A remote sensing intelligent monitoring method for zero-sample change detection of non-grain conversion of cultivated land in mountainous areas

Abstract

This invention provides a remote sensing intelligent monitoring method for non-grain crop conversion in mountainous areas using zero-sample change detection. The method collects high-resolution optical images, farmland vector data, and economic crop data to construct a typical example library containing seven types of farmland non-grain crop conversion changes. It calculates the change feature vectors for each change type, uses the SAM model for semantic segmentation of pre- and post-processing images, and designs a semantic information change detection operator based on spatially regularized diffusion learning clustering and dual-temporal latent spatial matching. This achieves zero-sample-based extraction of non-grain crop conversion patches in mountainous areas. This invention requires only a small number of typical examples to construct the example library, avoiding large-scale labeled training samples. It is low-cost and highly efficient. The spatially regularized diffusion learning clustering post-processing algorithm improves the segmentation accuracy of fragmented farmland in mountainous areas, and the semantic information change detection operator based on dual-temporal latent spatial matching enhances the detection accuracy of non-grain crop conversion. This provides a new solution for refined farmland identification and food security monitoring.

Description

Technical Field

[0001] This invention relates to the field of deep learning change detection technology, and in particular to a remote sensing intelligent monitoring method for the non-grain conversion of cultivated land in mountainous areas using zero-sample change detection. Background Technology

[0002] Monitoring the conversion of arable land to non-grain crops mainly refers to the prohibition of planting fruit trees, forage grasses, raising fish, piling up soil, or altering agricultural facilities and other facilities in grain-growing areas. It requires extracting such non-grain crop patches. Traditional monitoring of arable land conversion to non-grain crops relies primarily on manual on-site verification and reporting of arable land area and type at the village level. This method has two main drawbacks: first, manual on-site verification cannot be comprehensive and can only be conducted through sampling, which is time-consuming and labor-intensive. Furthermore, the accuracy of the verification is directly related to the operator's skill level and sense of responsibility; second, village-level reporting is prone to false reporting, as farmers may take chances and conceal information about planting cash crops or raising livestock.

[0003] Traditional monitoring of farmland conversion to non-grain crops, relying on manual on-site verification, is clearly insufficient to meet the current demands for large-scale, high-frequency, and refined farmland monitoring in farmland protection efforts. In recent years, improvements in spatial and spectral resolution of satellite remote sensing, along with the rapid development of intelligent remote sensing big data analysis technology, have created new opportunities for monitoring farmland conversion to non-grain crops. Currently, there are two main approaches to using intelligent remote sensing technology for monitoring farmland conversion to non-grain crops: one is to create intelligent remote sensing interpretation samples of farmland crops, train and construct a refined intelligent remote sensing model for identifying farmland crops, perform intelligent interpretation of multi-temporal images for each period, and then compare and analyze to extract non-grain crop patches; the other is to create farmland conversion to non-grain crop change detection samples, train and construct a change detection model, and extract non-grain crop patches from two periods of imagery. Both of these deep learning-driven approaches require the labeling of large-scale, high-quality training samples. Due to the fragmented and scattered nature of mountainous farmland, the complex land cover types, the mixing of farmland and garden grasslands, and the prevalence of crop rotation, intercropping, and relay cropping within farmland, it is extremely difficult to label large-scale, high-quality farmland crop planting or non-grain change detection sample data at the pixel-level in practical work. This requires enormous manpower and a long time. In recent years, to address the practical difficulties of producing large-scale deep learning training samples, zero-shot learning has been proposed by scholars to identify, segment, and detect changes in target objects without training samples. Summary of the Invention

[0004] Therefore, it is necessary to provide a remote sensing intelligent monitoring method for non-grain conversion of cultivated land in mountainous areas with zero-sample change detection to address the aforementioned technical problems.

[0005] A remote sensing intelligent monitoring method for zero-sample change detection of non-grain conversion of cultivated land in mountainous areas includes the following steps:

[0006] Step S1: Acquire high-resolution optical images of the area to be tested from different years, farmland survey monitoring vectors for the corresponding years, and economic crop survey data for the following year; perform geometric fine correction and pixel-level position registration on the high-resolution optical images to obtain image data from the earlier and later stages.

[0007] Step S2: Obtain the change characteristics of each change type in the typical sample library for detecting changes in arable land that are no longer used for grain production; wherein, the change types include: arable land becoming cash crops, arable land becoming forest land, arable land becoming nursery land, arable land becoming grassland, arable land becoming pits and ponds, arable land becoming buildings and structures, and arable land becoming excavated land.

[0008] Step S3: Perform sliding cropping on the pre- and post-image data to obtain a group of image patches; wherein, the group of image patches includes: a pre-image patch and a post-image patch; segment the group of image patches according to the SAM model and stitch the segmented patches together to obtain the image segmentation result;

[0009] Step S4: The segmentation results of the image are clustered using a spatial regularized diffusion learning clustering algorithm to obtain optimized segmentation and clustering results;

[0010] Step S5: Extract change patches from the optimized segmentation and clustering results using the semantic information change detection operator of dual-temporal latent space matching, obtain change patches, and calculate change features;

[0011] Step S6: Calculate the Pearson correlation coefficient based on the extracted changed patches and features and the change features of each change type in the typical sample library to obtain the patch change type.

[0012] In one embodiment, step S2 is preceded by:

[0013] By using overlay analysis, the map patches that have changed from cultivated land to non-cultivated land are filtered out from the cultivated land survey monitoring vector and economic crop survey data. Typical map patches that have changed from cultivated land to non-cultivated land are obtained to create a typical sample library for detecting changes in cultivated land conversion to non-grain production.

[0014] Obtain a typical sample library for detecting changes in cultivated land conversion to non-grain crops, and calculate the change characteristics of each change type in the typical sample library.

[0015] In one embodiment, a typical sample library for detecting changes in arable land conversion to non-grain crops is obtained, and the change characteristics of each change type in the typical sample library are calculated, including:

[0016] Calculate the variation characteristics of typical sample patches under the variation type. The variation characteristics of the typical sample patches include: pixel-level variation characteristics and object-level variation characteristics.

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027] in, Indicates the first A typical example of pixel-level variation characteristics of image patches The first The difference between the mean values ​​of the histograms of the red, green, and blue bands in the preceding and following image data within a typical sample patch. The first The difference in standard deviations of the red, green, and blue band histograms of pre- and post-image data within a typical sample patch. Indicates the first Typical example image features at the object level. The first The differences in roughness, contrast, directionality, linearity, regularity, and coarseness texture features in the red, green, and blue bands of the image data from different periods within a typical sample patch. The first The differences between the morphological top cap transformation features and the morphological bottom cap transformation features of the red, green, and blue bands in the image data of the previous and later stages of a typical sample patch. These represent the differences in roughness for the red, green, and blue bands, respectively. These represent the differences in contrast between the red, green, and blue bands, respectively. These represent the differences in directionality among the red, green, and blue bands, respectively. These represent the differences in linearity among the red, green, and blue bands, respectively. These represent the regularity differences between the red, green, and blue bands, respectively. These represent the differences in coarseness for the red, green, and blue bands, respectively. These represent the differences in morphological top-hat transformation characteristics for the red, green, and blue bands, respectively. These represent the differences in morphological cap transformation characteristics for the red, green, and blue bands, respectively.

[0028] The variation characteristics of the typical sample patch yield the variation characteristics of the variation type:

[0029] ;

[0030] ;

[0031] ;

[0032] in, This represents the average value of pixel-level variation features. This represents the number of all typical sample patches under the CT-type variation. Indicates the first A typical example image patch, Indicates object-level change characteristics. This represents the change characteristics of the CT-th type of change.

[0033] In one embodiment, step S4 includes:

[0034] The kernel density is calculated based on the segmentation results of the image, and representative pixels are selected according to the preset kernel density requirements. A spatially regularized KNN nearest neighbor graph is constructed based on the representative pixels to obtain a reduced spatially regularized diffusion graph.

[0035] Based on the reduced spatial regularization diffusion map, K pixels are located as cluster pattern centers. Cluster labels are propagated to unlabeled pixels in the reduced spatial regularization diffusion map. The cluster labels of the remaining pixels are determined based on the cluster labels of the pixels in the reduced spatial regularization diffusion map, thus obtaining the optimized segmentation and clustering results.

[0036] In one embodiment, the kernel density is calculated based on the segmentation result of the image, and representative pixels are selected according to a preset kernel density requirement. A spatially regularized KNN nearest neighbor graph is constructed based on the representative pixels to obtain a reduced spatially regularized diffusion graph, including:

[0037] Calculate the kernel density using the following formula:

[0038] ;

[0039] in, Represents the pixels within each patch object nuclear density, express The Middle 1 pixel, Represents pixels The set of nearest neighbor pixels within a radius of R pixels in the segmentation result, determined by Euclidean distance. , Representing pixels Pixel values ​​in the red band of the image, , Representing pixels Pixel values ​​in the green band of the image, , Representing pixels Pixel values ​​in the blue band of the image, This represents a scaling factor that controls the radius of interaction between pixels. Represents the regularization factor. make sure ;

[0040] Within each patch object, the kernel density of each pixel is sorted in descending order, and the top k pixels are selected as the representative pixels of the patch object.

[0041] The spatially regularized KNN nearest neighbor graph is constructed using the following formula, resulting in a reduced spatially regularized diffusion graph:

[0042] ;

[0043] in, This indicates a reduced spatial regularization diffusion map. Indicates the total number of objects in the map. Indicates the first A single image object, The number of pixels in .

[0044] In one embodiment, K pixels are located as cluster pattern centers based on the reduced spatial regularization diffusion map, cluster labels are propagated to unlabeled pixels in the reduced spatial regularization diffusion map, and cluster labels of the remaining pixels are determined based on the cluster labels of the pixels in the reduced spatial regularization diffusion map, resulting in optimized segmentation and clustering results including:

[0045] The cluster pattern center representation value is calculated according to the following formula, and the K pixels with the largest cluster pattern center representation values ​​are selected as the cluster pattern centers:

[0046] ;

[0047] ;

[0048] in, Represents the cluster pattern center representation value. Indicates time Time Pixel Compared with pixels in the reduced spatial regularization diffusion map With higher density The diffusion distance between nearest neighbors The time calculation is based on the spatially regularized diffusion map and Markov diffusion process. Time Pixel and Medium pixel diffusion distance, This represents a reduced spatial regularization diffusion plot;

[0049] Cluster labels are propagated to unlabeled pixels in the reduced spatial regularization diffusion map according to the following formula:

[0050] ;

[0051] ;

[0052] in, The cluster label function for pixel X. This indicates reducing the number of unlabeled pixels in the spatially regularized diffusion map. Indicates distance The most recent cluster center pixel, This represents the parameter that minimizes the function value. This indicates a reduced spatial regularization diffusion map. Indicates the cluster pattern center, The values ​​can be 1, 2, 3...K. Represents pixels and In time The diffusion distance;

[0053] The correlation score is calculated using the following formula, and the cluster label of the pixel with the highest correlation score is selected:

[0054] ;

[0055] in, Indicates the degree of correlation. Indicates unlabeled clustered pixels With each labeled cluster tag pixel in the reduced spatial regularization diffusion map Mahalanobis distance, Represents pixels Its own kernel density estimate;

[0056] Adjacent patches with the same cluster label are merged to obtain optimized segmentation and clustering results.

[0057] In one embodiment, step S5 includes:

[0058] The latent features of the previous and later images are extracted by the image encoder of the SAM model. The latent features are averaged in each channel within each patch of the optimized segmentation and clustering results to obtain the latent feature vector.

[0059] Calculate the confidence score of the patch change based on the latent feature vector to obtain the degree of semantic change of the patch in the previous and later images;

[0060] Obtain a confidence score threshold for the change of a patch; if the confidence score for the change of a patch is greater than the confidence score threshold for the change of a patch, then the patch is used as the extracted changed patch.

[0061] In one embodiment, a confidence score for patch change is calculated based on the latent feature vector to obtain the degree of semantic change of patches in the preceding and following images, including:

[0062] Calculate the confidence score for the change in map features using the following formula:

[0063] ;

[0064] ;

[0065] in, This represents the confidence score for changes in map features. This represents the confidence score of the change in any patch in a previous image patch. This represents the confidence score of the change in any patch in a later-stage image patch. This represents the confidence score of the change in any given patch. Indicates the map patch number, These represent the potential feature vectors of corresponding patches at the same location in the earlier and later images, respectively.

[0066] In one embodiment, step S6 includes:

[0067] The Pearson correlation coefficient is calculated using the following formula:

[0068] ;

[0069] in, This represents the Pearson correlation coefficient. Describing covariance, This represents the feature vector of the index-th changed patch. The variance representing the variation characteristics of the CT-th variation type;

[0070] Obtain the correlation coefficient threshold, and in response to the Pearson correlation coefficient being greater than the correlation coefficient threshold, define the type of the changed patch as the change type corresponding to the change feature.

[0071] A remote sensing intelligent monitoring system for the non-grain conversion of cultivated land in mountainous areas with zero-sample change detection, used to implement the remote sensing intelligent monitoring method for the non-grain conversion of cultivated land in mountainous areas with zero-sample change detection as described above, comprising:

[0072] The data acquisition module is used to acquire high-resolution optical images of the area to be measured in different years, farmland survey monitoring vectors of the corresponding years, and economic crop survey data of the following year; and to perform geometric fine correction and pixel-level position registration on the high-resolution optical images to obtain image data of the early and later stages.

[0073] The change sample acquisition module is used to acquire the change characteristics of each change type in the typical sample library for detecting changes in farmland to non-grain crops. The change types include: farmland to cash crops, farmland to forest, farmland to nursery, farmland to grassland, farmland to pits and ponds, farmland to buildings and structures, and farmland to excavated land.

[0074] The image segmentation module is used to perform sliding cropping on the pre- and post-image data to obtain image patch groups; wherein, the image patch group includes: a pre-image patch and a post-image patch; the image patch group is segmented according to the SAM model and the segmented patches are stitched together to obtain the image segmentation result;

[0075] The clustering optimization module is used to perform post-processing on the segmentation results of the image using a spatial regularized diffusion learning clustering algorithm to obtain optimized segmentation and clustering results.

[0076] The change extraction module is used to extract change patches from the optimized segmentation and clustering results by using a semantic information change detection operator based on bi-temporal latent space matching, thereby obtaining change patches and calculating change features;

[0077] The type acquisition module is used to calculate the Pearson correlation coefficient based on the change patch and change characteristics and the change characteristics of each change type in the typical sample library, so as to obtain the change type of the patch.

[0078] Compared with existing technologies, the advantages and beneficial effects of this invention are as follows: This invention cleverly combines the open-source SAM model with an innovative clustering algorithm, breaking through the bottleneck of traditional fully supervised learning relying on large-scale labeled samples. Only a small number of typical examples are needed to build a change detection example library, significantly reducing manpower and material costs. By optimizing the segmentation results through spatial regularized diffusion learning clustering algorithm, it effectively addresses the challenges of fragmented farmland and complex land cover in mountainous areas, improving the accuracy of patch segmentation. A semantic information change detection operator with dual-temporal potential spatial matching is designed, using cosine distance and bidirectional matching strategies to quantify the confidence of patch changes, enhancing the temporal symmetry and robustness of change detection. The overall process deeply integrates pixel-level and object-level feature analysis, realizing intelligent processing from change patch extraction to type recognition. This provides a low-cost, high-efficiency, and high-precision technical solution for monitoring the non-grain conversion of farmland in mountainous areas, and has important application value for the refined protection of farmland and food security. Attached Figure Description

[0079] Figure 1 This is a flowchart illustrating a remote sensing intelligent monitoring method for detecting the non-grain conversion of cultivated land in mountainous areas using zero-sample change detection, as shown in one embodiment.

[0080] Figure 2A This is a schematic diagram of farmland being converted into buildings and excavated land in one embodiment;

[0081] Figure 2B This is a schematic diagram illustrating the transformation of farmland into ponds in one embodiment;

[0082] Figure 3 This is a schematic diagram of the structure of a remote sensing intelligent monitoring system for detecting the non-grain conversion of cultivated land in mountainous areas with zero sample change, as described in one embodiment. Detailed Implementation

[0083] Before describing the specific embodiments of the present invention, the overall concept of the present invention will be explained as follows:

[0084] This invention is mainly developed for the traditional monitoring process of farmland non-grain conversion. The traditional farmland non-grain conversion monitoring, which relies on manual on-site verification, can obviously no longer meet the current farmland protection work's demand for large-scale, high-frequency, and refined farmland monitoring.

[0085] To address the shortcomings of existing technologies, this invention provides a remote sensing intelligent monitoring method for non-grain conversion of cultivated land in mountainous areas using zero-sample change detection. This method solves the problems of high manpower and material costs and low efficiency in labeling large-scale cultivated land non-grain conversion change detection samples due to complex mountainous terrain, fragmented cultivated land, and crop rotation, intercropping, and other factors within cultivated land, as well as the insufficient performance and robustness of traditional intelligent models. The method utilizes a zero-sample change detection strategy to identify changed patches, then extracts color, texture, and morphological features from previous and subsequent images of the changed patches, performs similarity determination with a cultivated land non-grain conversion change detection sample library, and finally filters out the changed cultivated land non-grain conversion patches.

[0086] After introducing the overall concept of the present invention, in order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below through specific embodiments and in conjunction with the accompanying drawings.

[0087] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of this specification should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of this specification do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word covers the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0088] In one embodiment, such as Figure 1 As shown, a remote sensing intelligent monitoring method for detecting the non-grain conversion of cultivated land in mountainous areas with zero-sample change is provided, including the following steps:

[0089] Step S1: Obtain high-resolution optical images of the area to be tested from different years, farmland survey monitoring vectors for the corresponding years, and economic crop survey data for the following year; perform geometric fine correction and pixel-level position registration on the high-resolution optical images to obtain image data from the earlier and later stages.

[0090] Specifically, high-resolution optical images of the area to be tested from two years, farmland survey monitoring vectors from the corresponding two years, and economic crop survey data from the following year are collected. The two years are considered as the early and late stages. Geometric fine correction and pixel-level position registration are performed on the images to obtain the early and late image data. Overlay analysis is used to filter out the patches where farmland has been converted to non-farmland from the farmland survey monitoring vectors from the two years and the economic crop survey data from the following year.

[0091] In this embodiment, 0.5-meter resolution images of Beibei District, Chongqing were collected. The time period of the early images was July 2020, and the time period of the later images was August 2024. The cultivated land survey monitoring vector in 2020 came from the results of the Third National Land Survey in 2020, and the cultivated land survey monitoring vector in 2024 came from the results of the 2024 National Land Change Survey. The economic crop survey data in 2024 came from the results of the special survey on cultivated forest space.

[0092] Step S2: Obtain the change characteristics of each change type in the typical sample library for detecting changes in arable land that are no longer used for grain production; wherein, the change types include: arable land becoming cash crops, arable land becoming forest land, arable land becoming nursery land, arable land becoming grassland, arable land becoming pits and ponds, arable land becoming buildings and structures, and arable land becoming excavated land.

[0093] Specifically, the change characteristics of each change type in the constructed typical sample library for detecting changes in arable land that are no longer used for grain production are obtained. There are a total of 7 types of changes in arable land that are no longer used for grain production. Among them, the types of arable land that are no longer used for grain production include arable land becoming cash crops, arable land becoming forest land, arable land becoming nursery land, arable land becoming grassland, arable land becoming pits and ponds, arable land becoming buildings and structures, and arable land becoming excavated land. 5 to 10 typical samples are selected for each type of change.

[0094] Based on this, step S2 also includes the following:

[0095] By using overlay analysis, the map patches that have changed from cultivated land to non-cultivated land are filtered out from the cultivated land survey monitoring vector and economic crop survey data. Typical map patches that have changed from cultivated land to non-cultivated land are obtained to create a typical sample library for detecting changes in cultivated land conversion to non-grain production.

[0096] Obtain a typical sample library for detecting changes in cultivated land conversion to non-grain crops, and calculate the change characteristics of each change type in the typical sample library.

[0097] Specifically, the typical sample library for detecting changes in arable land conversion to non-arable land includes typical sample plots of arable land converted to non-arable land, along with corresponding image data from earlier and later periods.

[0098] In this embodiment, based on high-resolution optical images from both the early and late stages, typical sample patches of farmland converted to non-farmland are visually selected, and marker information is generated and received. Combined with subsequent economic crop survey data, 10 typical sample patches of farmland converted to economic crops are found from the early and late images, and a typical sample library for detecting changes in farmland conversion to non-grain crops is created.

[0099] Furthermore, in this embodiment, the images used to create the sample library for detecting changes in arable land conversion to non-grain production are located in Longfengqiao Subdistrict, Xiema Town, Shijialiang Town, and Caijiagang Town of Beibei District.

[0100] Based on this, a typical sample library for detecting changes in arable land conversion to non-grain crops was obtained, and the change characteristics of each change type in the typical sample library were calculated, including:

[0101] Calculate the variation characteristics of typical sample patches under the variation type. The variation characteristics of the typical sample patches include: pixel-level variation characteristics and object-level variation characteristics.

[0102] ;

[0103] ;

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] in, Indicates the first A typical example of pixel-level variation characteristics of image patches The first The difference between the mean values ​​of the histograms of the red, green, and blue bands in the preceding and following image data within a typical sample patch. The first The difference in standard deviations of the red, green, and blue band histograms of pre- and post-image data within a typical sample patch. Indicates the first Typical example image patch object-level variation characteristics The first The differences in roughness, contrast, directionality, linearity, regularity, and coarseness texture features in the red, green, and blue bands of the image data from different periods within a typical sample patch. The first The differences between the morphological top cap transformation features and the morphological bottom cap transformation features of the red, green, and blue bands in the image data of the previous and later stages of a typical sample patch. These represent the differences in roughness for the red, green, and blue bands, respectively. These represent the differences in contrast between the red, green, and blue bands, respectively. These represent the differences in directionality among the red, green, and blue bands, respectively. These represent the differences in linearity among the red, green, and blue bands, respectively. These represent the regularity differences between the red, green, and blue bands, respectively. These represent the differences in coarseness for the red, green, and blue bands, respectively. These represent the differences in morphological top-hat transformation characteristics for the red, green, and blue bands, respectively. These represent the differences in morphological cap transformation characteristics for the red, green, and blue bands, respectively.

[0113] The variation characteristics of the typical sample patch yield the variation characteristics of the variation type:

[0114] ;

[0115] ;

[0116] ;

[0117] in, This represents the average value of pixel-level variation features. This represents the number of all typical sample patches under the CT-type variation. Indicates the first A typical example image patch, Indicates object-level change characteristics. This represents the change characteristics of the CT-th type of change.

[0118] Specifically, the calculation of change characteristics for each type of change in the typical sample library for detecting changes in arable land conversion to non-grain crops includes the following process:

[0119] First, calculate the th type of CT change. The variation characteristics of a typical sample image feature mainly include pixel-level variation characteristics and object-level variation characteristics. Pixel-level variation characteristics The main data points are the differences between the mean and standard deviation of the histograms of the three bands in the early and later stages of typical sample images, representing object-level variation characteristics. The main features include Tamura texture and morphological features in three bands of the early and late images within typical sample patches. Tamura texture features include roughness, contrast, directionality, linearity, regularity, and coarsness. Morphological features include morphological top-hat transformation and morphological bottom-hat transformation.

[0120] ;

[0121] ;

[0122] in, These represent the differences in the mean values ​​of the red, green, and blue band histograms of the early and later images within a typical sample image patch. These represent the difference in standard deviations of the histograms for the red, green, and blue bands of the early and later images within a typical sample patch. The values ​​represent the differences in six Tamura texture features (roughness, contrast, directionality, linearity, regularity, and coarsness) across the red, green, and blue bands of the early and later images within a typical sample patch. The differences between the morphological top-hat transformation features and the morphological bottom-hat transformation features of the red, green, and blue bands in the early and late stages of the typical sample image patch are calculated as follows:

[0123] ;

[0124] ;

[0125] ;

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] ;

[0131] in, This represents the roughness of the red, green, and blue bands of the previous image within a typical sample patch. This represents the roughness of the red, green, and blue bands in the later image within a typical sample patch. The roughness includes, but is not limited to, the standard deviation of the gray-level difference within the pixel neighborhood. This indicates the contrast of the red, green, and blue bands in the previous image within a typical sample patch. This represents the contrast of the red, green, and blue bands in the later image within a typical sample patch. The contrast includes, but is not limited to, calculations using the second moment of the grayscale histogram. This indicates the directionality of the red, green, and blue bands in the previous image within a typical sample patch. This indicates the directionality of the red, green, and blue bands in the later image within a typical sample patch. The directionality includes, but is not limited to, detection of peak values ​​in the gradient orientation histogram. This indicates the linearity of the red, green, and blue bands in the previous image within a typical sample patch. This represents the linearity of the red, green, and blue bands in the later image within a typical sample patch. The linearity includes, but is not limited to, calculations based on the variance of the edge direction distribution. This indicates the regularity of the red, green, and blue bands in the previous image within a typical sample patch. This indicates the regularity of the red, green, and blue bands in the later image within a typical sample patch. Regularity includes, but is not limited to, obtaining it through multi-scale autocorrelation analysis. This indicates the coarseness of the red, green, and blue bands in the previous image within a typical sample patch. This indicates the coarsness of the red, green, and blue bands in the later image within a typical sample patch. The coarsness includes, but is not limited to, calculations using wavelet transform energy coefficients. This indicates the morphological top-hat transformation characteristics of the red, green, and blue bands in the previous image within a typical sample patch. This represents the morphological top-hat transformation characteristics of the red, green, and blue bands in the later image within a typical sample patch. This indicates the morphological base-cap transformation characteristics of the red, green, and blue bands in the previous image within a typical sample patch. This shows the morphological base-cap transformation characteristics of the red, green, and blue bands in the later stages of a typical sample image patch.

[0132] Next, the variation characteristics of all typical sample patches under this variation type are calculated, and the average value is taken to obtain the variation characteristics of this variation type. These are then combined into a one-dimensional vector. ;

[0133] ;

[0134] ;

[0135] ;

[0136] Wherein, Nct represents the number of all typical sample patches under the CT-type variation.

[0137] Finally, following the steps outlined above, each of the seven variation types was calculated to obtain the variation feature vectors for all seven variation types. , , , , , , .

[0138] Step S3: Perform sliding cropping on the pre- and post-image data to obtain a group of image patches; wherein, the group of image patches includes: a pre-image patch and a post-image patch; segment the group of image patches according to the SAM model and stitch the segmented patches together to obtain the image segmentation result.

[0139] Specifically, the images from the pre-production and post-production stages are cropped into smaller pieces using a sliding window (1024*1024 pixels in this embodiment). The sliding window moves from left to right and from top to bottom to crop the images from the pre-production and post-production stages. The moving step of the sliding window is 1024 pixels, meaning that the image pieces do not overlap. This results in Sm groups of image pieces, each group consisting of one pre-production image piece and one post-production image piece.

[0140] For each image patch, the open-source SAM (Segment Anything Model) model is used to segment the earlier and later image patches separately, obtaining the segmentation results for the earlier and later image patches. Then, the segmentation results of all the earlier image patches are stitched together to obtain the earlier image segmentation result. The post-processing image segmentation result is obtained by stitching together the segmentation results of all the post-processing image segments. The image segmentation results include the segmentation results of the previous image and the segmentation results of the later image.

[0141] Furthermore, in this embodiment, the images used to cut and create small pieces are located in Tianfu Town, Jingguan Town, Shuitu Town, and Fuxing Town in Beibei District, which are spatially different from the images used to create the sample library for detecting changes in farmland conversion to non-grain production.

[0142] Step S4: The segmentation results of the image are clustered using a spatial regularized diffusion learning clustering algorithm to obtain optimized segmentation and clustering results.

[0143] Specifically, the segmentation results of the two images are clustered using a spatially regularized diffusion learning clustering algorithm to obtain the optimized segmentation and clustering results of the earlier image. And subsequent optimization of image segmentation and clustering results .

[0144] Based on this, the segmentation results of the image are clustered using a spatial regularized diffusion learning clustering algorithm to obtain optimized segmentation and clustering results, including:

[0145] The kernel density is calculated based on the segmentation results of the image, and representative pixels are selected according to the preset kernel density requirements. A spatially regularized KNN nearest neighbor graph is constructed based on the representative pixels to obtain a reduced spatially regularized diffusion graph.

[0146] Based on the reduced spatial regularization diffusion map, K pixels are located as cluster pattern centers. Cluster labels are propagated to unlabeled pixels in the reduced spatial regularization diffusion map. The cluster labels of the remaining pixels are determined based on the cluster labels of the pixels in the reduced spatial regularization diffusion map, thus obtaining the optimized segmentation and clustering results.

[0147] The kernel density is calculated based on the segmentation results of the image, and representative pixels are selected according to the preset kernel density requirements. A spatially regularized KNN nearest neighbor graph is constructed based on the representative pixels to obtain a reduced spatially regularized diffusion graph, including:

[0148] Calculate the kernel density using the following formula:

[0149] ;

[0150] in, Represents the pixels within each patch object nuclear density, express A certain pixel in, Represents pixels The set of nearest neighbor pixels within a radius of R pixels in the segmentation result, determined by Euclidean distance. , Representing pixels Pixel values ​​in the red band of the image, , Representing pixels Pixel values ​​in the green band of the image, , Representing pixels Pixel values ​​in the blue band of the image, This represents a scaling factor that controls the radius of interaction between pixels. Represents the regularization factor. make sure ;

[0151] Within each patch object, the kernel density of each pixel is sorted in descending order, and the top k pixels are selected as the representative pixels of the patch object.

[0152] The spatially regularized KNN nearest neighbor graph is constructed using the following formula, resulting in a reduced spatially regularized diffusion graph:

[0153] ;

[0154] in, This indicates a reduced spatial regularization diffusion map. Indicates the total number of objects in the map. Indicates the first A single image object, The number of pixels in .

[0155] Based on the reduced spatial regularization diffusion map, K pixels are located as cluster pattern centers. Cluster labels are propagated to unlabeled pixels in the reduced spatial regularization diffusion map. The cluster labels of the remaining pixels are determined based on the cluster labels of the pixels in the reduced spatial regularization diffusion map, resulting in optimized segmentation and clustering results, including:

[0156] The cluster pattern center representation value is calculated according to the following formula, and the K pixels with the largest cluster pattern center representation values ​​are selected as the cluster pattern centers:

[0157] ;

[0158] ;

[0159] in, Represents the cluster pattern center representation value. Indicates time Time Pixel Compared with pixels in the reduced spatial regularization diffusion map With higher density The diffusion distance between nearest neighbors The time calculation is based on the spatially regularized diffusion map and Markov diffusion process. Time Pixel and medium pixel diffusion distance, This represents a reduced spatial regularization diffusion plot;

[0160] Cluster labels are propagated to unlabeled pixels in the reduced spatial regularization diffusion map according to the following formula:

[0161] ;

[0162] ;

[0163] in, The cluster label function for pixel x. This indicates reducing the number of unlabeled pixels in the spatially regularized diffusion map. Indicates distance The most recent cluster center pixel, This represents the parameter that minimizes the function value. This indicates a reduced spatial regularization diffusion map. Indicates the cluster pattern center, The values ​​can be 1, 2, 3...K. Represents pixels and In time The diffusion distance;

[0164] The correlation score is calculated using the following formula, and the cluster label of the pixel with the highest correlation score is selected:

[0165] ;

[0166] in, Indicates the degree of correlation. Indicates unlabeled clustered pixels With each labeled cluster tag pixel in the reduced spatial regularization diffusion map Mahalanobis distance, Represents pixels Its own kernel density estimate;

[0167] Adjacent patches with the same cluster label are merged to obtain optimized segmentation and clustering results.

[0168] Specifically, the post-processing of the segmentation results of a certain period of image using the spatial regularized diffusion learning clustering algorithm mainly includes the following steps:

[0169] Step 4.1.1, constructing a spatially regularized diffusion graph, including calculating kernel density estimation, selecting representative pixels, and constructing a spatially regularized KNN nearest neighbor connection graph, including the following steps.

[0170] Step 4.1.10: Treat each patch in the segmentation result image of the open-source SAM model as a patch object, with a total number of patch objects. , For the first Each patch object is used to calculate pixels within each patch object. Density(x) of the kernel:

[0171] ;

[0172] in This refers to a specific pixel in the image of the patch object. , For all pixels of the image, Indicates that pixel x is in The set of nearest neighbor pixels within a radius of R pixels in the Euclidean distance (nearest neighbor radius is R pixels), where y is... A certain pixel in, Z is a scaling factor that controls the interaction radius between pixels, and Z is a regularization factor that ensures... =1, , For each pixel Pixel values ​​in the red band of the image, , For each pixel Pixel values ​​in the green band of the image, , For each pixel Pixel values ​​in the blue band of the image.

[0173] Furthermore, in this embodiment, the total number of objects in the early image patches... =1364, total number of objects in later image patches =1528, in this embodiment, the nearest neighbor radius R is set to 10, which is a scaling factor to control the interaction radius between pixels. Set it to 6.

[0174] Step 4.1.11: Within each patch object, according to the kernel density requirement (select a larger kernel density), sort the kernel density Density(x) of each pixel in descending order, and select the top k pixels as the representative pixels of the patch object.

[0175] Step 4.1.12, use the selected One representative pixel is used as the most representative high-density pixel. Each pixel is denoted as ,use To construct a diffusion graph with reduced spatial regularization, a KNN nearest neighbor graph is built from the pixels in the graph.

[0176] ;

[0177] Furthermore, in this embodiment, 6820 representative pixels were selected from the early-stage image and 7640 representative pixels were selected from the later-stage image.

[0178] Step 4.1.2 involves performing diffusion learning clustering using the constructed reduced spatial regularization diffusion map, including locating K pixels as cluster pattern centers and propagating cluster labels to... Unlabeled pixels in, based on The cluster labels of the middle pixels are used to determine the cluster labels of the remaining pixels, including the following steps.

[0179] Step 4.1.20, from K pixels are located in the cluster pattern and used as representative pixel samples of the potential cluster structure. A unique cluster label (label value 1, 2, 3...K) is assigned to each of these K pixels. Let one of these K pixels be denoted as... The corresponding tags are lb ,Right now Clustering pattern center is to make Maximize the K pixels, the cluster pattern centers of these K pixels are The highest density pixel in the middle that is furthest from other high-density pixels in terms of diffusion distance;

[0180] ;

[0181] in For pixel x at time t, Medium pixels (v) have a higher density. The diffusion distance between nearest neighbors The pixel x at time t is calculated based on the diffusion map with reduced spatial regularization and the Markov diffusion process. The diffusion distance of the middle pixel v.

[0182] Step 4.1.21, through calculation Each unlabeled pixel With the center pixel of each clustering pattern diffusion distance , Minimum value corresponding to The clustering label is Cluster tags enable the propagation of cluster tags to Unlabeled pixels in, i.e. ,in .

[0183] Step 4.1.22, based on The cluster labels of the middle pixels determine the cluster labels of the remaining pixels in the image, by calculating... Unlabeled clustered pixel and Each labeled cluster tag pixel Mahalanobis distance And then Sort in ascending order F( ) The maximum value corresponding to The clustering label is Cluster labels are used to determine the cluster labels of unlabeled pixels. The representativeness (kernel density) of an unlabeled pixel is weighted by the similarity (represented by the reciprocal of the Mahalanobis distance) between the unlabeled and labeled pixels. The larger the F value, the higher the probability that the unlabeled and labeled pixels belong to the same cluster.

[0184] Step 4.1.23, based on All pixels in the image are clustered with labels. Adjacent pixels with the same cluster label are merged to obtain the clustered post-processed pixels.

[0185] Step 4.1.3, following the steps above, for and After clustering, we obtain and .

[0186] Step S5: Extract change patches from the optimized segmentation and clustering results using the semantic information change detection operator of dual-temporal latent space matching, obtain change patches, and calculate change features.

[0187] Specifically, the semantic information change detection operator based on dual-temporal latent space matching is used to optimize the previous image segmentation and clustering results. and post-image segmentation and clustering optimization results Extract the changed patches.

[0188] Based on this, step S5 includes:

[0189] The latent features of the previous and later images are extracted by the image encoder of the SAM model. The latent features are averaged in each channel within each patch of the optimized segmentation and clustering results to obtain the latent feature vector.

[0190] Calculate the confidence score of the patch change based on the latent feature vector to obtain the degree of semantic change of the patch in the previous and later images;

[0191] Obtain a confidence score threshold for the change of a patch; if the confidence score for the change of a patch is greater than the confidence score threshold for the change of a patch, then the patch is used as the extracted changed patch.

[0192] Based on the latent feature vector, a confidence score for patch change is calculated to obtain the degree of semantic change in patches between pre- and post-image periods, including:

[0193] Calculate the confidence score for the change in map features using the following formula:

[0194] ;

[0195] ;

[0196] in, This represents the confidence score for changes in map features. This represents the confidence score of the change in any patch in a previous image patch. This represents the confidence score of the change in any patch in a later-stage image patch. This represents the confidence score for the change in any given patch. Indicates the map patch number, These represent the potential feature vectors of corresponding patches at the same location in the earlier and later images, respectively.

[0197] Specifically, step 5.1 involves using the encoder of the open-source SAM model to extract latent features from the earlier and later images. , ,exist Calculate within each patch object The average value is used as the early potential representation vector of the patch. ,exist Calculate within each patch object The average value is used as the latent representation vector of the patch. .

[0198] Step 5.2: Calculate using cosine distance and The similarity between the images is calculated, and the ChangeScore is used to assess the degree of semantic change of the images before and after the transition.

[0199] ;

[0200] Where index is the map patch number. This represents the potential representation vector of the corresponding patch at the same location in the previous and later images. The value ranges from 0 to 1, with higher values ​​indicating greater semantic change in the patch. The formula above is the complement of cosine similarity, used to measure the directional difference in the latent feature vectors of patches from different periods. When the vector directions are completely identical, the cosine value = 1. =0 (no change); when the vector directions are orthogonal, the cosine value = 0. =1 (maximum change).

[0201] Step 5.3: Design a bidirectional matching change detection strategy to calculate the change confidence score, ensuring the temporal symmetry of change detection. First, use... Using the given patch as a baseline, calculate the confidence score of the change in that patch. , and then with Using the given patch as a baseline, calculate the confidence score of the change in that patch. The final confidence score of the patch change = Set the confidence score threshold for patch changes. Exceeding the threshold This indicates that the patch is a variable patch.

[0202] Step S6: Calculate the Pearson correlation coefficient based on the changed patch and the change characteristics to obtain the patch change type.

[0203] Based on this, step S6 includes:

[0204] The Pearson correlation coefficient is calculated using the following formula:

[0205] ;

[0206] in, This represents the Pearson correlation coefficient. Describing covariance, This represents the feature vector of the index-th changed patch. The variance representing the variation characteristics of the CT-th variation type;

[0207] Obtain the correlation coefficient threshold, and in response to the Pearson correlation coefficient being greater than the correlation coefficient threshold, define the type of the changed patch as the change type corresponding to the change feature.

[0208] Specifically, pixel-level and object-level change features are calculated based on the index-th change patch to obtain the feature vector of the index-th change patch. Then, the Pearson correlation coefficient is calculated with the feature vector of each type of change. If the value exceeds the threshold PT, it indicates that the change type of the patch index is the CT-th change type.

[0209] ;

[0210] in, For covariance, , They are respectively The variance.

[0211] The changed patch image is converted into a vector, and an attribute field "change type" is added. Based on the Pearson correlation coefficient and threshold, the change type of each changed patch is determined. The change type is assigned to the attribute information of the corresponding changed patch. The attribute information values ​​are: cultivated land becomes cash crop, cultivated land becomes forest, cultivated land becomes nursery, cultivated land becomes grassland, cultivated land becomes pit, cultivated land becomes building, and cultivated land becomes excavated land. The resulting vector of remote sensing intelligent monitoring results of cultivated land non-grain conversion is obtained.

[0212] The remote sensing intelligent monitoring method for detecting zero-sample change in mountainous cultivated land non-grain production provided by this invention has significant effects:

[0213] (1) Based on the results of semantic segmentation of high-resolution remote sensing images before and after the open-source semantic segmentation model SAM, a zero-sample method for detecting changes in farmland non-grain conversion was designed, which includes two key steps: spatial regularization diffusion learning clustering algorithm and semantic information change detection operator with dual temporal potential spatial matching. This method uses the zero-sample change detection strategy to find changed patches, and then extracts color, texture, and morphological features from the images before and after the changed patches. The similarity is measured with the farmland non-grain conversion change detection sample library, and then the changed patches of farmland non-grain conversion are filtered out. This provides a new solution for the refined identification of farmland, prevention of farmland "non-grain conversion", and comprehensive consolidation of food security.

[0214] (2) The method of the present invention only requires a small number of manually labeled typical examples of farmland non-grain change areas, deeply explores the feature differences between the previous and later high-resolution images, avoids large-scale labeling of non-grain change area interpretation samples and training models, saves manpower and material resources, and has the advantages of lower cost, higher efficiency and better performance compared with fully supervised deep learning semantic segmentation algorithms.

[0215] It should be noted that the method of this embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this embodiment, and the multiple devices will interact with each other to complete the method described.

[0216] It should be noted that the above description describes some embodiments of the present invention. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0217] Based on the same inventive concept, and corresponding to any of the above embodiments, the present invention also provides a remote sensing intelligent monitoring system for zero-sample change detection of non-grain conversion of cultivated land in mountainous areas.

[0218] refer to Figure 3 The aforementioned remote sensing intelligent monitoring system for zero-sample change detection of non-grain conversion of cultivated land in mountainous areas includes:

[0219] The data acquisition module 201 is used to acquire high-resolution optical images of the area to be measured in different years, farmland survey monitoring vectors of the corresponding years, and economic crop survey data of the following year; and to perform geometric fine correction and pixel-level position registration on the high-resolution optical images to obtain image data of the early and later stages.

[0220] The change sample acquisition module 202 is used to acquire the change characteristics of each change type in the typical sample library for detecting changes in farmland to non-grain crops; among which, the change types include: farmland to cash crops, farmland to forest land, farmland to nursery, farmland to grassland, farmland to pits and ponds, farmland to buildings and structures, and farmland to excavated land.

[0221] The image segmentation module 203 is used to perform sliding cropping on the pre- and post-image data to obtain a group of image patches; wherein, the group of image patches includes: a pre-image patch and a post-image patch; the group of image patches is segmented according to the SAM model and the segmented patches are stitched together to obtain the image segmentation result;

[0222] The clustering optimization module 204 is used to perform post-clustering processing on the segmentation results of the image using a spatial regularized diffusion learning clustering algorithm to obtain optimized segmentation clustering results;

[0223] The change extraction module 205 is used to extract change patches from the optimized segmentation and clustering results by using a semantic information change detection operator of dual-temporal latent space matching, to obtain change patches and calculate change features;

[0224] The type acquisition module 206 is used to calculate the Pearson correlation coefficient based on the change features, change characteristics and change characteristics of each change type in the typical sample library, so as to obtain the change type of the patch.

[0225] For ease of description, the above system is described by dividing it into various modules based on their functions. Of course, in implementing this invention, the functions of each module can be implemented in one or more software and / or hardware components.

[0226] The system described in the above embodiments is used to implement a remote sensing intelligent monitoring method for non-grain conversion of cultivated land in mountainous areas with zero-sample change detection in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0227] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of the invention as described above, which are not provided in the details for the sake of brevity.

[0228] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0229] While specific details have been set forth to describe exemplary embodiments of the invention, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive. Although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description.

[0230] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this invention should be included within the protection scope of this invention.

Claims

1. A remote sensing intelligent monitoring method for non-grain conversion of cultivated land in mountainous areas using zero-sample change detection, characterized in that, include: Step S1: Obtain high-resolution optical images of the area to be tested from different years, farmland survey monitoring vectors for the corresponding years, and economic crop survey data for the following year. Geometric fine correction and pixel-level position registration are performed on the high-resolution optical image to obtain the image data of the early and later stages; Step S2: Obtain the change characteristics of each change type in the typical sample library for detecting changes in arable land that are no longer used for grain production; wherein, the change types include: arable land becoming cash crops, arable land becoming forest land, arable land becoming nursery land, arable land becoming grassland, arable land becoming pits and ponds, arable land becoming buildings and structures, and arable land becoming excavated land. Step S3: Perform sliding cropping on the pre- and post-image data to obtain a group of image patches; wherein, the group of image patches includes: a pre-image patch and a post-image patch; segment the group of image patches according to the SAM model and stitch the segmented patches together to obtain the image segmentation result; Step S4: The segmentation results of the image are clustered using a spatial regularized diffusion learning clustering algorithm to obtain optimized segmentation and clustering results; Step S5: Extract change patches from the optimized segmentation and clustering results using the semantic information change detection operator of dual-temporal latent space matching, obtain the change patches, and calculate the corresponding feature vectors; Step S6: Calculate the Pearson correlation coefficient based on the feature vector of the extracted changed patch and the change features of each change type in the typical sample library to obtain the patch change type; Step S5 includes: The latent features of the previous and later images are extracted by the image encoder of the SAM model. The latent features are averaged in each channel of each patch in the optimized segmentation and clustering results to obtain the latent feature vector. Calculate the confidence score of the patch change based on the latent feature vector to obtain the degree of semantic change of the patch in the previous and later images; Obtain a confidence score threshold for the change of a patch; if the confidence score for the change of a patch is greater than the confidence score threshold for the change of a patch, then the patch is used as the extracted changed patch.

2. The remote sensing intelligent monitoring method for non-grain conversion of cultivated land in mountainous areas with zero-sample change detection according to claim 1, characterized in that, The procedure preceding step S2 also includes: By using overlay analysis, the map patches that have changed from cultivated land to non-cultivated land are filtered out from the cultivated land survey monitoring vector and economic crop survey data. Typical map patches that have changed from cultivated land to non-cultivated land are obtained to create a typical sample library for detecting changes in cultivated land conversion to non-grain production. Obtain a typical sample library for detecting changes in cultivated land conversion to non-grain crops, and calculate the change characteristics of each change type in the typical sample library.

3. The remote sensing intelligent monitoring method for non-grain conversion of cultivated land in mountainous areas with zero-sample change detection according to claim 2, characterized in that, The process of obtaining a typical sample library for detecting changes in arable land conversion to non-grain crops, and calculating the change characteristics of each change type in the typical sample library, includes: Calculate the variation characteristics of typical sample patches under the variation type. The variation characteristics of the typical sample patches include: pixel-level variation characteristics and object-level variation characteristics. ， ， ， ， ， ， ， ， ， ， in, Indicates the first A typical example of pixel-level variation characteristics of image patches The first The difference between the mean values ​​of the histograms of the red, green, and blue bands in the preceding and following image data within a typical sample patch. The first The difference in standard deviations of the red, green, and blue band histograms of pre- and post-image data within a typical sample patch. Indicates the first Typical example image patch object-level variation characteristics The first The differences in roughness, contrast, directionality, linearity, regularity, and coarseness texture features in the red, green, and blue bands of the image data from different periods within a typical sample patch. The first The differences between the morphological top cap transformation features and the morphological bottom cap transformation features of the red, green, and blue bands in the image data of the previous and later stages of a typical sample patch. These represent the differences in roughness for the red, green, and blue bands, respectively. These represent the differences in contrast between the red, green, and blue bands, respectively. These represent the differences in directionality among the red, green, and blue bands, respectively. These represent the differences in linearity among the red, green, and blue bands, respectively. These represent the regularity differences between the red, green, and blue bands, respectively. These represent the differences in coarseness for the red, green, and blue bands, respectively. These represent the differences in morphological top-hat transformation characteristics for the red, green, and blue bands, respectively. These represent the differences in morphological cap transformation characteristics for the red, green, and blue bands, respectively. The variation characteristics of the typical sample patch yield the variation characteristics of the variation type: ， ， ， in, This represents the average value of pixel-level variation features. This represents the number of all typical sample patches under the CT-type variation. Indicates the first A typical example image patch, Indicates object-level change characteristics. This represents the change characteristics of the CT-th type of change.

4. The remote sensing intelligent monitoring method for non-grain conversion of cultivated land in mountainous areas with zero-sample change detection according to claim 1, characterized in that, Step S4 includes: The kernel density is calculated based on the segmentation results of the image, and representative pixels are selected according to the preset kernel density requirements. A spatially regularized KNN nearest neighbor graph is constructed based on the representative pixels to obtain a reduced spatially regularized diffusion graph. Based on the reduced spatial regularization diffusion map, K pixels are located as cluster pattern centers. Cluster labels are propagated to unlabeled pixels in the reduced spatial regularization diffusion map. The cluster labels of the remaining pixels are determined based on the cluster labels of the pixels in the reduced spatial regularization diffusion map, thus obtaining the optimized segmentation and clustering results.

5. The remote sensing intelligent monitoring method for non-grain conversion of cultivated land in mountainous areas with zero-sample change detection according to claim 4, characterized in that, The step of calculating the kernel density based on the segmentation results of the image, selecting representative pixels according to preset kernel density requirements, and constructing a spatially regularized KNN nearest neighbor graph based on the representative pixels to obtain a reduced spatially regularized diffusion graph includes: Calculate the kernel density using the following formula: ， in, Represents the pixels within each patch object nuclear density, express The Middle 1 pixel, Represents pixels The set of nearest neighbor pixels within a radius of R pixels in the segmentation result, determined by Euclidean distance. , Representing pixels Pixel values ​​in the red band of the image, , Representing pixels Pixel values ​​in the green band of the image, , Representing pixels Pixel values ​​in the blue band of the image, This represents a scaling factor that controls the radius of interaction between pixels. Represents the regularization factor. make sure ; Within each patch object, the kernel density of each pixel is sorted in descending order, and the top k pixels are selected as the representative pixels of the patch object. The spatially regularized KNN nearest neighbor graph is constructed using the following formula, resulting in a reduced spatially regularized diffusion graph: ， in, This indicates a reduced spatial regularization diffusion map. Indicates the total number of objects in the map. Indicates the first A single image object, The number of pixels in .

6. The remote sensing intelligent monitoring method for non-grain conversion of cultivated land in mountainous areas with zero-sample change detection according to claim 5, characterized in that, The steps of locating K pixels as cluster pattern centers based on the reduced spatial regularization diffusion map, propagating cluster labels to unlabeled pixels in the reduced spatial regularization diffusion map, and determining the cluster labels of the remaining pixels based on the cluster labels of the pixels in the reduced spatial regularization diffusion map to obtain the optimized segmentation and clustering results include: The cluster pattern center representation value is calculated according to the following formula, and the K pixels with the largest cluster pattern center representation values ​​are selected as the cluster pattern centers: ， ， in, Represents the cluster pattern center representation value. Indicates time Time Pixel With higher density HSI pixels The diffusion distance between nearest neighbors The time calculation is based on the spatially regularized diffusion map and Markov diffusion process. Time Pixel and Medium pixel diffusion distance, This represents a reduced spatial regularization diffusion plot; Cluster labels are propagated to unlabeled pixels in the reduced spatial regularization diffusion map according to the following formula: ， ， in, The cluster label function for pixel X. This indicates reducing the number of unlabeled pixels in the spatially regularized diffusion map. Indicates distance The most recent cluster center pixel, This represents the parameter that minimizes the function value. This indicates a reduced spatial regularization diffusion map. Indicates the cluster pattern center, The values ​​can be 1, 2, 3...K. Represents pixels and In time The diffusion distance; The correlation score is calculated using the following formula, and the cluster label of the pixel with the highest correlation score is selected: ， in, Indicates the degree of correlation. Indicates unlabeled clustered pixels With each labeled cluster tag pixel in the reduced spatial regularization diffusion map Mahalanobis distance, Represents pixels Its own kernel density estimate; Adjacent patches with the same cluster label are merged to obtain optimized segmentation and clustering results.

7. The remote sensing intelligent monitoring method for non-grain conversion of cultivated land in mountainous areas with zero-sample change detection according to claim 1, characterized in that, The step of calculating the confidence score of patch change based on the latent feature vector to obtain the degree of semantic change of patches in pre- and post-image images includes: Calculate the confidence score for the change in map features using the following formula: ， ， in, This represents the confidence score for changes in map features. This represents the confidence score of the change in any patch in a previous image patch. This represents the confidence score of the change in any patch in a later-stage image patch. This represents the confidence score for the change in any given patch. Indicates the map patch number, These represent the potential feature vectors of corresponding patches at the same location in the earlier and later images, respectively.

8. The remote sensing intelligent monitoring method for non-grain conversion of cultivated land in mountainous areas with zero-sample change detection according to claim 1, characterized in that, Step S6 includes: The Pearson correlation coefficient is calculated using the following formula: ， in, This represents the Pearson correlation coefficient. Describing covariance, This represents the feature vector of the index-th changed patch. express variance The variance representing the variation characteristics of the CT-th variation type; Obtain the correlation coefficient threshold, and in response to the Pearson correlation coefficient being greater than the correlation coefficient threshold, define the type of the changed patch as the change type corresponding to the change feature.

9. A remote sensing intelligent monitoring system for the non-grain conversion of cultivated land in mountainous areas with zero-sample change detection, characterized in that, A remote sensing intelligent monitoring method for detecting non-grain conversion of cultivated land in mountainous areas, as described in any one of claims 1-8, includes: The data acquisition module is used to acquire high-resolution optical images of the area to be measured in different years, farmland survey monitoring vectors of the corresponding years, and economic crop survey data of the following year; and to perform geometric fine correction and pixel-level position registration on the high-resolution optical images to obtain image data of the early and later stages. The change sample acquisition module is used to acquire the change characteristics of each change type in the typical sample library for detecting changes in farmland to non-grain crops. The change types include: farmland to cash crops, farmland to forest, farmland to nursery, farmland to grassland, farmland to pits and ponds, farmland to buildings and structures, and farmland to excavated land. The image segmentation module is used to perform sliding cropping on the pre- and post-image data to obtain image patch groups; wherein, the image patch group includes: a pre-image patch and a post-image patch; the image patch group is segmented according to the SAM model and the segmented patches are stitched together to obtain the image segmentation result; The clustering optimization module is used to perform post-processing on the segmentation results of the image using a spatial regularized diffusion learning clustering algorithm to obtain optimized segmentation and clustering results. The change extraction module is used to extract change patches from the optimized segmentation and clustering results by using a semantic information change detection operator based on bi-temporal latent space matching, thereby obtaining change patches and calculating the corresponding feature vectors. The type acquisition module is used to calculate the Pearson correlation coefficient based on the feature vector of the changed patch and the change characteristics of each change type in the typical sample library, so as to obtain the change type of the patch.

Images