] In the survey and evaluation of reserve farmland resources, existing technologies typically employ a cascaded processing flow: first, non-arable land patches are segmented based on remote sensing images, and then the segmentation results are used to evaluate the potential of arable land. This process relies on the cascaded deployment of convolutional neural networks (CNNs), where the front-end CNN model performs patch boundary recognition, and the back-end CNN model calculates key evaluation indicators such as terrain slope and contiguousness based on the boundary data, in order to achieve large-scale screening of reserve farmland resources.

[0003] However, the existing cascaded architecture has its flaws: the output error of the previous stage of patch segmentation is nonlinearly amplified during the transmission to the subsequent evaluation stage. Since the identification bias of the segmentation model in transitional zones such as cultivated land and forest land is unavoidable, such errors will significantly distort the calculation results of core indicators such as terrain slope after being processed by the evaluation model. Ultimately, this leads to a systematic inaccuracy in the grading conclusions of cultivated land reserve resources. The uncontrolled transmission of feature errors in the cascaded CNN causes the actual reliability of the survey and evaluation results to be lower than theoretical expectations. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a method and system for investigating and evaluating arable land reserve resources based on convolutional neural networks.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: This invention provides the following technical solution: Methods for surveying and evaluating arable land reserve resources based on convolutional neural networks include: S1. Obtain remote sensing image data of the area to be evaluated, and perform non-cultivated map patch segmentation processing through the pre-stage CNN model to generate an initial set of segmented patches. S2. For the boundary regions of each patch in the initial segmented patch set, when the rate of change of the fractal dimension of the boundary pixel relative to the standard cultivated land boundary exceeds the terrain abrupt change threshold and the transmission cost distance of the feature distribution on both sides of the boundary based on the optimal transmission theory exceeds the distribution difference threshold, the corresponding patch is marked as a transition zone patch. S3. Extract the feature activation trajectories of transition zone patches in the previous CNN model, construct a spatial topology network based on the feature correlation strength between convolutional layers, and calculate the error propagation weight coefficients. S4. Extract the boundary-related channel response values ​​from the feature map output by the previous CNN model, and calculate the weighted feature offset by combining the error propagation weight coefficients. S5. After generating an error compensation mask based on the weighted feature offset, perform pixel-level weighted fusion with the input feature map of the subsequent CNN model to generate a fused feature map. S6. Use the fused feature map to perform farmland potential evaluation, and output the slope calculation correction value and the reserve resource classification results.

[0006] Furthermore, remote sensing image data of the area to be evaluated is acquired, and non-cultivated land parcels are segmented using a pre-stage CNN model to generate an initial set of segmented parcels, including: Acquire multispectral remote sensing image data of the area to be evaluated; Preprocessing of multispectral remote sensing image data; The preprocessed multispectral remote sensing image data is input into the pre-stage CNN model, and the pre-stage CNN model outputs a binary segmentation mask for non-farmland and farmland. Connectivity analysis is performed on the binary segmentation mask to generate an initial set of segmentation patches, where each patch is defined by a sequence of boundary coordinates.

[0007] Furthermore, preprocessing includes radiometric calibration, atmospheric correction, and geometric registration.

[0008] Furthermore, for each boundary region of a patch in the initial segmented patch set, transition zone patches corresponding to the patches are identified and marked, including: Extract the feature space coordinates of pixels in the last convolutional layer of the preceding CNN model from the boundary regions of each patch in the initial segmentation patch set; Based on the extracted feature space coordinates, the fractal dimension value of the pixel feature vector of the boundary region is calculated by box counting. The calculated fractal dimension value is compared with the fractal dimension benchmark value of the standard cultivated land boundary to obtain the fractal dimension change rate; Extract the set of feature vectors from the last convolutional layer of the preceding CNN model from the cultivated land side of the map boundary; Extract the set of feature vectors from the pixel output of the last convolutional layer of the previous CNN model from the non-cultivated land side of the map boundary; The feature vector set on the cultivated land side of the boundary and the feature vector set on the non-cultivated land side of the boundary are respectively used as discrete probability distributions; Transmission cost distance is generated based on the optimal transmission theory derived from two discrete probability distributions; When the fractal dimension change rate exceeds the terrain mutation threshold and the transmission cost distance exceeds the distribution difference threshold, the current patch is marked as a transition zone patch.

[0009] Furthermore, the transmission cost distance is obtained by calculating the Wasserstein distance using the entropy-regularized Sinkhorn iterative algorithm.

[0010] Furthermore, the feature activation trajectories of transition zone patches in the preceding CNN model are extracted, a spatial topology network is constructed based on the feature correlation strength between convolutional layers, and error propagation weight coefficients are calculated, including: Obtain the feature maps of the pixels in the transition zone from the output of all convolutional layers of the previous CNN model; Extract the feature activation values ​​at the corresponding transition zone patch positions in the feature map of each convolutional layer to form a feature activation trajectory sequence; Calculate the mutual information value between the feature activation trajectory sequences of adjacent convolutional layers as the feature correlation strength between convolutional layers; Using the convolutional layers of the previous CNN model as nodes and the feature association strength as the edge weights, a spatial topology network is constructed. In spatial topological networks, the centrality value of each node is calculated using the eigenvector centrality algorithm; The calculated centrality values ​​are normalized to obtain the error propagation weight coefficients.

[0011] Furthermore, the boundary-related channel response values ​​are extracted from the feature map output by the previous CNN model, and the weighted feature offset is calculated by combining the error propagation weight coefficients, including: Identify target channels related to the boundary detection task in the feature map output by the previous CNN model; Extract the feature response values ​​of the target channel at the corresponding spatial location of the transition zone patch; The extracted feature response values ​​are multiplied by the error propagation weight coefficients using a weighted multiplication operation. The weighted multiplication result is subjected to mean pooling in the spatial dimension; The result of mean pooling is output as the weighted feature offset.

[0012] Furthermore, after generating an error-compensated mask based on the weighted feature offset, it is pixel-level weighted fused with the input feature map of the subsequent CNN model to generate a fused feature map, including: Convert the weighted feature offset into a compensation strength coefficient; Create a blank mask matrix with the same size as the input feature map of the subsequent CNN model; In the blank mask matrix, fill the corresponding regions of the transition zone patches with compensation intensity coefficients; Fill the non-transition zone patch locations in the blank mask matrix with fixed values; Use the filled blank mask matrix as the error compensation mask; The error compensation mask is multiplied with the input feature map of the subsequent CNN model according to the pixel position; The multiplication result is added to the input feature map of the subsequent CNN model according to the pixel position. The summation result is output as a fused feature map.

[0013] Furthermore, the fused feature map is used to perform arable land potential assessment, outputting slope calculation correction values ​​and reserve resource classification results, including: The fused feature maps are then input into the evaluation network layer of the subsequent CNN model; The response values ​​of terrain slope and farmland contiguousness were extracted by evaluating the network layers. The terrain slope characteristic response value is linearly scaled and transformed to obtain the slope calculation correction value; Input the characteristic response value of farmland contiguousness into the preset hierarchical rule library; The backup resource classification result is determined by matching the threshold range in the classification rule base; Output the slope calculation correction value and the backup resource classification results.

[0014] On the other hand, the present invention provides a system for surveying and evaluating arable land reserve resources based on convolutional neural networks, comprising: The image processing module is used to acquire remote sensing image data of the area to be evaluated, and to perform non-cultivated patch segmentation processing through the front-end CNN model to generate an initial set of segmented patches. The patch marking module is used to mark the boundary regions of each patch in the initial segmented patch set as transition zone patches when the rate of change of the fractal dimension of the boundary pixels relative to the standard cultivated land boundary exceeds the terrain abrupt change threshold and the transmission cost distance based on the optimal transmission theory of the feature distribution on both sides of the boundary exceeds the distribution difference threshold. The weight calculation module is used to extract the feature activation trajectories of transition zone patches in the previous CNN model, construct a spatial topology network based on the feature correlation strength between convolutional layers, and calculate the error propagation weight coefficients. The offset calculation module is used to extract the boundary-related channel response values ​​from the feature map output by the previous CNN model, and calculate the weighted feature offset by combining the error propagation weight coefficients. The feature fusion module is used to generate an error compensation mask based on the weighted feature offset and then perform pixel-level weighted fusion with the input feature map of the subsequent CNN model to generate a fused feature map. The potential assessment module is used to perform farmland potential assessment using fused feature maps, and output slope calculation correction values ​​and reserve resource classification results.

[0015] The beneficial effects of this invention are: 1. By constructing an error closed-loop correction mechanism, the problem of feature error propagation distortion in the cascaded CNN architecture is effectively solved. To address the distortion of evaluation metrics caused by transition zone patch recognition deviations, a dual-threshold transition zone judgment criterion (fractal dimension change rate + optimal transmission cost distance) is proposed. By fusing the dual criteria of geometric morphology and feature distribution, the blurred areas of farmland boundaries are accurately captured, significantly improving the transition zone localization accuracy compared to a single feature criterion. A feature activation trajectory tracking mechanism is designed, dynamically constructing an error propagation network based on the spatial topological association between convolutional layers, enabling quantifiable modeling of feature error propagation paths within the model. Weighted feature offset compensation is introduced, generating an error compensation mask that matches the size of the subsequent feature map, achieving adaptive correction of the feature space at the pixel level.

[0016] 2. A synergistic effect is formed, fundamentally suppressing the cascading amplification of previous segmentation errors: the transition zone marking stage blocks the source of error propagation, the weight coefficient calculation stage quantifies the intensity of error propagation, and the feature compensation stage achieves precise error cancellation. Ultimately, in the farmland potential evaluation stage, this is reflected in the following: the calculated topographic slope value is closer to the actual measurement data than traditional methods, especially in complex areas such as farmland-forest transition zones, where the reliability of the slope calculation correction value is significantly improved; the spatial distribution rationality of the reserve resource classification results is enhanced, eliminating misjudgments of farmland contiguousness caused by segmentation errors, providing a more credible decision-making basis for land spatial planning, and realizing the controllability and transparency of the error propagation path of the cascading architecture. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for investigating and evaluating arable land reserve resources based on convolutional neural networks according to the present invention. Figure 2 This is a schematic diagram of the structure of the arable land reserve resource survey and evaluation system based on convolutional neural networks of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Figure 1 This invention presents a method for investigating and evaluating arable land reserve resources based on convolutional neural networks, comprising: S1. Obtain remote sensing image data of the area to be evaluated, and perform non-cultivated map patch segmentation processing through the pre-stage CNN model to generate an initial set of segmented patches. S2. For the boundary regions of each patch in the initial segmented patch set, when the rate of change of the fractal dimension of the boundary pixel relative to the standard cultivated land boundary exceeds the terrain abrupt change threshold and the transmission cost distance of the feature distribution on both sides of the boundary based on the optimal transmission theory exceeds the distribution difference threshold, the corresponding patch is marked as a transition zone patch. S3. Extract the feature activation trajectories of transition zone patches in the previous CNN model, construct a spatial topology network based on the feature correlation strength between convolutional layers, and calculate the error propagation weight coefficients. S4. Extract the boundary-related channel response values ​​from the feature map output by the previous CNN model, and calculate the weighted feature offset by combining the error propagation weight coefficients. S5. After generating an error compensation mask based on the weighted feature offset, perform pixel-level weighted fusion with the input feature map of the subsequent CNN model to generate a fused feature map. S6. Use the fused feature map to perform farmland potential evaluation, and output the slope calculation correction value and the reserve resource classification results.

[0020] S1. Obtain remote sensing image data of the area to be evaluated, and perform non-cultivated patch segmentation processing using a pre-stage CNN model to generate an initial set of segmented patches. The specific implementation is as follows: Multispectral remote sensing imagery data of the area to be evaluated is acquired. Specifically, visible light, near-infrared, and shortwave infrared imagery data of the target geographic area are collected via a satellite remote sensing platform. The spatial resolution of the images is controlled within, for example, 10 to 30 meters, covering the complete geographic area to be evaluated. Acquisition of multispectral remote sensing imagery data must ensure temporal synchronization of data across all bands. Image acquisition should be conducted during, for example, the peak vegetation growth period from May to September each year, under clear weather conditions with cloud cover, for example, less than 10%. After acquisition, the data is stored as GeoTIFF format files containing geographic coordinate reference information, using the National Geodetic Coordinate System 2000.

[0021] The acquired multispectral remote sensing image data underwent preprocessing. First, radiometric calibration was performed, converting the raw digital quantization values ​​recorded by the sensor into atmospheric top-layer reflectance. This was achieved using a calibration coefficient file provided by the sensor manufacturer, which contained band response functions and radiometric calibration parameters. Radiometric calibration was performed using ENVI software, selecting the reflectance calibration type. Next, atmospheric correction was performed, employing the 6S radiative transfer model to eliminate atmospheric scattering and absorption effects. Input parameters included the acquisition time, geographic location (latitude and longitude coordinates) extracted from image metadata, and the atmospheric model type and rural aerosol model parameters determined according to the season. Finally, geometric registration was performed, using orthophoto maps from the National Geographic Information Database as a reference. A quadratic polynomial correction model was used, selecting, for example, 20 spatially uniformly distributed ground control points for precise registration, with the planar position error controlled within, for example, 0.5 pixel dimensions.

[0022] The preprocessed multispectral remote sensing image data is input into a pre-processing CNN model. This model employs an encoder-decoder convolutional neural network architecture. The encoder consists of five convolutional computation modules, each composed of a 3×3 pixel convolutional kernel, a batch normalization layer, and a ReLU activation function in sequence. The decoder uses transposed convolutions for upsampling, and the final output layer uses a Sigmoid activation function to generate a probability map. The pre-processing CNN model initializes its network parameters through transfer learning. The base model uses a ResNet34 network pre-trained on the ImageNet dataset, followed by fine-tuning training using, for example, 100,000 labeled farmland samples. During training, the input data size is adjusted to, for example, 512×512 pixels, and the batch size is set to, for example, 16. Cross-entropy loss and the Adam optimizer are used for parameter updates. The pre-processing CNN model outputs a binary segmentation mask for non-farmland and farmland, where pixel value 0 represents non-farmland and pixel value 1 represents farmland.

[0023] Connectivity analysis is performed on the binary segmentation mask output by the preceding CNN model, specifically using the 8-neighborhood connectivity criterion to scan the entire mask image. First, morphological closing operations are performed on the binary image, using 3×3 pixel circular structuring elements to eliminate small void regions. Then, a two-pass scanning algorithm is used to label connected regions: the first pass assigns temporary labels row-by-row and records equivalence relations; the second pass merges the temporary labels according to the equivalence relation table to generate independent patches. Each connected region corresponds to an independent patch object, with an area threshold set (e.g., 400 square meters) to filter out noise patches with areas smaller than this threshold. Finally, the boundary coordinate sequence of each patch object is extracted. Specifically, a boundary tracing algorithm is used, starting from the outermost top-left pixel and traversing clockwise to record the planar coordinates of the center point of the boundary pixels. The coordinate values ​​are preserved to two decimal places, and the planar coordinate system is consistent with the original image's geographic coordinate system. This generates an initial set of segmented patches. Each patch data structure contains a unique identifier, a boundary coordinate sequence, and an area attribute value, and is finally stored and output in GeoJSON format.

[0024] S2. For the boundary regions of each patch in the initial segmented patch set, when the rate of change of the fractal dimension of the boundary pixels relative to the standard cultivated land boundary exceeds the terrain abrupt change threshold and the transmission cost distance based on the optimal transmission theory of the feature distribution on both sides of the boundary exceeds the distribution difference threshold, the corresponding patch is marked as a transition zone patch. The specific implementation is as follows: The feature space coordinates of pixels in the boundary regions of each patch in the initial segmented patch set are extracted from the output of the last convolutional layer of the preceding CNN model. Specifically, for each patch's boundary region, a buffer region is expanded outwards from the boundary coordinate sequence as the center line. The buffer width is set to, for example, a distance of 5 pixels. All pixels within this buffer region are traversed, and the feature vector corresponding to the spatial location is obtained from the feature map output by the last convolutional layer of the preceding CNN model. The feature vector has a dimension of, for example, 512, with each dimension corresponding to an abstract feature extracted by the convolutional kernel. The feature space coordinates are composed of the values ​​of each dimension of the feature vector, and the coordinate system is the real coordinate system of the CNN feature space. The spatial resolution of the feature map maintains the same proportional relationship as the input image, and the feature value range is controlled, for example, between -10.0 and +10.0. The boundary region range is controlled, for example, between 5% and 15% of the patch perimeter, to ensure that the analysis area is statistically significant.

[0025] Based on the extracted feature space coordinates, the fractal dimension of the pixel feature vectors in the boundary region is calculated using box counting. The specific implementation process is as follows: A cubic mesh system is established in the feature space. The initial mesh side length is set to, for example, 1 / 10 of the feature value range, and the side length is gradually reduced in multiples of 2 to form a geometric progression. For each mesh size, the number of meshes containing at least one feature point is counted, excluding empty meshes. A Cartesian coordinate system is established, representing the logarithm of the mesh size and the logarithm of the number of meshes. A straight line is fitted using the least squares method, and the absolute value of the slope of this line is the fractal dimension. During the calculation, the mesh size variation range is controlled between, for example, 0.5 to 3 times the standard deviation of the eigenvalues, ensuring that at least 5 data points are fitted. The standard deviation of the eigenvalues ​​is calculated by statistically analyzing the variances of each dimension of all feature points in the current boundary region.

[0026] The calculated fractal dimension value is compared with the baseline fractal dimension value of the standard cultivated land boundary to obtain the fractal dimension change rate. The baseline fractal dimension value of the standard cultivated land boundary is obtained by statistically analyzing typical cultivated land samples. Specifically, 1000 standard cultivated land patches are selected from the National Cultivated Land Resources Database. The fractal dimension value of each patch is calculated using the aforementioned box counting method. After removing outliers, the arithmetic mean is taken as the baseline value. The formula for calculating the fractal dimension change rate is: (current patch fractal dimension value minus the baseline value) divided by the baseline value and then multiplied by 100%. The calculation result is retained to two decimal places. The baseline value is stored in a parameter database and supports dynamic updates according to different agricultural ecological zones.

[0027] The feature vector set of pixels from the cultivated land side of the map patch boundary is extracted from the output of the last convolutional layer of the preceding CNN model. Specifically, the buffer region is extended inward from the map patch boundary line, with a buffer width set to, for example, 5 pixels, defined as the cultivated land side analysis area. All pixels within this region are traversed, and feature vectors corresponding to their spatial locations are extracted from the feature map of the last layer of the preceding CNN model. Each feature vector contains, for example, 512 floating-point elements, forming a vector set of the same dimension. Simultaneously, a feature vector set is extracted from the non-cultivated land side of the map patch boundary. The non-cultivated land side is defined as the area extending outward from the boundary line by the same pixel distance. The extraction method is completely consistent with that of the cultivated land side, ensuring symmetrical widths on both sides. When the number of pixels in the boundary region is less than, for example, 50, the buffer width is automatically expanded to meet the minimum sample size requirement.

[0028] The feature vector sets on the cultivated land side and the non-cultivated land side of the boundary are respectively treated as discrete probability distributions. The transformation method is as follows: First, each feature vector is normalized using the L2 norm, calculated as the square root of the sum of the squares of each element of the feature vector. Then, each element of the original vector is divided by the norm value to make the feature vector magnitude 1. The normalized feature vectors are then considered as discrete points in the probability space. All feature vectors on the cultivated land side constitute the first discrete probability distribution, and those on the non-cultivated land side constitute the second discrete probability distribution. Each distribution contains the same number of points as the number of pixels in the corresponding region, for example, 100 to 1000 points. The total mass of the probability distribution is 1, and the mass of each point is uniformly distributed.

[0029] Based on two discrete probability distributions, the Wasserstein distance is calculated using the entropy-regularized Sinkhorn iterative algorithm, serving as the transmission cost distance based on optimal transmission theory. The specific implementation steps are as follows: First, construct a cost matrix C, where each element Cij represents the squared Euclidean distance between the i-th point in the first distribution and the j-th point in the second distribution. The row index i represents the i-th feature point in the first distribution (i ranges from 1 to M); the column index j represents the j-th feature point in the second distribution (j ranges from 1 to N); the matrix size is M×N (M and N are the number of feature points on both sides, respectively). Set the entropy regularization coefficient λ to, for example, 0.1. This value is adjusted according to the feature dimension; the higher the dimension, the smaller the value of λ. Initialize the transmission matrix T as an all-one matrix. The iterative process is performed as follows: First, keep the transmission matrix T unchanged and update the row normalization factor vector u. The specific operations are as follows: For each feature point i in the first distribution (i ranges from 1 to M), calculate the sum of the products of all elements in the i-th row of the transfer matrix T and the corresponding elements of the column normalization factor vector v. Use this sum as the denominator, and take the reciprocal of the denominator as the i-th element value of u. Then, update the column normalization factor vector v based on the updated row normalization factor vector u. The specific operations are as follows: For each feature point j in the second distribution (j ranges from 1 to N), calculate the sum of the products of all elements in the j-th column of the transfer matrix T and the corresponding elements of the row normalization factor vector u. Use this sum as the denominator, and take the reciprocal of the denominator as the j-th element value of v. This alternating update process is executed cyclically until the change in the transfer matrix generated by two adjacent iterations satisfies the preset convergence condition. Alternating iterations continue until the rate of change of the Frobenius norm of the transfer matrix between two adjacent iterations is less than, for example, 0.01%. The final Wasserstein distance calculation formula is: sum(T*C), which is the sum of the element-wise products of the transfer matrix and the cost matrix, with the result rounded to three decimal places. An upper limit of 1000 iterations is set to prevent infinite loops.

[0030] When the fractal dimension change rate exceeds the topographic abrupt change threshold and the transmission cost distance exceeds the distribution difference threshold, the current patch is marked as a transition zone patch. The topographic abrupt change threshold is set, for example, to 25%, which is derived from statistical analysis of typical regional samples with a topographic slope change rate greater than 15 degrees, specifically calculated using the median of the fractal dimension change rate of the sample region. The distribution difference threshold is set, for example, to 1.5, which is determined based on the 75th percentile of the Wasserstein distance distribution of typical cultivated-non-cultivated land sample pairs. The dual threshold judgment uses a logical AND operation, meaning that both of the following conditions must be met simultaneously: the fractal dimension change rate is greater than the topographic abrupt change threshold, and the transmission cost distance is greater than the distribution difference threshold. For patches that meet the conditions, a "transition zone" marker field is added to the patch attribute table, with the field value being a Boolean true value. The threshold parameters are stored in a configuration file and can be dynamically adjusted according to regional geographic characteristics, with an adjustment step size of, for example, 5%.

[0031] All calculations are performed within the feature space of the preceding CNN model. The feature space coordinate system is mapped to the original image geographic coordinate system via an affine transformation matrix, the parameters of which are extracted from the network structure of the preceding CNN model. The boundary region range dynamically adapts to the patch size; small patches use a fixed pixel width, while large patches use a percentage width. When the number of feature points is insufficient, an automatic resampling mechanism is triggered to ensure the effectiveness of the analysis.

[0032] S3. Extract the feature activation trajectories of transition zone patches in the previous CNN model, construct a spatial topology network based on the feature correlation strength between convolutional layers, and calculate the error propagation weight coefficients. The specific implementation is as follows: To obtain the feature maps of transition zone patches from the outputs of all convolutional layers in the preceding CNN model, the specific steps are as follows: For each region marked as a transition zone patch, extract all pixels within its boundary coordinate sequence. Iterate through the feature maps output by each convolutional layer in the preceding CNN model. The feature map size decreases with network depth; for example, the first layer is 512×512 pixels, and the last layer is 32×32 pixels. For each convolutional layer, based on the correspondence between the original image coordinates of the pixels and the feature map downsampling ratio, determine the corresponding position on the feature map using bilinear interpolation, and obtain the feature vector at that position. The preceding CNN model contains, for example, 5 convolutional layers, with each feature map having, for example, 64 to 512 channels, and feature values ​​ranging from, for example, -5.0 to +5.0. All feature maps are stored as a three-dimensional tensor data structure in convolutional layer order, with tensor dimensions of layer number × height × width × number of channels. When the patch area is less than, for example, 100 pixels, a full-pixel calculation mode is used to avoid information loss.

[0033] The feature activation values ​​corresponding to the transition zone patch positions in the feature map of each convolutional layer are extracted to form a feature activation trajectory sequence. Specifically, for each pixel within the transition zone patch, feature vectors at the same spatial position are extracted from the shallowest to the deepest layers according to the convolutional layer order. The dimension of the feature vectors is consistent with the number of channels in the corresponding convolutional layer. The feature vectors of the same pixel in different convolutional layers are concatenated in layer order to form the feature activation trajectory sequence for that pixel. For example, when there are 5 convolutional layers, each feature activation trajectory sequence contains 5 feature vectors. The same operation is performed on all pixels within the transition zone patch region, ultimately generating a three-dimensional data matrix with dimensions of number of pixels × number of convolutional layers × number of channels, where each element is the feature activation value of a specific pixel in a specific convolutional layer. The feature activation values ​​retain the original data precision, such as 32-bit single-precision floating-point format. Spatial coordinate indices are appended when storing the feature activation trajectory sequence.

[0034] The mutual information value between the feature activation trajectory sequences of adjacent convolutional layers is calculated as the feature association strength between convolutional layers. The specific steps are as follows: First, for each pair of adjacent convolutional layers, extract the feature vector pairs corresponding to all pixels in the two layers. The feature vectors are discretized preprocessed, dividing continuous feature values ​​into, for example, 16 equal-width intervals. The joint probability distribution is calculated: traversing all pixels, the frequency of the feature values ​​of the two layers falling into each interval combination is counted, the frequency value is smoothed by adding 1, and then divided by the total number of pixels to obtain the joint probability distribution. The edge probability distribution of each layer is calculated: the frequency ratio of the first layer feature value in each interval and the frequency ratio of the second layer feature value in each interval are counted, and smoothed by adding 1. The mutual information value is calculated as follows: summing all interval combinations [joint probability × natural logarithm (joint probability ÷ (first layer edge probability × second layer edge probability))]. Finally, the mutual information value between adjacent layers is obtained, with a range of, for example, between 0.15 and 0.85, retaining three decimal places. When the feature dimension is high, the number of intervals is automatically expanded to, for example, 32 to ensure statistical significance.

[0035] Using convolutional layers of a previous CNN model as nodes and feature association strength as edge weights, a spatial topological network is constructed. Specifically, each convolutional layer is abstracted as a network node, with node numbers incrementing from 1 according to the network layer sequence. Undirected edges are created between each pair of adjacent nodes, with edge weights calculated based on the mutual information between the two layers. For example, five convolutional layers form nodes 1 to 5, and four edges are established between nodes 1-2, 2-3, 3-4, and 4-5. Non-adjacent layers are not directly connected, and the network topology is a linear chain graph. Node attributes include the convolutional layer number and feature map size, while edge attributes include the mutual information value and the layer pair identifiers. The topological network is stored as an adjacency matrix, where matrix elements represent the connection weights between layers, with non-adjacent layers having a weight of 0.

[0036] In a spatial topological network, the centrality value of each node is calculated using the eigenvector centrality algorithm. The specific iterative process is as follows: Initialize the centrality value of all nodes to 1.0, forming an initial centrality vector. Construct an adjacency matrix W, with dimensions equal to the number of nodes multiplied by the number of nodes. The iterative update formula is: New centrality vector = W × Current centrality vector. After each iteration, the centrality vector is normalized using the L2 norm, with the normalization factor being the square root of the sum of the squares of the vector's elements. Convergence conditions are set as follows: the Euclidean distance between the centrality vectors of two adjacent iterations is less than, for example, 0.001, or the number of iterations reaches an upper limit, for example, 100. The final centrality value characterizes the relative importance of a node in the network, and the calculation result is retained to four decimal places. For example, the centrality value of nodes in the middle layer is usually higher than that of nodes in the edge layer. When the network has only one node, the centrality value is directly set to 1.0.

[0037] The calculated centrality values ​​are normalized to obtain the error propagation weight coefficients. Specifically, the centrality values ​​of all nodes in the network are summed. The weight coefficient for each node is then calculated using the linear normalization formula: Node weight coefficient = Node centrality value ÷ Sum of centrality values ​​of all nodes. For example, if there are 5 nodes and the sum of their centrality values ​​is 2.5, and a node's centrality value is 0.5, then its weight coefficient is 0.2. Finally, the error propagation weight coefficients for each convolutional layer are output, with a total sum of 1.0000. The weight coefficients are stored as floating-point vectors, with the vector index corresponding to the convolutional layer number, used in subsequent weighted calculations. Normalization ensures that the weight coefficients conform to probability distribution characteristics, supporting the reasonable distribution of errors during backpropagation.

[0038] All calculations are based on the feature activation values ​​of pixels within the transition zone patches. When the patch area is too small, full-pixel calculation mode is automatically enabled. The number of feature discretization intervals is dynamically adjusted according to the sample size: 8 intervals for less than 1000 samples, 16 intervals for 1000-5000 samples, and 32 intervals for more than 5000 samples. Mutual information calculation uses a smoothing process with increments of 1 to avoid zero-probability issues. The centrality algorithm has a maximum number of iterations to prevent non-convergence; if convergence is not achieved after more than 100 iterations, the result of the last iteration is taken and normalized. The weight coefficient vector is appended with a timestamp and a source patch identifier, supporting a mechanism for tracking calculation results and dynamic updates.

[0039] S4. Extract the boundary-related channel response values ​​from the feature map output by the previous CNN model, and calculate the weighted feature offset by combining the error propagation weight coefficients. The specific implementation is as follows: The process of identifying target channels relevant to the boundary detection task in the feature map output by the preceding CNN model is as follows: Analyze the semantic meaning of each channel in the feature map output by the last convolutional layer of the preceding CNN model. Calculate the contribution of each channel to the classification of boundary pixels using a gradient-based activation mapping method. This method includes: constructing a cross-entropy loss function for the boundary detection task in the model output layer; calculating the absolute gradient value of each channel in the feature map through backpropagation; and calculating the average of the absolute gradient values ​​at all positions for each channel. Select the top 10% of channels with the highest average gradient values ​​as the candidate channel set. Further verify the boundary response characteristics of the candidate channels: on a standard farmland boundary sample set, calculate the intersection-union ratio (IU) of each channel's response value with the real boundary mask, and retain channels with an IU greater than, for example, 0.3. The final determined target channel numbers are stored as an index list, such as channel 12, channel 35, channel 48, etc. The target channel identification process is completed before model deployment, and the identification results are stored in a configuration file for subsequent use.

[0040] The feature response values ​​of the target channels at their corresponding spatial locations in the transition zone patches are extracted as follows: A binary spatial location mask is generated based on the boundary coordinate sequence of the transition zone patches, with pixel values ​​of 1 within the mask area and 0 outside the area. The feature maps of each target channel are traversed, and the spatial resolution of the feature maps is aligned to the original image resolution using bilinear interpolation. Within the transition zone patch area, the feature values ​​corresponding to the locations where the mask value is 1 are extracted. The feature response value at each spatial location is a scalar, representing the activation intensity of that point in the target channel. For multiple target channels, their respective feature response value matrices are extracted, with the matrix dimension being the number of pixels in the patch × the number of target channels. The feature response values ​​retain their original numerical range, for example, between -3.5 and +4.2. When a patch spans multiple feature map blocks, patch stitching is performed to ensure spatial continuity.

[0041] The extracted feature response values ​​are multiplied with the error propagation weight coefficients using a weighted multiplication operation. The specific steps are as follows: First, obtain the error propagation weight coefficient vector generated in step S3. The vector length should match the number of convolutional layers in the preceding CNN model; for example, 5 convolutional layers correspond to a vector of length 5. For each target channel, determine its corresponding convolutional layer index. Extract the coefficient values ​​of the corresponding convolutional layer from the weight coefficient vector. Then, perform a broadcast multiplication operation on the feature response value matrix along the channel dimension with the weight coefficients: the feature response value of each channel is multiplied by the weight coefficient of the convolutional layer to which that channel belongs. The calculation formula is: Weighted Feature Response Value = Original Feature Response Value × Error Propagation Weight Coefficient [Convolutional Layer Index]. After the operation, the dimension of the feature response value matrix remains unchanged, while the numerical range is scaled proportionally. The weighted multiplication operation supports GPU parallel acceleration, improving processing efficiency.

[0042] The weighted multiplication result is subjected to mean pooling in the spatial dimension. Specifically, the weighted feature response value matrix is ​​split into multiple two-dimensional matrices according to the target channel dimension. Each matrix corresponds to the spatial feature response distribution (height × width) of a target channel. For each target channel's feature response value two-dimensional matrix, the arithmetic mean of the feature values ​​at all spatial locations is calculated. The mean calculation formula is: Channel feature offset = Sum of feature response values ​​at all spatial locations ÷ Total number of spatial locations. The total number of spatial locations equals the number of valid pixels within the transition zone patch, excluding invalid value markers. This is calculated independently for each target channel, ultimately generating a feature offset vector in the channel number dimension. For example, when three target channels are identified, the output is a feature offset vector containing three elements. Mean pooling preserves the original numerical precision, such as 32-bit single-precision floating-point format.

[0043] The mean pooling result is output as a weighted feature offset. Specifically, the feature offsets calculated for each target channel are arranged in ascending order of channel number to construct a one-dimensional feature offset vector. The order of vector elements matches the target channel index list. The output vector is standardized by calculating the arithmetic mean and standard deviation of the vector elements and performing z-score standardization. The formula is: Standardized value = (Original value - Mean) ÷ Standard deviation. The standardized feature offset vector is the final output weighted feature offset, stored as a floating-point array. Metadata information, including the source patch identifier, calculation timestamp, and target channel list, is appended to the weighted feature offset. The output result is transmitted to subsequent processing stages via a data interface, and the data is encapsulated using a binary serialization format such as Protocol Buffers.

[0044] All operations are based on the feature space data of the preceding CNN model, and the feature map resolution maintains a geometric correspondence with the original image. A caching mechanism is implemented in the target channel recognition process, requiring only one recognition operation for the same model configuration. The weighted multiplication operation supports dynamic weight updates, automatically triggering recalculation when new error propagation weight coefficients are generated in step S3. Mean pooling automatically excludes invalid pixel locations, triggering an anomaly alarm when the proportion of valid pixels falls below, for example, 50%. Output normalization uses sliding window statistical parameters, with the window containing, for example, the calculation results of the 100 most recently processed patches, ensuring parameter timeliness. Invalid pixels include specially marked locations such as cloud-covered areas and shadowed areas.

[0045] S5. After generating an error compensation mask based on the weighted feature offset, perform pixel-level weighted fusion with the input feature map of the subsequent CNN model to generate a fused feature map. The specific implementation is as follows: The weighted feature offsets are converted into compensation intensity coefficients. The specific implementation process is as follows: Obtain the weighted feature offset vector output from step S4. This vector contains, for example, three elements corresponding to the offsets of the target channels. Perform a Sigmoid function transformation on each element of the vector, calculated as 1 divided by (1 plus the negative power of the natural exponent), mapping the output value to the interval 0 to 1. Calculate the arithmetic mean of the transformed vector elements to obtain the scalar compensation intensity coefficient. For example, when the weighted feature offsets are [0.8, -0.2, 1.5], the Sigmoid transformation results in [0.69, 0.45, 0.82], with an average value of 0.65. The compensation intensity coefficients retain two decimal places and are stored as floating-point variables. When the calculated result is greater than, for example, 0.9, it is automatically truncated to 0.9 to prevent overcompensation. The conversion process is executed on the GPU to accelerate computational efficiency.

[0046] Create a blank mask matrix with the same dimensions as the input feature map of the subsequent CNN model. The specific steps are as follows: Obtain the dimensions of the input feature map of the first convolutional layer of the subsequent CNN model, including height, width, and number of channels, for example, 512 pixels high, 512 pixels wide, and 64 channels. Initialize the blank mask matrix as an all-zero matrix, with dimensions identical to the input feature map (height × width × number of channels). Set the matrix data type to 32-bit floating-point and use the last channel format for memory layout. Record the feature map's spatial resolution and geographic coordinate reference information during creation to ensure alignment with the original image coordinate system. Automatically trigger a re-creation mechanism when the feature map size changes.

[0047] The process of filling the regions corresponding to transition zone patches in the blank mask matrix with compensation intensity coefficients involves the following steps: First, based on the boundary coordinate sequence of the transition zone patches, generate a binary mask for the polygonal region, where pixels within the region have a value of 1 and pixels outside the region have a value of 0. Then, iterate through each spatial position in the mask matrix; when the corresponding mask value is 1, fill that position with compensation intensity coefficients in all channels of the blank mask matrix. For example, if the compensation intensity coefficient is 0.65, set the values ​​of all 64 channels at the coordinate position to 0.65. The filling operation is optimized using parallel computation, processing in rows and blocks to accelerate large-scale matrix operations. When multiple transition zone patches exist, fill their respective regions sequentially according to their identifiers.

[0048] In the blank mask matrix, non-transition zone patch locations are filled with fixed values. Specifically, the method is as follows: traverse all spatial positions in the blank mask matrix; where the transition zone patch mask value is 0, fill all channels with a fixed value of 1.0. The fixed value is selected based on the multiplicative identity property to ensure that the original feature values ​​are not affected. The filling operation is performed after the transition zone region filling is complete, using a matrix-wide assignment optimization strategy. The fixed value can be configured to other values ​​within the range of 0.8 to 1.2, with a default value of 1.0. After filling, data verification is performed by randomly sampling, for example, 100 locations to verify the correctness of the filling.

[0049] The filled blank mask matrix is ​​used as the error compensation mask. Specifically, after filling all locations, the consistency of the matrix data is verified. The matrix is ​​encapsulated as a tensor data structure, with additional metadata including: feature map size, filling timestamp, compensation intensity coefficient value, and transition zone patch identifier. The error compensation mask is stored as an HDF5 format file, supporting fast read and write operations. A final verification is performed before output: the transition zone region value equals the compensation intensity coefficient, and the non-transition zone region value equals a fixed value. Regeneration is triggered when the error rate exceeds 0.1%.

[0050] The error compensation mask is multiplied pixel-wise with the input feature map of the subsequent CNN model. Specifically, the input feature map of the current batch of the subsequent CNN model is loaded, with a size strictly matching the error compensation mask. Element-wise multiplication is performed for each channel k at each position (a, b) according to spatial coordinate indices: the multiplication result [a, b, k] = input feature map [a, b, k] × error compensation mask [a, b, k]. The multiplication operation is executed in parallel on the GPU, using CUDA kernel functions to optimize computational efficiency. When the feature map sizes do not match, bilinear interpolation alignment is triggered. The multiplication result retains the original numerical precision.

[0051] The multiplication result is added to the input feature map of the subsequent CNN model at corresponding pixel positions. Specifically, while maintaining the same spatial coordinate indices, element-wise addition is performed on each channel k at each position (a, b): Addition result [a, b, k] = Multiplication result [a, bj, k] + Input feature map a, b, k. The addition operation is performed immediately after the multiplication operation to avoid intermediate result storage overhead. An overflow protection mechanism is implemented: when the absolute value of the calculation result is greater than, for example, 1000, it is automatically scaled to the range of -1.0 to +1.0. The addition operation is optimized for parallel processing using SIMD instructions.

[0052] The summation result is output as a fused feature map. Specifically, the addition result matrix is ​​encapsulated as a feature map tensor with the same dimensions as the input feature map. Data standardization is performed: the mean and standard deviation of the entire feature map are calculated, and a z-score transformation is applied ((original value - mean) ÷ standard deviation). The standardization parameters are dynamically calculated based on the current training batch. The fused feature map output is in a tensor format acceptable to subsequent CNN models. An additional gradient calculation hook is added to the output to support backpropagation. Finally, it is stored as a binary file and transmitted to the model computation pipeline. The file header contains the feature map size, data type, and spatial reference information.

[0053] All operations are performed in the feature map space, where spatial coordinates are associated with geographic coordinates through affine transformations. A threshold limit is set for the compensation intensity coefficient conversion. Mask filling operations support batch processing of multiple patches. Feature map operations employ a double-buffer mechanism: the next batch of data is loaded asynchronously while the current batch is being processed. The output fused feature map enables a checksum mechanism, using, for example, CRC32 to verify data transmission integrity. When processing large-scale regions, automatic block processing is performed, with a block size set to, for example, 512×512 pixels.

[0054] S6. Utilize the fused feature map to perform arable land potential assessment, outputting slope calculation correction values ​​and reserve resource classification results. The specific implementation is as follows: The fused feature map is input into the evaluation network layer of the subsequent CNN model. The specific implementation process is as follows: Obtain the fused feature map tensor output from step S5, which contains spatial and channel dimension information. The evaluation network layer is located at the end of the subsequent CNN model and consists of, for example, three convolutional layers and two fully connected layers. The convolutional layers use 1×1 pixel kernels to extract spatially invariant features, and the fully connected layers perform feature compression. A dimension matching check is performed before input: if the spatial dimension of the fused feature map is inconsistent with the input requirements of the evaluation network layer, a bilinear interpolation algorithm is triggered to adjust the spatial dimension to a matching state. The evaluation network layer loads pre-trained parameters, which are obtained through training on a historical farmland evaluation dataset, using a mean squared error loss function and the Adam optimizer during training. The feature map input process uses batch processing mode, with a batch size set to, for example, 16, supporting GPU-accelerated computation. The input data channel order is converted to the format required by the evaluation network layer.

[0055] The evaluation network layers extract topographic slope and farmland contiguousness feature response values. The specific steps are as follows: Feature values ​​are obtained from specific output nodes of the evaluation network layers. The topographic slope feature response value originates from the first output node of the last fully connected layer of the network; the weights of this node are obtained through optimized training using slope-labeled samples. The farmland contiguousness feature response value originates from the second output node of the same fully connected layer; its weights are obtained through optimized training using contiguousness-labeled samples. Feature response values ​​are scalar floating-point numbers, with slope response values ​​ranging from, for example, -5.0 to +5.0, and contiguousness response values ​​ranging from 0 to 10.0. The extraction process records node indices and response value timestamps to ensure traceability. When a response value exceeds the normal range, such as a slope response value less than -6.0 or greater than +6.0, an anomaly detection mechanism is triggered and logged.

[0056] The slope characteristic response value is linearly scaled to obtain the slope calculation correction value. The specific steps are as follows: Obtain the slope characteristic response value (raw_value). Calculate the correction value using the linear transformation formula: Correction value = q × raw_value + p. Parameters q and p are determined through standard slope sample calibration: collect, for example, 100 known slope measurement points, and establish a linear regression model between the measured values ​​and the characteristic response value. The regression equation is Y = qX + p, where X is the characteristic response value and Y is the actual slope value. For example, when q = 0.25 and p = 0.5, a response value of 2.0 corresponds to a correction value of 1.0. The correction value is in degrees, retaining one decimal place. Boundary protection is implemented during the conversion process: a calculation result less than 0 is set to 0 degrees, and a result greater than 45 degrees is set to 45 degrees. The conversion parameters are stored in a configuration file, supporting dynamic adjustment of parameter groups according to different terrain types such as plains, hills, and mountains.

[0057] The contiguousness characteristic response value of cultivated land is input into a preset hierarchical rule base. Specifically, the hierarchical rule base is a key-value pair data structure that stores the mapping relationship between the contiguousness value range and the hierarchical results. The rule base contains, for example, five threshold intervals: 0 to less than 2.0 corresponds to Level 1 backup resources, 2.0 to less than 4.0 corresponds to Level 2 backup resources, 4.0 to less than 6.0 corresponds to Level 3 backup resources, 6.0 to less than 8.0 corresponds to Level 4 backup resources, and 8.0 to 10.0 corresponds to Level 5 backup resources. After inputting the characteristic response value, the rule base is traversed for interval matching: when the response value satisfies the condition that the lower limit ≤ value < upper limit, the corresponding hierarchical result is returned. The matching algorithm uses a binary search method to optimize efficiency, with a time complexity of O(log n). The rule base supports dynamic updates; when agricultural policies change, new rules are imported through the management interface, and update operations require security verification such as password authentication.

[0058] Based on threshold interval matching in the hierarchical rule base, the hierarchical results of backup resources are determined. Specifically, after performing interval matching, a hierarchical identifier is obtained. A confidence assessment is added to the hierarchical results: the distance between the current response value and the median of the matching interval is calculated; the smaller the distance, the higher the confidence. The confidence calculation formula is: 1 - |Response Value - Interval Median| ÷ Interval Width. For example, in a level 3 interval of 4.0-6.0, where the median is 5.0, a response value of 5.2 has a confidence of 0.9. The hierarchical results are encapsulated as structured data, containing the hierarchical identifier, confidence value, and original response value. When the response value exceeds the rule base range, such as being less than 0 or greater than 10.0, the nearest neighbor interval result is returned and marked as an estimated value. A confidence level below, for example, 0.5, triggers a manual review process.

[0059] The output includes slope calculation correction values ​​and backup resource grading results. The specific steps are as follows: Slope calculation correction values ​​are converted to JSON format fields, containing the value, unit, and precision information (e.g., {"value":12.5,"unit":"degree","precision":0.1}). Backup resource grading results are converted to structured objects (e.g., {"level":"Level 3","confidence":0.92,"raw_value":5.2}). Spatial location information is appended to the output data packet, associating geographic regions through the coordinate sequence of map patch boundaries. The final output is written to a geographic database and stored as spatial attribute table records. Each record contains a map patch identifier, slope correction value, and grading result. A farmland potential evaluation report document is also generated, including statistical charts and textual analysis. The report template supports customization. Data validation is performed before output: slope values ​​must be within the 0-45 degree range and grading identifiers must be in a preset set. An alarm is triggered if validation fails.

[0060] All calculations are performed at the evaluation network layer, and network parameters are initialized through transfer learning. A memory mechanism is used for slope correction value transformation: the average of the most recent 10 transformation results is used as a reference benchmark. Rule base matching implements multi-level caching: hot rules reside in memory, while all rules are stored on disk. When processing batches of polygons, parallel computing optimization is enabled, with the number of threads set to, for example, 80% of the CPU cores. Evaluation report generation uses a template engine, and templates support customization. The output interface supports multiple protocols, including HTTP API and direct database connection. Data storage employs a dual backup mechanism: the primary copy is stored in a relational database, and the backup copy is stored in a distributed file system.

[0061] It is worth noting that the pre-stage CNN model refers to the convolutional neural network used for initial feature extraction. Its structure adopts existing publicly available encoder architectures (such as variants of VGGNet or ResNet), initialized with pre-trained weights, and is responsible for extracting multi-scale boundary features from the input image. The post-stage CNN model refers to a dedicated network for farmland evaluation, employing a combination of fully connected layers and regression layers. Its input size is aligned with the output of the pre-stage model, and its parameters are fine-tuned through transfer learning. It is responsible for mapping features to slope and contiguousness indices. The cascading relationship between the two forms a processing pipeline: the pre-stage focuses on local boundary feature detection (such as enhancing edges through the Sobel operator), and the post-stage performs global semantic parsing (such as integrating context through spatial pyramid pooling).

[0062] Example 2: Figure 2 A schematic diagram of the structure of the arable land reserve resource survey and evaluation system based on convolutional neural networks of the present invention is given. The arable land reserve resource survey and evaluation system based on convolutional neural networks includes: The image processing module is used to acquire remote sensing image data of the area to be evaluated, and to perform non-cultivated patch segmentation processing through the front-end CNN model to generate an initial set of segmented patches. The patch marking module is used to mark the boundary regions of each patch in the initial segmented patch set as transition zone patches when the rate of change of the fractal dimension of the boundary pixels relative to the standard cultivated land boundary exceeds the terrain abrupt change threshold and the transmission cost distance based on the optimal transmission theory of the feature distribution on both sides of the boundary exceeds the distribution difference threshold. The weight calculation module is used to extract the feature activation trajectories of transition zone patches in the previous CNN model, construct a spatial topology network based on the feature correlation strength between convolutional layers, and calculate the error propagation weight coefficients. The offset calculation module is used to extract the boundary-related channel response values ​​from the feature map output by the previous CNN model, and calculate the weighted feature offset by combining the error propagation weight coefficients. The feature fusion module is used to generate an error compensation mask based on the weighted feature offset and then perform pixel-level weighted fusion with the input feature map of the subsequent CNN model to generate a fused feature map. The potential assessment module is used to perform farmland potential assessment using fused feature maps, and output slope calculation correction values ​​and reserve resource classification results.

[0063] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0064] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0065] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center containing one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0066] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0067] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0068] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0069] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0070] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0071] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0072] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for investigating and evaluating arable land reserve resources based on convolutional neural networks, characterized in that, include: S1. Obtain remote sensing image data of the area to be evaluated, and perform non-cultivated map patch segmentation processing through the pre-stage CNN model to generate an initial set of segmented patches. S2. For the boundary regions of each patch in the initial segmented patch set, when the rate of change of the fractal dimension of the boundary pixel relative to the standard cultivated land boundary exceeds the terrain abrupt change threshold and the transmission cost distance of the feature distribution on both sides of the boundary based on the optimal transmission theory exceeds the distribution difference threshold, the corresponding patch is marked as a transition zone patch. S3. Extract the feature activation trajectories of transition zone patches in the previous CNN model, construct a spatial topology network based on the feature correlation strength between convolutional layers, and calculate the error propagation weight coefficients, including: Obtain the feature maps of the pixels in the transition zone from the output of all convolutional layers of the previous CNN model; Extract the feature activation values ​​at the corresponding transition zone patch positions in the feature map of each convolutional layer to form a feature activation trajectory sequence; Calculate the mutual information value between the feature activation trajectory sequences of adjacent convolutional layers as the feature correlation strength between convolutional layers; Using the convolutional layers of the previous CNN model as nodes and the feature association strength as the edge weights, a spatial topology network is constructed. In spatial topological networks, the centrality value of each node is calculated using the eigenvector centrality algorithm; The calculated centrality values ​​are normalized to obtain the error propagation weight coefficients; S4. Extract the boundary-related channel response values ​​from the feature map output by the previous CNN model, and calculate the weighted feature offset by combining the error propagation weight coefficients. S5. After generating an error compensation mask based on the weighted feature offset, perform pixel-level weighted fusion with the input feature map of the subsequent CNN model to generate a fused feature map, including: The weighted feature offset is converted into a compensation intensity coefficient. The compensation intensity coefficient is obtained by: obtaining a weighted feature offset vector, which contains the offset of the target channel corresponding to each element; performing a sigmoid function transformation on each element of the weighted feature offset vector to map the output value to the interval between 0 and 1; and calculating the arithmetic mean of the transformed vector elements to obtain the scalar compensation intensity coefficient. Create a blank mask matrix with the same size as the input feature map of the subsequent CNN model; Fill the corresponding regions of the transition zone patches in the blank mask matrix with compensation intensity coefficients; Fill the non-transition zone patch locations in the blank mask matrix with fixed values; Use the filled blank mask matrix as the error compensation mask; The error compensation mask is multiplied with the input feature map of the subsequent CNN model according to the pixel position; The multiplication result is added to the input feature map of the subsequent CNN model according to the pixel position. The summation result is output as a fused feature map; S6. Use the fused feature map to perform farmland potential evaluation, and output the slope calculation correction value and the reserve resource classification results.

2. The method for investigating and evaluating arable land reserve resources based on convolutional neural networks according to claim 1, characterized in that, Remote sensing image data of the area to be evaluated is acquired, and non-cultivated land parcels are segmented using a pre-stage CNN model to generate an initial set of segmented parcels, including: Acquire multispectral remote sensing image data of the area to be evaluated; Preprocessing of multispectral remote sensing image data; The preprocessed multispectral remote sensing image data is input into the pre-stage CNN model, and the pre-stage CNN model outputs a binary segmentation mask for non-farmland and farmland. Connectivity analysis is performed on the binary segmentation mask to generate an initial set of segmentation patches, where each patch is defined by a sequence of boundary coordinates.

3. The method for investigating and evaluating arable land reserve resources based on convolutional neural networks according to claim 2, characterized in that, Preprocessing includes radiometric calibration, atmospheric correction, and geometric registration.

4. The method for investigating and evaluating arable land reserve resources based on convolutional neural networks according to claim 1, characterized in that, For each boundary region of a patch in the initial segmented patch set, determine the transition zone patches that mark the corresponding patches, including: Extract the feature space coordinates of pixels in the last convolutional layer of the preceding CNN model from the boundary regions of each patch in the initial segmentation patch set; Based on the extracted feature space coordinates, the fractal dimension value of the pixel feature vector of the boundary region is calculated by box counting. The calculated fractal dimension value is compared with the fractal dimension benchmark value of the standard cultivated land boundary to obtain the fractal dimension change rate; Extract the set of feature vectors from the last convolutional layer of the preceding CNN model from the cultivated land side of the map boundary; Extract the set of feature vectors from the pixel output of the last convolutional layer of the previous CNN model from the non-cultivated land side of the map boundary; The feature vector set on the cultivated land side of the boundary and the feature vector set on the non-cultivated land side of the boundary are respectively used as discrete probability distributions; Transmission cost distance is generated based on the optimal transmission theory derived from two discrete probability distributions; When the fractal dimension change rate exceeds the terrain mutation threshold and the transmission cost distance exceeds the distribution difference threshold, the current patch is marked as a transition zone patch.

5. The method for investigating and evaluating arable land reserve resources based on convolutional neural networks according to claim 4, characterized in that, The transmission cost distance is obtained by calculating the Wasserstein distance using the entropy-regularized Sinkhorn iterative algorithm.

6. The method for investigating and evaluating arable land reserve resources based on convolutional neural networks according to claim 1, characterized in that, Extract the boundary-related channel response values ​​from the feature map output by the previous CNN model, and calculate the weighted feature offset by combining the error propagation weight coefficients, including: Identify target channels related to the boundary detection task in the feature map output by the previous CNN model; Extract the feature response values ​​of the target channel at the corresponding spatial location of the transition zone patch; The extracted feature response values ​​are multiplied by the error propagation weight coefficients using a weighted multiplication operation. The weighted multiplication result is subjected to mean pooling in the spatial dimension; The result of mean pooling is output as the weighted feature offset.

7. The method for investigating and evaluating arable land reserve resources based on convolutional neural networks according to claim 1, characterized in that, The fused feature map is used to perform farmland potential assessment, outputting slope calculation correction values ​​and reserve resource classification results, including: The fused feature maps are then input into the evaluation network layer of the subsequent CNN model; The response values ​​of terrain slope and farmland contiguousness were extracted by evaluating the network layers. The terrain slope characteristic response value is linearly scaled and transformed to obtain the slope calculation correction value; Input the characteristic response value of farmland contiguousness into the preset hierarchical rule library; The backup resource classification result is determined by matching the threshold range in the classification rule base; Output the slope calculation correction value and the backup resource classification results.

8. A system for investigating and evaluating reserve farmland resources based on convolutional neural networks, used to implement the method for investigating and evaluating reserve farmland resources based on convolutional neural networks as described in any one of claims 1-7, characterized in that, include: The image processing module is used to acquire remote sensing image data of the area to be evaluated, and to perform non-cultivated patch segmentation processing through the front-end CNN model to generate an initial set of segmented patches. The patch marking module is used to mark the boundary regions of each patch in the initial segmented patch set as transition zone patches when the rate of change of the fractal dimension of the boundary pixels relative to the standard cultivated land boundary exceeds the terrain abrupt change threshold and the transmission cost distance based on the optimal transmission theory of the feature distribution on both sides of the boundary exceeds the distribution difference threshold. The weight calculation module is used to extract the feature activation trajectories of transition zone patches in the previous CNN model, construct a spatial topology network based on the feature correlation strength between convolutional layers, and calculate the error propagation weight coefficients. The offset calculation module is used to extract the boundary-related channel response values ​​from the feature map output by the previous CNN model, and calculate the weighted feature offset by combining the error propagation weight coefficients. The feature fusion module is used to generate an error compensation mask based on the weighted feature offset and then perform pixel-level weighted fusion with the input feature map of the subsequent CNN model to generate a fused feature map. The potential assessment module is used to perform farmland potential assessment using fused feature maps, and output slope calculation correction values ​​and reserve resource classification results.

Images