A cultivated land resource protection evaluation method and system based on big data analysis

Abstract

The application discloses a cultivated land resource protection evaluation method and system based on big data analysis, which comprises the following steps: S1, acquiring multi-period remote sensing image data and cultivated land vector boundary data; S2, generating a cultivated land segmentation map; S3, inputting an improved PerceiverIO model, introducing a cross-temporal and cross-spatial stability mechanism, and generating cross-period cultivated land potential features; S4, calculating the feature distance between different time cultivated land plots based on the cross-period cultivated land potential features, and constructing a cross-period plot matching cost matrix; S5, performing iterative solving by combining an optimal transport algorithm and a Sinkhorn iteration algorithm; S6, constructing a plot correction matrix and generating stable plot units; and S7, constructing a comprehensive evaluation index and generating a cultivated land resource protection evaluation result. The application realizes stable identification and accurate correspondence of multi-period cultivated land plots, improves the cultivated land boundary stability analysis capability and the reliability of the cultivated land resource protection evaluation result.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image analysis and land resource monitoring technology, and in particular to a method and system for evaluating the protection of arable land resources based on big data analysis. Background Technology

[0002] With the increasing demand for remote sensing observation technology and land resource monitoring, the analysis and protection assessment technology of arable land resource changes based on multi-period remote sensing imagery has received widespread attention. Existing arable land monitoring and assessment methods mainly rely on the classification results of single-period remote sensing images or simple area statistics to analyze changes in arable land resources. However, these methods generally suffer from the following problems in practical applications:

[0003] Remote sensing images are easily affected by factors such as sensor differences, changes in observation angles, and atmospheric conditions during acquisition at different times. This leads to radiometric differences and spatial offsets between images, causing significant changes in the boundary position and texture information of the same cultivated land plot in multiple images. This reduces the stability of cultivated land identification results and affects the reliability of cross-period cultivated land change analysis. Existing cultivated land identification methods mostly use traditional image classification or single deep learning models for pixel-level identification, lacking the ability to jointly model the boundary structure features, texture features, and spatial structure features of cultivated land plots. Under complex surface environments or image noise interference, missegmentation or plot boundary breakage are prone to occur, making it difficult to accurately reflect the true structure of cultivated land plots. In the process of multi-period cultivated land plot matching and change analysis, traditional matching methods based on geometric distance or simple feature differences cannot effectively handle the problems of plot morphological changes and boundary position offsets, resulting in inaccurate identification of cross-period plot correspondences. At the same time, there is a lack of stability evaluation indicators to comprehensively measure the consistency and spatial continuity of cultivated land boundaries, thus affecting the objectivity and accuracy of cultivated land resource protection evaluation results.

[0004] Therefore, how to provide a method and system for evaluating the protection of arable land resources based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for evaluating farmland resource protection based on big data analysis. This invention constructs a multi-period farmland image sequence by acquiring multi-period remote sensing image data and farmland vector boundary data. It then uses the SegFormer model to extract farmland plot boundary features, plot texture features, and plot spatial structure features to form a multi-period farmland feature set. This multi-period farmland feature set is input into an improved PerceiverIO model. By introducing a cross-temporal stabilization mechanism in the cross-attention writing layer, it generates cross-period farmland potential features. Based on these cross-period farmland potential features, a cross-period plot matching algorithm is constructed. The method calculates the inter-period plot matching matrix by combining the valence matrix with the optimal transmission algorithm and the Sinkhorn iterative algorithm. It then calculates the spatial offset vector of cultivated land plots through the inter-period plot matching matrix and constructs the plot correction matrix. Spatial correction is performed on the boundaries of cultivated land plots in multi-period cultivated land image sequences to generate stable plot units. Furthermore, based on the stable plot units, the inter-period stability index, boundary consistency index, and spatial continuity index are calculated, and a comprehensive evaluation index is constructed to generate the cultivated land resource protection evaluation results. This method has the advantages of high accuracy in inter-period cultivated land plot matching, strong ability to analyze the stability of cultivated land boundaries, and high reliability of cultivated land resource protection evaluation results.

[0006] According to an embodiment of the present invention, a method for evaluating the protection of arable land resources based on big data analysis includes the following steps:

[0007] S1. Acquire multi-period remote sensing image data and cultivated land vector boundary data, perform preprocessing on the multi-period remote sensing image data, and perform spatial cropping on the multi-period remote sensing image data based on the cultivated land vector boundary data to generate a multi-period cultivated land image sequence.

[0008] S2. Input the multi-period farmland image sequence into the SegFormer model, extract multi-scale semantic features and generate a farmland segmentation map, and extract the boundary features, texture features and spatial structure features of farmland plots based on the farmland segmentation map to form a multi-period farmland feature set.

[0009] S3. Input the set of farmland features from multiple periods into the improved PerceiverIO model. Introduce a cross-temporal stabilization mechanism in the cross-attention writing layer of the improved PerceiverIO model to map farmland features from different times to a unified latent space and generate cross-period farmland latent features.

[0010] S4. Calculate the feature distance between farmland plots at different times based on the potential features of farmland across periods, and construct the cross-period plot matching cost matrix;

[0011] S5. The inter-period land parcel matching cost matrix is ​​obtained by iteratively solving the inter-period land parcel matching matrix by combining the optimal transmission algorithm and the Sinkhorn iterative algorithm.

[0012] S6. Calculate the spatial offset vector between cultivated land parcels based on the cross-period parcel matching matrix, construct the parcel correction matrix, perform spatial correction on the boundaries of cultivated land parcels in the multi-period cultivated land image sequence, and generate stable parcel units.

[0013] S7. Calculate the intertemporal stability index, boundary consistency index, and spatial continuity index based on stable plot units, construct a comprehensive evaluation index, and generate farmland resource protection evaluation results.

[0014] Optionally, S1 specifically includes:

[0015] Acquire remote sensing image data and farmland vector boundary data acquired at different times, perform radiometric correction, geometric correction and coordinate unification processing on the remote sensing image data, and generate corrected image data;

[0016] The corrected image data is processed to unify resolution and standardize pixel values ​​to generate standard image data.

[0017] The extent of cultivated land area is determined based on cultivated land vector boundary data. Spatial cropping is then performed on standard image data according to the cultivated land area extent to generate cropped image data.

[0018] The cropped image data were arranged according to the acquisition time of the remote sensing image data to construct a multi-period farmland image sequence.

[0019] Optionally, S2 specifically includes:

[0020] The multi-period farmland image sequence is divided into multiple image blocks according to a set size, and the image blocks are subjected to location encoding processing to generate image block feature sequences with location identifiers.

[0021] The image patch feature sequence is input into the SegFormer model, and hierarchical feature extraction and feature fusion processing are performed on the image patch feature sequence to generate multi-scale semantic features.

[0022] Pixel-level classification calculations are performed based on multi-scale semantic features to generate farmland segmentation maps;

[0023] Extract the set of pixels representing the boundaries of cultivated land parcels from the cultivated land segmentation map, calculate the coordinates of the parcel boundaries based on the set of pixels, and generate the parcel boundary features.

[0024] Calculate the plot texture features based on the distribution of farmland pixel values ​​in the farmland segmentation map;

[0025] Calculate the spatial structure features of land parcels based on the spatial adjacency relationship of cultivated land pixels in the cultivated land segmentation map;

[0026] The characteristics of land parcel boundaries, land parcel textures, and land parcel spatial structures are integrated according to the time sequence of the images to form a multi-period set of cultivated land characteristics.

[0027] Optionally, the improved PerceiverIO model specifically includes an input encoding layer, a cross-attention writing layer, a latent self-attention layer, and an output mapping layer;

[0028] The input encoding layer performs feature splicing processing on the plot boundary features, plot texture features and plot spatial structure features in the multi-period cultivated land feature set to generate a feature vector sequence, and performs position encoding processing on the feature vector sequence to generate an encoded feature sequence;

[0029] The dimensionality of the encoded feature sequence is transformed and a potential vector sequence is generated according to a set length.

[0030] The cross-attention writing layer introduces a cross-spatiotemporal stabilization mechanism to perform element-wise multiplication between each feature vector in the encoded feature sequence and each latent vector in the latent vector sequence, and sums the product values ​​to obtain a dot product value. The dot product values ​​are then exponentially calculated and divided by the sum of the exponent values ​​to obtain a weight value. The weight values ​​are then used to perform numerical multiplication on the feature vectors in the encoded feature sequence and summed. The calculated value is added to the latent vector value at the corresponding position in the latent vector sequence. The numerical difference between the feature vector sequences corresponding to different times is calculated, and the values ​​in the latent vector sequence are updated according to the numerical difference in a proportional reduction manner to generate a cross-period latent representation.

[0031] The latent self-attention layer performs element-wise multiplication between latent vectors in the intertemporal latent representation and sums the product values ​​to obtain the dot product value. It then performs exponential calculation on all dot product values ​​and divides them by the sum of the exponential values ​​to obtain the weight value. Finally, it uses the weight value to perform numerical multiplication on the latent vectors and sums the results to generate the latent feature representation.

[0032] The output mapping layer performs matrix multiplication calculations on the latent feature representations to generate potential features of cultivated land across different periods.

[0033] Optionally, S4 specifically includes:

[0034] Potential features of cultivated land spanning different periods are grouped and arranged according to the order of image acquisition time to form a set of potential features of plots corresponding to different times;

[0035] For the potential vectors in the potential feature set of the land parcels corresponding to different times, perform element-by-element difference calculation and square the difference values. Then, sum all the squared values ​​to obtain the distance value, which is equal to the sum of the squared differences of all elements.

[0036] Arrange the distance values ​​between the potential features of the land parcels corresponding to all time combinations, and then arrange the distance values ​​in a matrix according to the row direction representing the potential features of the land parcels corresponding to the previous time image in the time sequence and the column direction representing the potential features of the land parcels corresponding to the next time image in the time sequence to generate a cross-period land parcel matching cost matrix.

[0037] The numerical elements in the cross-period land parcel matching cost matrix are indexed and recorded according to the parcel order to obtain the feature distance matrix structure between cultivated land parcels at different times.

[0038] Optionally, S5 specifically includes:

[0039] The exponential matrix is ​​obtained by performing an exponential operation on all the numerical elements in the cross-period land parcel matching cost matrix. The row direction normalization coefficient is obtained by summing the numerical elements in each row of the exponential matrix. The row normalization matrix is ​​obtained by dividing each row numerical element by the corresponding row direction normalization coefficient.

[0040] The column-direction normalization coefficient is obtained by summing the numerical elements of each column in the row normalization matrix. The column-direction normalization coefficient is obtained by dividing each numerical element of the column by the corresponding column-direction normalization coefficient.

[0041] The column-normalized matrix is ​​then normalized again in the row direction and in the column direction. This process is repeated multiple times in an alternating manner of row and column normalization until the sum of the numerical elements in each row and the sum of the numerical elements in each column of the matrix reach the set values.

[0042] The matrix obtained from the final iteration is used as the cross-period land parcel matching matrix. The matrix is ​​arranged according to the row direction representing the cultivated land parcel corresponding to the previous time image in the time sequence and the column direction representing the cultivated land parcel corresponding to the next time image in the time sequence, so as to obtain the matching relationship matrix structure between cultivated land parcels at different times.

[0043] Optionally, the step of calculating the spatial offset vector between cultivated land parcels based on the inter-period parcel matching matrix and constructing the parcel correction matrix specifically involves:

[0044] The numerical elements in the cross-period land parcel matching matrix are traversed in the order of image acquisition time, and the matching relationship between cultivated land parcels at different times is determined based on the magnitude of the numerical elements.

[0045] Based on the matching relationship, obtain the center coordinates of the land parcel boundaries in the land parcel boundary features corresponding to different times;

[0046] The coordinate difference is calculated for the center coordinates of the plot boundary at different times to obtain the horizontal coordinate difference and the vertical coordinate difference; the horizontal coordinate difference and the vertical coordinate difference are combined to form a spatial offset vector;

[0047] All spatial offset vectors are arranged in the order of land parcels. The differences in horizontal and vertical coordinates in the spatial offset vectors are written into the matrix structure as numerical elements to generate the land parcel correction matrix.

[0048] Optionally, the step of performing spatial correction on the boundaries of cultivated land parcels in the multi-period cultivated land image sequence to generate stable parcel units specifically involves:

[0049] Obtain the spatial offset vector in the land parcel correction matrix and sequentially map it to the boundary coordinates of farmland parcels in the multi-phase farmland image sequence according to the image acquisition time order;

[0050] For each farmland plot boundary pixel coordinate, perform coordinate translation calculation, add the boundary pixel horizontal coordinate value to the horizontal coordinate difference in the spatial offset vector, and add the boundary pixel vertical coordinate value to the vertical coordinate difference in the spatial offset vector to obtain the corrected boundary pixel coordinates.

[0051] Arrange all corrected boundary pixel coordinates according to the image row and column order to generate a set of corrected farmland plot boundaries;

[0052] Spatial overlay of the corrected farmland plot boundary sets corresponding to different times, and pixel-by-pixel statistics of the boundary pixel coordinates corresponding to multiple times;

[0053] The coordinates of boundary pixels that have been counted more than or equal to a set number are retained and a stable set of boundary pixels is formed.

[0054] Based on the set of stable boundary pixels, the boundary range of cultivated land plots is constructed, and stable plot units are generated.

[0055] Optionally, S7 specifically includes:

[0056] The boundary pixel coordinates of stable plot units in multi-period cultivated land image sequences are statistically analyzed pixel by pixel. The number of times each boundary pixel appears in images at different times is recorded. The number of times the boundary pixel appears is divided by the total number of images to obtain the pixel stability ratio. The pixel stability ratios corresponding to all boundary pixels in stable plot units are summed and divided by the number of boundary pixels to obtain the inter-period stability index.

[0057] The set of farmland plot boundaries of stable plot units in images at different times is compared pairwise. The absolute values ​​of the differences in horizontal and vertical coordinates between corresponding boundary pixel coordinates are calculated. All difference values ​​are accumulated and divided by the number of boundary pixels involved in the calculation to obtain the average boundary offset value. The average boundary offset value is normalized according to a preset range to obtain the boundary consistency index.

[0058] Adjacency statistics are performed on the coordinates of all cultivated land pixels in the stable plot unit. The number of adjacent cultivated land pixels in the horizontal, vertical and diagonal directions of each cultivated land pixel is counted. The pixel adjacency ratio is obtained by dividing the number of adjacent cultivated land pixels by the total number of adjacent positions. The pixel adjacency ratios of all cultivated land pixels are summed and divided by the number of cultivated land pixels to obtain the spatial continuity index.

[0059] The intertemporal stability index, boundary consistency index, and spatial continuity index are multiplied according to set weights and then summed to generate a comprehensive evaluation index.

[0060] Based on the magnitude of the comprehensive evaluation index, the farmland resource protection status of the stable plot units is classified into different levels, and farmland resource protection evaluation results are generated.

[0061] Optionally, a farmland resource protection assessment system based on big data analysis includes the following modules:

[0062] The data acquisition module is used to acquire multi-period remote sensing image data and cultivated land vector boundary data, perform preprocessing on the multi-period remote sensing image data, and perform spatial cropping on the multi-period remote sensing image data based on the cultivated land vector boundary data to generate a multi-period cultivated land image sequence.

[0063] The segmentation and extraction module is used to input multi-period farmland image sequences into the SegFormer model, extract multi-scale semantic features and generate farmland segmentation maps, and extract boundary features, texture features and spatial structure features of farmland plots based on the farmland segmentation maps to form a multi-period farmland feature set.

[0064] The feature mapping module is used to input the multi-period farmland feature set into the improved PerceiverIO model. A cross-temporal stabilization mechanism is introduced in the cross-attention writing layer of the improved PerceiverIO model to map farmland features from different times to a unified latent space, generate cross-period farmland latent features, calculate the feature distance between farmland plots from different times, and construct the cross-period plot matching cost matrix.

[0065] The matching solution module is used to perform iterative solution on the cross-period land parcel matching cost matrix by combining the optimal transmission algorithm and the Sinkhorn iterative algorithm to obtain the cross-period land parcel matching matrix, calculate the spatial offset vector between cultivated land parcels, and construct the land parcel correction matrix.

[0066] The evaluation generation module is used to perform spatial correction on the boundaries of cultivated land plots in multi-period cultivated land image sequences to generate stable plot units; based on the stable plot units, the inter-period stability index, boundary consistency index and spatial continuity index are calculated to construct a comprehensive evaluation index and generate cultivated land resource protection evaluation results.

[0067] The beneficial effects of this invention are:

[0068] (1) Input the set of multi-period cultivated land features into the improved PerceiverIO model, introduce a cross-temporal stabilization mechanism in the cross-attention writing layer, and realize the mapping of cross-period cultivated land features to a unified latent space through cross-period feature difference calculation and proportional reduction update, effectively suppress the influence of remote sensing image acquisition differences and short-term fluctuations on cultivated land feature expression, and improve the stability and expression accuracy of cross-period cultivated land latent features.

[0069] (2) Based on the potential characteristics of cross-period cultivated land, construct the cross-period plot matching cost matrix, and combine the optimal transmission algorithm and the Sinkhorn iterative algorithm to calculate the cross-period plot matching matrix. Through the cross-period plot matching matrix, determine the correspondence between cultivated land plots at different times, realize high-precision matching of cross-period cultivated land plots, and improve the accuracy of multi-period cultivated land plot change analysis.

[0070] (3) Calculate the spatial offset vector of cultivated land plots based on the inter-period plot matching matrix and construct the plot correction matrix. Perform spatial correction on the boundaries of cultivated land plots in the multi-period cultivated land image sequence to generate stable plot units. Construct a comprehensive evaluation index by calculating the inter-period stability index, boundary consistency index and spatial continuity index to achieve an objective quantitative evaluation of the state of cultivated land resource protection and improve the reliability and scientificity of the cultivated land resource protection evaluation results. Attached Figure Description

[0071] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0072] Figure 1 This is a flowchart of a farmland resource protection evaluation method based on big data analysis proposed in this invention;

[0073] Figure 2 This is a schematic diagram of the improved PerceiverIO model proposed in this invention;

[0074] Figure 3 This is a schematic diagram of a farmland resource protection and evaluation system based on big data analysis proposed in this invention. Detailed Implementation

[0075] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0076] refer to Figures 1-2 A method for evaluating the protection of arable land resources based on big data analysis includes the following steps:

[0077] S1. Acquire multi-period remote sensing image data and cultivated land vector boundary data, perform preprocessing on the multi-period remote sensing image data, and perform spatial cropping on the multi-period remote sensing image data based on the cultivated land vector boundary data to generate a multi-period cultivated land image sequence.

[0078] S2. Input the multi-period farmland image sequence into the SegFormer model, extract multi-scale semantic features and generate farmland segmentation maps. Based on the farmland segmentation maps, extract the boundary features, texture features and spatial structure features of farmland plots to form a multi-period farmland feature set.

[0079] S3. Input the set of farmland features from multiple periods into the improved PerceiverIO model. Introduce a cross-temporal stabilization mechanism in the cross-attention writing layer of the improved PerceiverIO model to map farmland features from different times to a unified latent space and generate cross-period farmland latent features.

[0080] S4. Calculate the feature distance between farmland plots at different times based on the potential features of farmland across periods, and construct the cross-period plot matching cost matrix;

[0081] S5. By combining the optimal transmission algorithm and the Sinkhorn iterative algorithm, the cross-period land parcel matching cost matrix is ​​solved iteratively to obtain the cross-period land parcel matching matrix;

[0082] S6. Calculate the spatial offset vector between cultivated land parcels based on the cross-period parcel matching matrix, construct the parcel correction matrix, perform spatial correction on the boundaries of cultivated land parcels in the multi-period cultivated land image sequence, and generate stable parcel units.

[0083] S7. Calculate the intertemporal stability index, boundary consistency index, and spatial continuity index based on stable plot units, construct a comprehensive evaluation index, and generate farmland resource protection evaluation results.

[0084] In this embodiment, S1 specifically refers to:

[0085] Acquire remote sensing image data and farmland vector boundary data acquired at different times, perform radiometric correction, geometric correction and coordinate unification processing on the remote sensing image data, and generate corrected image data;

[0086] The corrected image data is subjected to resolution unification processing and pixel value standardization processing. Resolution unification processing converts images with different spatial resolutions into a unified resolution through resampling. Pixel value standardization processing generates standard image data by normalizing the image pixel values ​​according to the range of maximum and minimum pixel values.

[0087] The scope of cultivated land area is determined based on cultivated land vector boundary data. The polygon coordinates in the cultivated land vector boundary data are used as the clipping boundary. Spatial clipping processing is performed on the standard image data according to the cultivated land area to generate clipped image data.

[0088] The cropped image data is arranged according to the acquisition time of the remote sensing image data, and the cropped image data from different times are stacked in chronological order to construct a multi-period farmland image sequence.

[0089] In this embodiment, S2 specifically refers to:

[0090] The multi-period farmland image sequence is divided into multiple image blocks according to a set size. The set size is a rectangular area composed of a fixed pixel width and a fixed pixel height. The farmland image sequence is uniformly divided by sequential scanning in the row and column directions to generate multiple image blocks. The row coordinate position and column coordinate position of each image block in the original image are recorded. The row coordinate position and column coordinate position are combined to form a position code value. The position code value is equal to the sum of the row coordinate position value and the column coordinate position value. The position code value is then appended to the corresponding position in the pixel value matrix of the image block to form an image block feature sequence with position identification.

[0091] The image patch feature sequence is input into the SegFormer model, which performs hierarchical feature extraction and feature fusion processing on the image patch feature sequence. Hierarchical feature extraction involves extracting the pixel distribution information of the image patch feature sequence at different spatial scales. Each level of feature is formed by aggregating adjacent pixel regions to form a new pixel representation. The pixel representation value is equal to the sum of the pixel values ​​participating in the aggregation divided by the number of pixels participating in the aggregation. Feature fusion processing involves adding the feature maps generated at different levels pixel by pixel according to their spatial corresponding positions. The addition result forms a multi-scale semantic feature.

[0092] Pixel-level classification calculations are performed based on multi-scale semantic features. The pixel category is determined by comparing the feature values ​​corresponding to each pixel coordinate in the multi-scale semantic features. When the pixel feature value is greater than the set farmland threshold, the pixel is marked as a farmland pixel. When the pixel feature value is less than or equal to the farmland threshold, the pixel is marked as a non-farmland pixel, thus forming a farmland segmentation map.

[0093] Based on the farmland segmentation map, the set of boundary pixels of farmland plots is extracted. By scanning pixel by pixel, it is determined whether there are both farmland pixels and non-farmland pixels in the adjacent area of ​​a pixel. When both types of pixels exist in the adjacent area, the current pixel is recorded as a boundary pixel. All boundary pixels are arranged in the order of the rows and columns of the image to form a boundary pixel set. The center coordinates of the plot boundary are determined by calculating the average value of the pixel coordinates in the boundary pixel set. The range of the plot boundary is determined by calculating the difference between the maximum and minimum values ​​of the horizontal coordinates and the difference between the maximum and minimum values ​​of the vertical coordinates of the boundary pixels, thus forming the plot boundary features.

[0094] The plot texture features are calculated based on the distribution of cultivated land pixel values ​​in the cultivated land segmentation map. The sum of all cultivated land pixel values ​​in the cultivated land segmentation map is calculated and divided by the number of cultivated land pixels to obtain the average cultivated land pixel value. The squared difference between each cultivated land pixel value and the average cultivated land pixel value is calculated and accumulated. The accumulated result is divided by the number of cultivated land pixels to obtain the degree of pixel change. The degree of pixel change is used as the plot texture feature.

[0095] The spatial structure features of the land parcels are calculated based on the spatial adjacency relationships of cultivated land pixels in the cultivated land segmentation map. For each cultivated land pixel in the cultivated land segmentation map, adjacent pixels are determined in the horizontal, vertical, and diagonal directions. The number of cultivated land pixels in the adjacent positions around each cultivated land pixel is counted, and the total number of adjacent positions is recorded. The spatial adjacency degree of the pixels is obtained by dividing the number of adjacent cultivated land pixels by the total number of adjacent positions. The average adjacency degree is obtained by summing the spatial adjacency degrees of all cultivated land pixels and dividing by the number of cultivated land pixels. The average adjacency degree is used as the spatial structure feature of the land parcels.

[0096] The feature data of land parcel boundary features, land parcel texture features, and land parcel spatial structure features are arranged sequentially from early to late according to the image acquisition time. In the time dimension, the feature data of different times are connected sequentially in the row direction to form a new feature matrix, forming a multi-period farmland feature set.

[0097] In this embodiment, the improved PerceiverIO model specifically includes an input encoding layer, a cross-attention writing layer, a latent self-attention layer, and an output mapping layer;

[0098] The input encoding layer performs feature stitching on the plot boundary features, plot texture features, and plot spatial structure features in the multi-period cultivated land feature set. The plot boundary features, plot texture features, and plot spatial structure features are arranged in a fixed order and numerically connected in the row direction to form a feature vector sequence. The feature vector sequence is then subjected to position encoding processing, incorporating the image time position value into the corresponding value in the feature vector sequence to form an encoded feature sequence. The value of the encoded feature sequence is equal to the value of the feature vector sequence plus the image time position value.

[0099] By introducing a cross-temporal stabilization mechanism, the dimension of the encoded feature sequence is transformed and a potential vector sequence is generated according to a set length. The vector values ​​in the encoded feature sequence are multiplied with the values ​​in the transformation matrix and the product values ​​are accumulated to obtain new vector values. The new vector values ​​are arranged in order according to the set length to form a potential vector sequence.

[0100] The cross-attention writing layer performs element-wise multiplication of the feature vectors in the encoded feature sequence with the latent vectors in the latent vector sequence, and accumulates the product values ​​to obtain the dot product value. All dot product values ​​are then exponentially multiplied, and each exponent value is divided by the sum of all exponent values ​​to obtain the weight value. The weight value is then multiplied by the feature vector in the encoded feature sequence and accumulated. The accumulated value is added to the corresponding latent vector value in the latent vector sequence. The difference between the feature vector sequences corresponding to different times is calculated. The difference value is equal to the absolute value of the feature vector value at the later time step minus the absolute value of the feature vector value at the earlier time step. The difference value is compared with a stability threshold. When the difference value is greater than the stability threshold, a reduction ratio is calculated. The reduction ratio is equal to the difference value divided by the sum of the difference value and the stability threshold. The corresponding value in the latent vector sequence is multiplied by 1 and the reduction ratio is subtracted, and the latent vector sequence value is updated to generate a cross-period latent representation.

[0101] The latent self-attention layer performs element-wise multiplication on the latent vectors in the intertemporal latent representation and accumulates the product values ​​to obtain the dot product value. It then performs exponential operation on all dot product values ​​and divides each exponential value by the sum of all exponential values ​​to obtain the weight value. The weight values ​​are then used to perform numerical multiplication on the latent vectors and accumulated to obtain new latent vector values. The new latent vector values ​​are then arranged in the original order to form the latent feature representation.

[0102] The output mapping layer performs matrix multiplication on the latent feature representation, multiplies the vector values ​​in the latent feature representation with the values ​​in the mapping matrix, and accumulates the multiplied values ​​to obtain new vector values. The new vector values ​​are then arranged in the order of the latent vectors to form the latent features of cultivated land across periods.

[0103] In this embodiment, the improved PerceiverIO model introduces intertemporal stability constraint calculation in the cross-attention writing layer. When writing multi-period farmland feature sets between the encoded feature sequence and the latent vector sequence via cross-attention, the numerical differences between feature vectors at different times are calculated simultaneously. The values ​​in the latent vector sequence are then proportionally reduced and updated based on the magnitude of the differences, ensuring that the numerical changes in the intertemporal latent representation remain within a stable range. By performing multiple attention aggregations on the intertemporal latent representation in the latent self-attention layer, farmland features from different times are given a unified expression in the latent space. Combined with feature splicing and location encoding processing in the input encoding layer, parcel boundary features, parcel texture features, and parcel spatial structure features are jointly modeled in the same feature vector sequence. This improves the consistency and stability of intertemporal farmland feature representation, reduces the impact of remote sensing image noise and short-term fluctuations on farmland feature representation, and enhances the accuracy of intertemporal farmland latent feature representation and the reliability of farmland resource protection assessment results.

[0104] In this embodiment, S4 specifically refers to:

[0105] Potential features of cultivated land spanning different periods are grouped and arranged according to the order of image acquisition time to form a set of potential features of plots corresponding to different times;

[0106] For the potential vectors in the potential feature set of the land parcels corresponding to different times, perform element-by-element difference calculation and square the difference values. Then, sum all the squared values ​​to obtain the distance value, which is equal to the sum of the squared differences of all elements.

[0107] Arrange the distance values ​​between the potential features of the land parcels corresponding to all time combinations, and then arrange the distance values ​​in a matrix according to the row direction representing the potential features of the land parcels corresponding to the previous time image in the time sequence and the column direction representing the potential features of the land parcels corresponding to the next time image in the time sequence to generate a cross-period land parcel matching cost matrix.

[0108] The numerical elements in the cross-period land parcel matching cost matrix are indexed and recorded according to the parcel order to obtain the feature distance matrix structure between cultivated land parcels at different times.

[0109] In this embodiment, S5 specifically refers to:

[0110] Obtain the numerical elements in the cross-period land parcel matching cost matrix, perform an exponential operation on each numerical element, and the result of the exponential operation is equal to the numerical element raised to the power of the natural constant to obtain the exponential matrix; sum the numerical elements in each row of the exponential matrix to obtain the row direction normalization coefficient, which is equal to the sum of all numerical elements in the same row; divide each numerical element in each row of the exponential matrix by the corresponding row direction normalization coefficient to obtain the row normalization matrix, and the numerical element in the row normalization matrix is ​​equal to the numerical element at the corresponding position in the exponential matrix divided by the corresponding row direction normalization coefficient;

[0111] Obtain the numerical elements in the row normalized matrix. Summate the numerical elements in each column of the row normalized matrix to obtain the column direction normalization coefficient. The column direction normalization coefficient is equal to the sum of all numerical elements in the same column. Divide each column of the numerical elements in the row normalized matrix by the corresponding column direction normalization coefficient to obtain the column normalized matrix. The numerical elements in the column normalized matrix are equal to the numerical elements at the corresponding positions in the row normalized matrix divided by the corresponding column direction normalization coefficient.

[0112] Obtain the numerical elements in the column normalized matrix. Then, perform row-direction normalization calculation on the column normalized matrix again. Row-direction normalization calculation involves summing the numerical elements in each row to obtain a new row-direction normalization coefficient, and then dividing each row's numerical elements by the new row-direction normalization coefficient to obtain a new row normalized matrix. Perform column-direction normalization calculation on the new row normalized matrix again. Column-direction normalization calculation involves summing the numerical elements in each column to obtain a new column-direction normalization coefficient, and then dividing each column's numerical elements by the new column-direction normalization coefficient to obtain a new column normalized matrix. Repeat this process multiple times, alternating between row-direction and column-direction normalization calculations, until the sum of the numerical elements in each row and the sum of the numerical elements in each column of the matrix both reach the set values.

[0113] The matrix obtained from the final iteration is used as the cross-period land parcel matching matrix. The matrix is ​​arranged according to the row direction representing the cultivated land parcel corresponding to the previous time image in the time sequence and the column direction representing the cultivated land parcel corresponding to the next time image in the time sequence, so as to obtain the matching relationship matrix structure between cultivated land parcels at different times.

[0114] In this embodiment, the spatial offset vector between cultivated land parcels is calculated based on the inter-period parcel matching matrix, and a parcel correction matrix is ​​constructed, specifically as follows:

[0115] The numerical elements in the cross-period land parcel matching matrix are traversed row by row according to the image acquisition time order. The numerical elements in each row of the cross-period land parcel matching matrix are compared according to their numerical values. The numerical element with the largest value is selected as the matching numerical element. The row index position and column index position of the matching numerical element are recorded as the land parcel correspondence, and the matching relationship between cultivated land parcels corresponding to images at different times is determined.

[0116] Based on the matching relationship, the center coordinates of the land parcel boundary in the land parcel boundary features corresponding to different times are obtained. The center coordinates of the land parcel boundary are composed of horizontal coordinate numerical elements and vertical coordinate numerical elements. The center coordinates of the land parcel boundary corresponding to the previous time image in the time sequence and the center coordinates of the land parcel boundary corresponding to the next time image in the time sequence are recorded to form a numerical sequence of the center coordinates of the land parcel boundary.

[0117] The coordinate difference is calculated for the numerical sequence of the center coordinates of the plot boundary. The horizontal coordinate difference element is equal to the horizontal coordinate element corresponding to the later time image in the time sequence minus the horizontal coordinate element corresponding to the earlier time image in the time sequence. The vertical coordinate difference element is equal to the vertical coordinate element corresponding to the later time image in the time sequence minus the vertical coordinate element corresponding to the earlier time image in the time sequence. The horizontal and vertical coordinate difference elements are combined in order to form a spatial offset vector numerical sequence, which is composed of horizontal and vertical coordinate difference elements.

[0118] All spatial offset vector numerical sequences are arranged in order of plot number. The horizontal coordinate difference numerical elements of each spatial offset vector numerical sequence are written into the first column of the matrix, and the vertical coordinate difference numerical elements of each spatial offset vector numerical sequence are written into the second column of the matrix to form a plot correction matrix. Each row of numerical elements in the plot correction matrix represents the spatial offset vector numerical sequence of the corresponding cultivated land plot.

[0119] In this embodiment, spatial correction is performed on the boundaries of cultivated land parcels in a multi-period cultivated land image sequence to generate stable parcel units, specifically as follows:

[0120] The numerical elements in the plot correction matrix are obtained. Each row of the plot correction matrix corresponds to a spatial offset vector numerical sequence. The spatial offset vector numerical sequence is composed of horizontal coordinate difference numerical elements and vertical coordinate difference numerical elements. The spatial offset vector numerical sequence is recorded in accordance with the image acquisition time sequence and the boundary coordinates of cultivated land plots in the multi-period cultivated land image sequence to form a correspondence between the boundary pixel coordinate numerical sequence and the spatial offset vector numerical sequence.

[0121] Obtain the horizontal and vertical coordinate elements of the boundary pixels from the numerical sequence of boundary pixel coordinates of cultivated land plots. Add the horizontal coordinate elements of the boundary pixels to the horizontal coordinate difference elements in the numerical sequence of spatial offset vectors. Add the vertical coordinate elements of the boundary pixels to the vertical coordinate difference elements in the numerical sequence of spatial offset vectors to obtain the corrected horizontal and vertical coordinate elements of the boundary pixels. The corrected horizontal coordinate elements of the boundary pixels are equal to the horizontal coordinate elements of the boundary pixels plus the horizontal coordinate difference elements, and the corrected vertical coordinate elements of the boundary pixels are equal to the vertical coordinate elements of the boundary pixels plus the vertical coordinate difference elements.

[0122] Arrange all the corrected boundary pixel horizontal coordinate values ​​and vertical coordinate values ​​according to the image row direction and image column direction to form a set of corrected farmland plot boundaries.

[0123] Obtain the corrected farmland plot boundary set corresponding to different times, and spatially overlay the corrected farmland plot boundary set corresponding to different times according to the pixel coordinate position. Perform pixel-by-pixel statistics on the boundary pixel coordinates corresponding to the same horizontal coordinate value element and the same vertical coordinate value element. The statistical value is equal to the number of times the same pixel coordinate position appears in the images at different times.

[0124] The statistical value is compared with the set number of times. When the statistical value is greater than or equal to the set number of times, the corresponding boundary pixel coordinates are retained. The retained boundary pixel coordinates are arranged in the order of pixel coordinates to form a stable boundary pixel set.

[0125] Obtain the pixel coordinate values ​​from the stable boundary pixel set. Determine the horizontal range of the boundary based on the maximum and minimum values ​​of the horizontal coordinate values ​​in the stable boundary pixel set. Determine the vertical range of the boundary based on the maximum and minimum values ​​of the vertical coordinate values ​​in the stable boundary pixel set. Combine the horizontal and vertical ranges of the boundary to form the boundary range of the cultivated land plot, thus generating a stable plot unit.

[0126] In this embodiment, S7 specifically refers to:

[0127] Obtain the set of stable boundary pixels corresponding to stable plot units. Perform pixel-by-pixel statistics on the coordinates of boundary pixels in the stable boundary pixel set across multiple cultivated land image sequences. The statistical value is equal to the number of times the boundary pixel corresponding to the same horizontal coordinate value and the same vertical coordinate value appears in images at different times. Divide the number of times the boundary pixel appears by the total number of images to obtain the pixel stability ratio value. The pixel stability ratio value is equal to the number of times the boundary pixel appears divided by the total number of images. Accumulate all the pixel stability ratio values ​​in the stable boundary pixel set and divide by the number of boundary pixels to obtain the inter-period stability index value. The inter-period stability index value is equal to the sum of all pixel stability ratio values ​​divided by the number of boundary pixels.

[0128] The process involves obtaining stable boundary pixel sets of farmland plot boundaries in images from different time periods. These farmland plot boundaries are then compared pairwise based on pixel coordinates. The difference between the horizontal and vertical coordinate values ​​of the corresponding boundary pixels is calculated, and the absolute value of this difference is used to obtain the horizontal and vertical coordinate offset values. All horizontal and vertical coordinate offset values ​​are summed and divided by the number of boundary pixels involved in the calculation to obtain the average boundary offset value. The average boundary offset value is equal to the sum of all horizontal and vertical coordinate offset values ​​divided by the number of boundary pixels. Finally, the average boundary offset value is divided by a preset range to obtain the boundary consistency index value.

[0129] Obtain the set of cultivated land pixel coordinates corresponding to stable plot units. For each cultivated land pixel in the set of cultivated land pixel coordinates, count the adjacent cultivated land pixels in the horizontal, vertical, and diagonal directions. The total number of adjacent positions is equal to the sum of the number of horizontal, vertical, and diagonal adjacent positions. The number of adjacent cultivated land pixels is equal to the number of cultivated land pixels among the adjacent positions. Divide the number of adjacent cultivated land pixels by the total number of adjacent positions to obtain the pixel adjacency ratio. The pixel adjacency ratio is equal to the number of adjacent cultivated land pixels divided by the total number of adjacent positions. Accumulate all pixel adjacency ratios and divide by the number of cultivated land pixels to obtain the spatial continuity index. The spatial continuity index is equal to the sum of all pixel adjacency ratios divided by the number of cultivated land pixels.

[0130] Obtain the numerical elements of the intertemporal stability index, the boundary consistency index, and the spatial continuity index. Multiply the intertemporal stability index by the stability weight element, the boundary consistency index by the consistency weight element, and the spatial continuity index by the continuity weight element. Summate the three product elements to obtain the comprehensive evaluation index. The comprehensive evaluation index is equal to the stable weight element multiplied by the intertemporal stability index, the consistency weight element multiplied by the boundary consistency index, the continuity weight element multiplied by the spatial continuity index.

[0131] The comprehensive evaluation index numerical elements are obtained and compared with the numerical elements of the preset evaluation interval. When the comprehensive evaluation index numerical elements are within different evaluation intervals, the farmland resource protection status corresponding to the stable plot unit is divided into different levels, and the farmland resource protection evaluation result is generated.

[0132] refer to Figure 3 In this embodiment, a farmland resource protection evaluation system based on big data analysis includes the following modules:

[0133] The data acquisition module is used to acquire multi-period remote sensing image data and cultivated land vector boundary data, perform preprocessing on the multi-period remote sensing image data, and perform spatial cropping on the multi-period remote sensing image data based on the cultivated land vector boundary data to generate a multi-period cultivated land image sequence.

[0134] The segmentation and extraction module is used to input multi-period farmland image sequences into the SegFormer model, extract multi-scale semantic features and generate farmland segmentation maps, and extract boundary features, texture features and spatial structure features of farmland plots based on the farmland segmentation maps to form a multi-period farmland feature set.

[0135] The feature mapping module is used to input the multi-period farmland feature set into the improved PerceiverIO model. A cross-temporal stabilization mechanism is introduced in the cross-attention writing layer of the improved PerceiverIO model to map farmland features from different times to a unified latent space, generate cross-period farmland latent features, calculate the feature distance between farmland plots from different times, and construct the cross-period plot matching cost matrix.

[0136] The matching solution module is used to perform iterative solution of the cross-period land parcel matching cost matrix by combining the optimal transmission algorithm and the Sinkhorn iterative algorithm, obtain the cross-period land parcel matching matrix, calculate the spatial offset vector between cultivated land parcels, and construct the land parcel correction matrix;

[0137] The evaluation generation module is used to perform spatial correction on the boundaries of cultivated land plots in multi-period cultivated land image sequences to generate stable plot units; based on the stable plot units, the inter-period stability index, boundary consistency index and spatial continuity index are calculated to construct a comprehensive evaluation index and generate cultivated land resource protection evaluation results.

[0138] Example 1: To verify the feasibility of this invention in practice, it was applied to a farmland resource protection assessment scenario in a major grain-producing area. This area is located in the central part of the North China Plain, where farmland use is mainly based on a wheat-corn rotation system. Farmland is concentrated in scale and has relatively regular plot boundaries. However, due to years of agricultural production activities and infrastructure construction, some farmland plots have experienced frequent boundary changes, slight spatial shifts, and a decline in farmland continuity. Traditional farmland change monitoring methods based on single-period remote sensing images are insufficient to accurately determine the stability relationship of farmland plots at different times, easily misjudging surface changes caused by normal agricultural rotation as farmland destruction, thus affecting the reliability of the farmland resource protection assessment results. To address the aforementioned issues, this embodiment utilizes multi-period remote sensing image data to construct a multi-period cultivated land image sequence. Combined with cultivated land vector boundary data, spatial constraints are applied to cultivated land areas. A cultivated land segmentation map is generated using the SegFormer model. From this map, plot boundary features, plot texture features, and plot spatial structure features are extracted to form a multi-period cultivated land feature set. This feature set is then input into an improved PerceiverIO model. Through cross-attention writing layers and cross-spatiotemporal stabilization mechanisms, cultivated land features from different times are mapped to a unified latent space to generate cross-period cultivated land latent features. In this unified latent space, feature distances between cultivated land plots at different times are calculated, and a cross-period plot matching cost matrix is ​​constructed. This cross-period plot matching matrix is ​​obtained by combining the optimal transmission algorithm and the Sinkhorn iterative algorithm. Furthermore, spatial offset vectors between cultivated land plots are calculated to construct a plot correction matrix. Spatial correction is performed on the cultivated land plot boundaries in the multi-period cultivated land image sequence to generate stable plot units. Finally, based on these stable plot units, a cross-period stability index, a boundary consistency index, and a spatial continuity index are calculated, and a comprehensive evaluation index is constructed, thereby generating a cultivated land resource protection evaluation result.

[0139] This embodiment selects four periods of Sentinel-2 remote sensing image data, each with a spatial resolution of 10m, covering a study area of ​​approximately 320 square kilometers and containing a total of 2184 farmland plots. After radiometric and geometric correction, the images are used to generate a multi-period farmland image sequence. In the farmland segmentation stage, farmland segmentation maps are generated and farmland plot features are extracted. The improved PerceiverIO model is used to generate potential features of farmland across periods. In the feature distance calculation stage, the size of the cross-period plot matching cost matrix is ​​2184×2184. The cross-period plot matching matrix is ​​obtained through Sinkhorn iteration, and the spatial offset vector is calculated. After spatial correction, 2076 stable plot units are formed, while the remaining 108 plots are identified as low-stability plot units due to significant boundary changes.

[0140] Table 1. Statistical Table of Multi-Plot Matching and Spatial Correction Effects of Cultivated Land

[0141] time Number of land parcels included in the calculation Average feature distance Average spatial offset (m) Corrected average offset (m) Matching accuracy Phase 1 2184 0.318 6.42 2.15 0.932 Phase 2 2184 0.305 6.11 2.03 0.941 Phase 3 2184 0.297 5.86 1.94 0.948 4 cases 2184 0.336 6.88 2.34 0.927

[0142] As shown in Table 1, without spatial correction, the average spatial offset of cultivated land parcels between multiple image periods ranges from 5.86m to 6.88m. This offset mainly originates from factors such as geometric errors in remote sensing images, agricultural machinery operations, and land consolidation. When directly using the original images for cultivated land change analysis, parcel boundary misalignment is prone to occur. After constructing a cross-period parcel matching matrix and calculating the spatial offset vector using this invention, spatial correction is performed on the cultivated land parcel boundaries. The average offset after correction decreases to between 1.94m and 2.34m, reducing the spatial offset error by more than 65%. At the same time, the cross-period parcel matching accuracy remains stable above 0.93, indicating that the improved PerceiverIO model can effectively represent the characteristics of cultivated land parcels at different times in a unified latent space. Furthermore, by establishing stable cross-period matching relationships through optimal transmission and Sinkhorn iterative solutions, a reliable data foundation is provided for the subsequent construction of stable parcel units.

[0143] After the stable parcel units were constructed, the intertemporal stability index, boundary consistency index, and spatial continuity index were calculated for the 2076 stable parcel units. A comprehensive evaluation index was constructed based on the set weights, and the evaluation results are shown in Table 2.

[0144] Table 2 Statistical Table of Cultivated Land Resource Protection Evaluation Index

[0145] Plot Number Intertemporal stability index Boundary Consistency Index Spatial continuity index Comprehensive evaluation index Protection status level A102 0.92 0.88 0.91 0.904 High stability A318 0.89 0.84 0.87 0.872 High stability B207 0.83 0.79 0.82 0.813 medium stability B462 0.80 0.76 0.78 0.780 medium stability C115 0.74 0.69 0.73 0.720 generally C309 0.71 0.67 0.70 0.693 generally

[0146] As shown in Table 2, the comprehensive evaluation index can effectively reflect the stability of cultivated land plots across multiple image periods. The inter-period stability index mainly reflects the stability of plot boundaries at different times, the boundary consistency index reflects the magnitude of changes in plot boundary positions, and the spatial continuity index reflects the spatial connectivity of cultivated land pixels. For plots with a comprehensive evaluation index higher than 0.85, both boundary stability and spatial continuity remain at a high level, indicating stable cultivated land use and good protection. Plots with a comprehensive evaluation index between 0.75 and 0.85 show some boundary changes, but the overall spatial structure remains continuous. Plots with a comprehensive evaluation index lower than 0.75 exhibit significant boundary changes or cultivated land fragmentation. The above analysis demonstrates that this invention, by constructing an inter-period plot matching matrix and stable plot units, achieves accurate quantitative evaluation of the stability of cultivated land plots across multiple periods, providing reliable technical support for cultivated land resource protection supervision and cultivated land quality assessment.

[0147] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for evaluating the protection of arable land resources based on big data analysis, characterized in that, Includes the following steps: S1. Acquire multi-period remote sensing image data and cultivated land vector boundary data, perform preprocessing on the multi-period remote sensing image data, and perform spatial cropping on the multi-period remote sensing image data based on the cultivated land vector boundary data to generate a multi-period cultivated land image sequence. S2. Input the multi-period farmland image sequence into the SegFormer model, extract multi-scale semantic features and generate a farmland segmentation map, and extract the boundary features, texture features and spatial structure features of farmland plots based on the farmland segmentation map to form a multi-period farmland feature set. S3. Input the set of farmland features from multiple periods into the improved PerceiverIO model. Introduce a cross-temporal stabilization mechanism in the cross-attention writing layer of the improved PerceiverIO model to map farmland features from different times to a unified latent space and generate cross-period farmland latent features. S4. Calculate the feature distance between farmland plots at different times based on the potential features of farmland across periods, and construct the cross-period plot matching cost matrix; S5. The inter-period land parcel matching cost matrix is ​​obtained by iteratively solving the inter-period land parcel matching matrix by combining the optimal transmission algorithm and the Sinkhorn iterative algorithm. S6. Calculate the spatial offset vector between cultivated land parcels based on the cross-period parcel matching matrix, construct the parcel correction matrix, perform spatial correction on the boundaries of cultivated land parcels in the multi-period cultivated land image sequence, and generate stable parcel units. S7. Calculate the intertemporal stability index, boundary consistency index, and spatial continuity index based on stable plot units, construct a comprehensive evaluation index, and generate farmland resource protection evaluation results.

2. The method for evaluating the protection of arable land resources based on big data analysis according to claim 1, characterized in that, Specifically, S1 is: Acquire remote sensing image data and farmland vector boundary data acquired at different times, perform radiometric correction, geometric correction and coordinate unification processing on the remote sensing image data, and generate corrected image data; The corrected image data is processed to unify resolution and standardize pixel values ​​to generate standard image data. The extent of cultivated land area is determined based on cultivated land vector boundary data. Spatial cropping is then performed on standard image data according to the cultivated land area extent to generate cropped image data. The cropped image data were arranged according to the acquisition time of the remote sensing image data to construct a multi-period farmland image sequence.

3. The method for evaluating and protecting arable land resources based on big data analysis according to claim 1, characterized in that, Specifically, S2 is: The multi-period farmland image sequence is divided into multiple image blocks according to a set size, and the image blocks are subjected to location encoding processing to generate image block feature sequences with location identifiers. The image patch feature sequence is input into the SegFormer model, and hierarchical feature extraction and feature fusion processing are performed on the image patch feature sequence to generate multi-scale semantic features. Pixel-level classification calculations are performed based on multi-scale semantic features to generate farmland segmentation maps; Extract the set of pixels representing the boundaries of cultivated land parcels from the cultivated land segmentation map, calculate the coordinates of the parcel boundaries based on the set of pixels, and generate the parcel boundary features. Calculate the plot texture features based on the distribution of farmland pixel values ​​in the farmland segmentation map; Calculate the spatial structure features of land parcels based on the spatial adjacency relationship of cultivated land pixels in the cultivated land segmentation map; The characteristics of land parcel boundaries, land parcel textures, and land parcel spatial structures are integrated according to the time sequence of the images to form a multi-period set of cultivated land characteristics.

4. The method for evaluating the protection of arable land resources based on big data analysis according to claim 1, characterized in that, The improved PerceiverIO model specifically includes an input encoding layer, a cross-attention writing layer, a latent self-attention layer, and an output mapping layer; The input encoding layer performs feature splicing processing on the plot boundary features, plot texture features and plot spatial structure features in the multi-period cultivated land feature set to generate a feature vector sequence, and performs position encoding processing on the feature vector sequence to generate an encoded feature sequence; The dimensionality of the encoded feature sequence is transformed and a potential vector sequence is generated according to a set length. The cross-attention writing layer introduces a cross-spatiotemporal stabilization mechanism to perform element-wise multiplication between each feature vector in the encoded feature sequence and each latent vector in the latent vector sequence, and sums the product values ​​to obtain a dot product value. The dot product values ​​are then exponentially calculated and divided by the sum of the exponent values ​​to obtain a weight value. The weight values ​​are then used to perform numerical multiplication on the feature vectors in the encoded feature sequence and summed. The calculated value is added to the latent vector value at the corresponding position in the latent vector sequence. The numerical difference between the feature vector sequences corresponding to different times is calculated, and the values ​​in the latent vector sequence are updated according to the numerical difference in a proportional reduction manner to generate a cross-period latent representation. The latent self-attention layer performs element-wise multiplication between latent vectors in the intertemporal latent representation and sums the product values ​​to obtain the dot product value. It then performs exponential calculation on all dot product values ​​and divides them by the sum of the exponential values ​​to obtain the weight value. Finally, it uses the weight value to perform numerical multiplication on the latent vectors and sums the results to generate the latent feature representation. The output mapping layer performs matrix multiplication calculations on the latent feature representations to generate potential features of cultivated land across different periods.

5. The method for evaluating and protecting arable land resources based on big data analysis according to claim 1, characterized in that, Specifically, S4 is: Potential features of cultivated land spanning different periods are grouped and arranged according to the order of image acquisition time to form a set of potential features of plots corresponding to different times; For the potential vectors in the potential feature set of the land parcels corresponding to different times, perform element-by-element difference calculation and square the difference values. Then, sum all the squared values ​​to obtain the distance value, which is equal to the sum of the squared differences of all elements. Arrange the distance values ​​between the potential features of the land parcels corresponding to all time combinations, and then arrange the distance values ​​in a matrix according to the row direction representing the potential features of the land parcels corresponding to the previous time image in the time sequence and the column direction representing the potential features of the land parcels corresponding to the next time image in the time sequence to generate a cross-period land parcel matching cost matrix. The numerical elements in the cross-period land parcel matching cost matrix are indexed and recorded according to the parcel order to obtain the feature distance matrix structure between cultivated land parcels at different times.

6. The method for evaluating the protection of arable land resources based on big data analysis according to claim 1, characterized in that, Specifically, S5 is: The exponential matrix is ​​obtained by performing an exponential operation on all the numerical elements in the cross-period land parcel matching cost matrix. The row direction normalization coefficient is obtained by summing the numerical elements in each row of the exponential matrix. The row normalization matrix is ​​obtained by dividing each row numerical element by the corresponding row direction normalization coefficient. The column-direction normalization coefficient is obtained by summing the numerical elements of each column in the row normalization matrix. The column-direction normalization coefficient is obtained by dividing each numerical element of the column by the corresponding column-direction normalization coefficient. The column-normalized matrix is ​​then normalized again in the row direction and in the column direction. This process is repeated multiple times in an alternating manner of row and column normalization until the sum of the numerical elements in each row and the sum of the numerical elements in each column of the matrix reach the set values. The matrix obtained from the final iteration is used as the cross-period land parcel matching matrix. The matrix is ​​arranged according to the row direction representing the cultivated land parcel corresponding to the previous time image in the time sequence and the column direction representing the cultivated land parcel corresponding to the next time image in the time sequence, so as to obtain the matching relationship matrix structure between cultivated land parcels at different times.

7. The method for evaluating the protection of arable land resources based on big data analysis according to claim 1, characterized in that, The step of calculating the spatial offset vector between cultivated land parcels based on the inter-period parcel matching matrix and constructing the parcel correction matrix is ​​as follows: The numerical elements in the cross-period land parcel matching matrix are traversed in the order of image acquisition time, and the matching relationship between cultivated land parcels at different times is determined based on the magnitude of the numerical elements. Based on the matching relationship, obtain the center coordinates of the land parcel boundaries in the land parcel boundary features corresponding to different times; The coordinate difference is calculated for the center coordinates of the plot boundary at different times to obtain the horizontal coordinate difference and the vertical coordinate difference; the horizontal coordinate difference and the vertical coordinate difference are combined to form a spatial offset vector; All spatial offset vectors are arranged in the order of land parcels. The differences in horizontal and vertical coordinates in the spatial offset vectors are written into the matrix structure as numerical elements to generate the land parcel correction matrix.

8. The method for evaluating the protection of arable land resources based on big data analysis according to claim 1, characterized in that, The process of performing spatial correction on the boundaries of cultivated land parcels in multi-period cultivated land image sequences to generate stable parcel units specifically involves: Obtain the spatial offset vector in the land parcel correction matrix and sequentially map it to the boundary coordinates of farmland parcels in the multi-phase farmland image sequence according to the image acquisition time order; For each farmland plot boundary pixel coordinate, perform coordinate translation calculation, add the boundary pixel horizontal coordinate value to the horizontal coordinate difference in the spatial offset vector, and add the boundary pixel vertical coordinate value to the vertical coordinate difference in the spatial offset vector to obtain the corrected boundary pixel coordinates. Arrange all corrected boundary pixel coordinates according to the image row and column order to generate a set of corrected farmland plot boundaries; Spatial overlay of the corrected farmland plot boundary sets corresponding to different times, and pixel-by-pixel statistics of the boundary pixel coordinates corresponding to multiple times; The coordinates of boundary pixels that have been counted more than or equal to a set number are retained and a stable set of boundary pixels is formed. Based on the set of stable boundary pixels, the boundary range of cultivated land plots is constructed, and stable plot units are generated.

9. The method for evaluating the protection of arable land resources based on big data analysis according to claim 1, characterized in that, Specifically, S7 is: The boundary pixel coordinates of stable plot units in multi-period cultivated land image sequences are statistically analyzed pixel by pixel. The number of times each boundary pixel appears in images at different times is recorded. The number of times the boundary pixel appears is divided by the total number of images to obtain the pixel stability ratio. The pixel stability ratios corresponding to all boundary pixels in stable plot units are summed and divided by the number of boundary pixels to obtain the inter-period stability index. The set of farmland plot boundaries of stable plot units in images at different times is compared pairwise. The absolute values ​​of the differences in horizontal and vertical coordinates between corresponding boundary pixel coordinates are calculated. All difference values ​​are accumulated and divided by the number of boundary pixels involved in the calculation to obtain the average boundary offset value. The average boundary offset value is normalized according to a preset range to obtain the boundary consistency index. Adjacency statistics are performed on the coordinates of all cultivated land pixels in the stable plot unit. The number of adjacent cultivated land pixels in the horizontal, vertical and diagonal directions of each cultivated land pixel is counted. The pixel adjacency ratio is obtained by dividing the number of adjacent cultivated land pixels by the total number of adjacent positions. The pixel adjacency ratios of all cultivated land pixels are summed and divided by the number of cultivated land pixels to obtain the spatial continuity index. The intertemporal stability index, boundary consistency index, and spatial continuity index are multiplied according to set weights and then summed to generate a comprehensive evaluation index. Based on the magnitude of the comprehensive evaluation index, the farmland resource protection status of the stable plot units is classified into different levels, and farmland resource protection evaluation results are generated.

10. A farmland resource protection assessment system based on big data analysis, comprising implementing the farmland resource protection assessment method based on big data analysis as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The data acquisition module is used to acquire multi-period remote sensing image data and cultivated land vector boundary data, perform preprocessing on the multi-period remote sensing image data, and perform spatial cropping on the multi-period remote sensing image data based on the cultivated land vector boundary data to generate a multi-period cultivated land image sequence. The segmentation and extraction module is used to input multi-period farmland image sequences into the SegFormer model, extract multi-scale semantic features and generate farmland segmentation maps, and extract boundary features, texture features and spatial structure features of farmland plots based on the farmland segmentation maps to form a multi-period farmland feature set. The feature mapping module is used to input the multi-period farmland feature set into the improved PerceiverIO model. A cross-temporal stabilization mechanism is introduced in the cross-attention writing layer of the improved PerceiverIO model to map farmland features from different times to a unified latent space, generate cross-period farmland latent features, calculate the feature distance between farmland plots from different times, and construct the cross-period plot matching cost matrix. The matching solution module is used to perform iterative solution on the cross-period land parcel matching cost matrix by combining the optimal transmission algorithm and the Sinkhorn iterative algorithm to obtain the cross-period land parcel matching matrix, calculate the spatial offset vector between cultivated land parcels, and construct the land parcel correction matrix. The evaluation generation module is used to perform spatial correction on the boundaries of cultivated land plots in multi-period cultivated land image sequences to generate stable plot units; based on the stable plot units, the inter-period stability index, boundary consistency index and spatial continuity index are calculated to construct a comprehensive evaluation index and generate cultivated land resource protection evaluation results.

Images