A platform for remote sensing monitoring of cultivated land protection under multi-source information

Abstract

The application discloses a kind of cultivated land protection remote sensing monitoring platform under multi-source information, belong to cultivated land resource protection technical field, comprising: multi-source data preprocessing, multidimensional feature extraction, cultivated land outflow monitoring module, pre-processing module integrates multi-source remote sensing image data and forms standardized monitoring dataset, and independent space map spot with homogeneous spectral and texture features is divided out;Multidimensional feature extraction module extracts features from four dimensions of feature change intensity, vegetation growth cycle, tillage texture entropy value and ridge linearity, generates multidimensional core feature dataset after correcting local disturbance data;Cultivated land outflow monitoring module constructs coupling weighted discrimination model to complete spot preliminary judgment, filters false change interference, and then realizes fine classification by support vector machine, outputs standardized monitoring results and connects cultivated land supervision port;The application can comprehensively characterize cultivated land state, reduce monitoring misjudgment, and give consideration to global monitoring efficiency and classification effect, to provide reliable technical support for cultivated land normalization control.

Description

Technical Field

[0001] This invention belongs to the field of farmland resource protection technology, specifically relating to a remote sensing monitoring platform for farmland protection based on multi-source information. Background Technology

[0002] Remote sensing monitoring of arable land is a key support for the protection and supervision of arable land. Currently, most methods rely on single-dimensional features for analysis and judgment, and have been gradually applied in the supervision of arable land across the entire region. However, existing methods have shortcomings in feature characterization, change identification, and monitoring applications, making it difficult to meet the needs of routine protection and management of arable land. Therefore, this invention urgently needs to solve the following technical problems: The existing monitoring features are limited in dimension and cannot fully reflect the dynamic changes of cultivated land, vegetation growth and cultivation patterns. The data is easily affected by local anomalies and cannot truly reflect the actual condition of cultivated land, which has an adverse impact on the identification of cultivated land status. The identification of farmland outflow lacks a reasonable configuration of feature weights and cannot effectively filter out interference from non-real outflow such as cloud shadows and isolated local changes. The monitoring process is prone to misjudgment and cannot support the preliminary screening of farmland outflow over a large area. Remote sensing monitoring of arable land struggles to balance the efficiency of large-scale monitoring with the effectiveness of refined classification. It is difficult to distinguish between illegal outflow of arable land and regular crop rotation. The monitoring results are not well integrated with daily supervision, and it cannot provide support for the normalized management and control of arable land. Therefore, we propose a remote sensing monitoring platform for arable land protection based on multi-source information. Summary of the Invention

[0003] The purpose of this invention is to provide a remote sensing monitoring platform for farmland protection based on multi-source information, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a remote sensing monitoring platform for farmland protection based on multi-source information, comprising: Multi-source data preprocessing module: acquires multi-source fusion and registration remote sensing images, digital differential images, land cover interpretation rules and time-series standardized image data of the entire Xishuangbanna Prefecture, integrates them to form a standardized basic dataset for farmland monitoring, and performs independent spatial patch division processing on the multi-source fusion and registration remote sensing images; Multidimensional feature extraction module: Constructs spatial mapping between independent spatial patches and global digital differential imagery, calculates the intensity value of land cover change in independent spatial patches; analyzes vegetation characteristics of independent spatial patches, constructs a subset of vegetation growth periodicity features; extracts the tillage texture entropy value and field ridge linearity of independent spatial patches and completes local perturbation correction; integrates various features to generate a multidimensional core feature dataset; Farmland outflow monitoring module: Construct a multi-feature coupled weighted discrimination model, analyze the feature weights of each independent spatial patch, calculate the farmland outflow discrimination score of each independent spatial patch and complete the initial judgment of the patch, organize and construct a set of farmland outflow patches to be confirmed; train a support vector machine model to classify and discriminate the set of farmland outflow patches to be confirmed, determine the final farmland outflow patches, and monitor the farmland outflow status of the entire region.

[0005] Preferably, the specific process of integrating and forming a standardized basic dataset for farmland monitoring, and performing independent spatial patch delineation on multi-source fused and registered remote sensing images is as follows: The project aims to acquire multi-source fusion registered remote sensing images of the entire Xishuangbanna Prefecture after spatial registration and multi-source fusion processing, and simultaneously acquire digital differential images of the entire prefecture, land cover interpretation rules, and time-series standardized image data, and integrate them to form a standardized basic dataset for farmland monitoring. Existing multi-scale segmentation methods are used to process multi-source fusion and registration remote sensing images, resulting in independent spatial patches with homogeneous spectral and textural features. All independent spatial patches are then numbered according to their spatial coordinates.

[0006] Preferably, the specific process for constructing a spatial mapping between independent spatial patches and the global digital differential image, and calculating the land cover change intensity value of the independent spatial patches, is as follows: For each independent spatial patch, the corresponding spatial boundary coordinates are obtained. Based on a unified spatial reference system, the spatial boundary coordinates of the patch are projected and matched with the pixel coordinates of the global digital differential image to establish a one-to-one spatial mapping relationship. Traverse the entire digital differential image to extract all pixels within the patch area to form a basic pixel set of land cover changes; extract the radiance value of each pixel in the two adjacent time-series standardized images, and calculate the absolute value of the difference to obtain the time-series change value of a single pixel; Calculate the initial mean and standard deviation of the temporal change values ​​within the patch, iteratively remove outliers and recalculate until the mean converges, and take the mean of the temporal change values ​​of the final effective pixels to obtain the ground feature change intensity value.

[0007] Preferably, the specific process of analyzing the vegetation characteristics of independent spatial patches and constructing a subset of vegetation growth periodicity characteristics is as follows: Based on a unified spatial reference system, the coordinates of independent spatial patches and time-normalized images are matched, and the image regions corresponding to the patches in each time phase are extracted. The normalized vegetation index of the pixels in the region is calculated and converted into vegetation cover to obtain the multi-time phase vegetation cover sequence of the patches. The mean and standard deviation of vegetation cover in multiple time phases are calculated to obtain the time series variation coefficient. The peak and trough values ​​of the sequence are extracted and the duration of the peak values ​​is statistically analyzed. The above parameters are integrated to construct a subset of vegetation growth periodicity features.

[0008] Preferably, the specific process for extracting the tillage texture entropy value and field ridge linearity of independent spatial patches is as follows: Based on the land cover interpretation rules, the optical band image region corresponding to the independent spatial patch is extracted from the multi-source fusion and registration remote sensing image. The optical band image region is then processed into grayscale to obtain the single-channel grayscale image corresponding to the independent spatial patch. Based on the analysis of single-channel grayscale images using the gray-level co-occurrence matrix, the cultivation texture entropy value of independent spatial patches is calculated. An edge detection algorithm is used to extract the linear contours of a single-channel grayscale image, and candidate contours of field ridges are obtained by screening. The candidate contours of field ridges are fitted with straight lines by the least squares method and the linearity is calculated. The arithmetic mean of the linearity of all candidate contours of field ridges is taken to obtain the linearity of field ridges of independent spatial patches.

[0009] Preferably, the specific process for completing local perturbation correction and generating a multidimensional core feature dataset is as follows: Centered on each independent spatial patch, the cultivation texture entropy value and field ridge linearity of the surrounding adjacent independent spatial patches are extracted; Calculate the correlation coefficients of the tillage texture entropy values ​​and the correlation coefficients of the field ridge linearity between the independent spatial patch and its adjacent independent spatial patches, respectively. The number of correlation coefficients of tillage texture entropy and field ridge linearity that are less than the preset correlation threshold is counted separately. When the preset judgment condition is met, it is determined that the independent spatial patch has local disturbance. For map patches with local disturbances, the arithmetic mean of the corresponding features of the surrounding adjacent map patches is used for replacement; the intensity value of land cover change, the temporal variation coefficient of vegetation coverage, the subset of vegetation growth periodic features, the entropy value of tillage texture and the linearity of field ridges are integrated and normalized to generate a multidimensional core feature dataset.

[0010] Preferably, the specific process of constructing a multi-feature coupled weighted discrimination model and analyzing the feature weights of each independent spatial patch is as follows: A training sample set was constructed by selecting farmland patch samples that had been verified in the field and confirmed by land type attributes. The samples included four types of farmland patches and corresponding multi-dimensional feature datasets. The multi-dimensional feature datasets and corresponding land type attributes were input into a random forest model to construct a binary classification model for farmland outflow discrimination. The out-of-bag error method is used to calculate the overall importance score of each feature. The overall importance score is then normalized to obtain the overall adaptive weight coefficients for each feature.

[0011] Preferably, the specific process of calculating the farmland outflow discrimination score for each independent spatial patch and completing the initial patch judgment, and compiling and constructing the set of farmland outflow patches to be confirmed is as follows: A multi-feature coupled weighted discrimination model is constructed based on comprehensive adaptive weight coefficients. The normalized feature data of each independent spatial patch are substituted into the model to calculate the farmland outflow discrimination score. Based on the preset discrimination threshold, the patches are divided into normal farmland patches, pseudo-change patches to be judged, and suspected farmland outflow patches. For pseudo-change patches to be determined, the intensity index of land cover change, the consistency index of temporal change, and the spectral feature matching index are obtained, and the comprehensive determination value of pseudo-change is calculated. Pseudo-change patches are removed according to the determination threshold. Pseudo-change patches that are not determined as pseudo-changes are merged with suspected farmland outflow patches to form a set of farmland outflow patches to be confirmed.

[0012] Preferably, the specific process of calculating the farmland outflow discrimination score for each independent spatial patch and completing the initial patch judgment, and compiling and constructing the set of farmland outflow patches to be confirmed is as follows: The model training sample set is divided into a model training set and a model validation set according to a preset ratio. Multimodal classification features are extracted based on multi-source fusion and registration remote sensing images. The feature data is input into the support vector machine model to complete training and accuracy verification, and an optimized support vector machine classification model is obtained. The model is used to classify and identify the farmland outflow plots to be confirmed, and the final farmland outflow plots are determined after removing normal crop rotation plots; Extract the attribute information of the final farmland outflow patches to form a standard attribute dataset of farmland outflow patches, generate a full-area farmland outflow monitoring vector layer and complete classification and statistics, integrate relevant data and push it to the farmland dynamic monitoring business port to complete the full-area farmland outflow remote sensing monitoring.

[0013] Compared with the prior art, the beneficial effects of the present invention are: (1) This multi-source information remote sensing monitoring platform for farmland protection extracts farmland features from four dimensions: intensity of land cover change, vegetation growth cycle pattern, entropy value of cultivated texture, and linearity of field ridges. It corrects the feature data affected by local anomalies by using the average feature value of surrounding adjacent patches, and integrates and normalizes various features to form multi-dimensional core feature data. It can completely depict the static attributes and dynamic changes of farmland, weaken the data deviation caused by local anomalies, and truly restore the actual state of farmland. It provides sufficient and stable feature basis for farmland status judgment and improves the problem of single feature dimension and insufficient data credibility of traditional monitoring.

[0014] (2) This multi-source information-based remote sensing monitoring platform for farmland protection assigns adaptive weights to the actual contribution of various features to the judgment of farmland outflow, constructs a multi-feature collaborative coupling discrimination model to complete the initial judgment of farmland outflow, and combines the intensity of land cover change, consistency of temporal change, and spectral feature matching index to filter out false change interference such as cloud shadows and local isolated changes, optimizes the farmland outflow identification logic, reduces monitoring misjudgment, and makes the initial screening results of farmland outflow in a large area more consistent with the actual change scenario, thereby improving the overall credibility and practicality of farmland outflow discrimination work.

[0015] (3) This multi-source information-based remote sensing monitoring platform for farmland protection adopts a hierarchical discrimination process that combines preliminary screening and precise confirmation. It first completes the rapid preliminary screening and false change removal of all map patches in the entire area, and then uses a multi-modal remote sensing feature training model to achieve fine classification of map patches, clearly distinguishing between illegal outflow of farmland and regular crop rotation behavior. At the same time, it transforms the monitoring results into standardized results and connects them with the daily supervision of farmland, taking into account both the efficiency of monitoring the entire area and the classification effect, and providing solid technical support for the normalized and standardized management and control of farmland resources. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1; Please see Figure 1 This invention provides a remote sensing monitoring platform for farmland protection based on multi-source information, comprising: Multi-source data preprocessing module: Acquires multi-source fused and registered remote sensing images, digital differential images, land cover interpretation rules, and standardized temporal image data for the entire Xishuangbanna Prefecture, integrates them to form a standardized basic dataset for cultivated land monitoring, and performs independent spatial patch delineation processing on the multi-source fused and registered remote sensing images. The specific process is as follows: Acquire multi-source fusion registered remote sensing images of the entire Xishuangbanna Prefecture after spatial registration and multi-source fusion processing, and simultaneously acquire digital differential images of the entire prefecture, land cover interpretation rules, and time-series standardized image data, and integrate them to form a standardized basic dataset for farmland monitoring; Multi-source fusion and registration remote sensing imagery: refers to a comprehensive remote sensing imagery that integrates optical, radar, and hyperspectral multimodal features and has a unified spatial location. Global digital differential imagery: refers to image data formed through multi-temporal image difference operations, which can enhance the subtle changes in ground features; Land use interpretation rules: These are the reference criteria for distinguishing land features such as cultivated land, forest land, construction land, and water areas. They include the spectral thresholds, texture parameters, spatial distribution, and temporal variation characteristics of various land features. Time-series standardized image data: refers to image data that has undergone radiometric correction, missing data completion and normalization, is continuous in time and has uniform radiometric brightness, and can be used for time-series variation analysis; The existing multi-scale segmentation method is used to process the multi-source fusion and registration remote sensing image (including selecting eight neighboring pixels as the judgment object, setting the spectral Euclidean distance and texture similarity judgment thresholds, and aggregating and classifying the image pixels), to divide the image into independent spatial patches with homogeneous spectral and texture features, and to number all independent spatial patches in order of spatial coordinates (this process is the existing conventional processing method for remote sensing image patch division, and will not be described in detail). Among them, independent spatial patches refer to blocky image units in multi-source fusion and registration remote sensing images that have uniform and consistent internal spectral and textural features and independent and non-overlapping spatial boundaries.

[0019] It should be noted that the existing conventional data processing and spatial division methods are used to complete the integration of multi-source data and the division of independent spatial patches, mainly to lay the groundwork for the subsequent farmland monitoring process. By integrating multiple types of remote sensing data into a standardized dataset, the data used in subsequent steps such as feature extraction and model discrimination can be more consistent and have better adaptability. Delineating standardized independent spatial patches can also fit the actual spatial distribution of cultivated land, providing suitable spatial units for subsequent operations such as land feature change analysis and vegetation characteristic calculation, allowing the entire cultivated land remote sensing monitoring process to proceed smoothly and orderly.

[0020] The multi-dimensional feature extraction module constructs a spatial mapping between independent spatial patches and the global digital differential image, calculates the intensity of land cover changes in independent spatial patches, analyzes the vegetation characteristics of independent spatial patches, and constructs a subset of vegetation growth periodicity features. It extracts the tillage texture entropy and field ridge linearity of independent spatial patches and corrects for local perturbations. Finally, it integrates various features to generate a multi-dimensional core feature dataset. The specific process is as follows: For each independent spatial patch, the corresponding spatial boundary coordinates are obtained. Based on the unified spatial reference system shared by the independent spatial patches and the global digital differential image, the spatial boundary coordinates of the independent spatial patches are projected and matched with the pixel coordinates of the global digital differential image to establish a one-to-one spatial mapping relationship between the boundary range of the independent spatial patches and the pixel set of the digital differential image. Traverse the entire digital differential imagery and extract all pixels that fall within the spatial boundary coordinate range of an independent spatial patch to form the basic pixel set of land cover changes for that independent spatial patch. For each pixel in the basic pixel set of ground feature changes, the radiance value of the pixel in the two adjacent time-series standardized images is extracted; wherein, the radiance value is obtained by reading the radiance stored value corresponding to the pixel in the time-series standardized image, and the time-series standardized image pre-stores the radiance data of each pixel in each time phase. The absolute value of the difference between the two radiance values ​​corresponding to a pixel in two adjacent time-normalized images is calculated to obtain the single-pixel time-series variation value of that pixel. For the temporal variation values ​​of all single pixels within the independent spatial patch, calculate the initial arithmetic mean and initial standard deviation of the temporal variation values ​​of all single pixels within the independent spatial patch; Remove extreme outliers that exceed a preset multiple of the initial standard deviation; recalculate the arithmetic mean and standard deviation based on the temporal variation values ​​of the remaining valid pixels; repeat the above removal and calculation process until the difference between the arithmetic means obtained from two consecutive calculations is less than the preset convergence threshold. The arithmetic mean of the temporal variation values ​​of all remaining valid pixels is used to obtain the land feature variation intensity value of this independent spatial patch; Based on time-series standardized image data, and combined with the same spatial reference frame of independent spatial patches and time-series standardized images, the spatial boundary coordinates of independent spatial patches are matched with the pixel coordinates of time-series standardized images. Image blocks in time-series standardized images that completely overlap with the spatial boundary range of independent spatial patches are marked as the image regions corresponding to independent spatial patches. According to the acquisition time sequence of time-series standardized images, the image regions corresponding to independent spatial patches in each time phase are extracted. For the corresponding image region of the independent spatial patch under a single time phase, each pixel in the region is traversed, and the near-infrared band reflectance and red band reflectance corresponding to the pixel are extracted respectively. The difference between the near-infrared band reflectance and the red band reflectance is used as the numerator, and the sum of the two is used as the denominator. The ratio of the numerator to the denominator is calculated to obtain the normalized vegetation index corresponding to each pixel. A linear correspondence between the normalized vegetation index and vegetation cover is preset (this linear correspondence can be extracted from the land cover interpretation rules, which store the spectral characteristics of various land features, the correspondence between vegetation index and vegetation cover, and the linear correspondence between the normalized vegetation index and vegetation cover can be preset based on this correspondence). By matching the normalized vegetation index of each pixel in the corresponding image area of ​​an independent spatial patch with the preset linear correspondence, the normalized vegetation index of each pixel is converted into the corresponding vegetation cover value. The arithmetic mean of the vegetation cover values ​​of all pixels within the image area corresponding to the independent spatial patch is calculated to obtain the average vegetation cover of the independent spatial patch in the current time phase. According to the temporal sequence of the time-standardized images, the average vegetation cover of independent spatial patches in each temporal phase is integrated sequentially to obtain the multi-temporal vegetation cover sequence of independent spatial patches. The arithmetic mean and standard deviation of the vegetation cover sequence in the independent spatial patches were calculated respectively. The ratio of the arithmetic mean to the standard deviation was calculated to obtain the temporal variation coefficient of vegetation cover in the independent spatial patches. Traverse the multi-temporal vegetation cover sequence of independent spatial patches, extract the maximum value in the sequence as the vegetation growth peak value of the independent spatial patch, and the minimum value as the vegetation growth valley value of the independent spatial patch. A preset peak ratio threshold is used to multiply the peak vegetation growth rate by the preset peak ratio threshold to obtain the critical value of the peak vegetation growth rate. The number of consecutive time-series nodes in a multi-temporal vegetation cover sequence with vegetation cover greater than the critical value of vegetation growth peak is counted, and this count is recorded as the duration of vegetation growth peak. By integrating the vegetation growth peak, vegetation growth valley and peak duration of independent spatial patches, a subset of vegetation growth periodicity features of independent spatial patches is constructed. For each independent spatial patch, based on the pre-determined criteria for farmland texture and field ridge morphology within the land use interpretation rules, image regions containing only optical band data and consistent with the spatial extent of the independent spatial patch are selected and extracted from the multi-source fused and registered remote sensing images, and recorded as the optical band image region corresponding to the independent spatial patch. The land use interpretation rules store texture feature standards, field ridge morphology judgment indicators, and matching analysis band types for various land features. This type of information can be directly retrieved from the standardized farmland monitoring basic dataset constructed by the multi-source data preprocessing module. The optical band image regions corresponding to independent spatial patches are preprocessed by grayscale conversion (removing multi-band color spectral information and retaining only the pixel brightness and darkness grayscale numerical characteristics) to obtain single-channel grayscale images corresponding to independent spatial patches; where single-channel grayscale images refer to simplified image data containing only a single grayscale channel and without redundant color band information. The window size, gray level, and step size parameters required for gray-level co-occurrence matrix analysis are preset (which can be determined based on the typical spatial scale of cultivated land texture features in the land use interpretation rules, and should be consistent with the cultivated land spatial scale standard used in the aforementioned criteria for determining cultivated land texture and field ridge morphology features). The image is traversed window by window according to the preset window size. The gray values ​​of pixels in each window are statistically analyzed according to the preset gray level. The frequency of the combination of gray values ​​of adjacent pixels is recorded in combination with the preset step size. The frequency of occurrence of gray-level combinations of adjacent pixels corresponding to each window is summarized to form a gray-level co-occurrence matrix that can reflect the texture distribution pattern of the image area. Traverse the probability values ​​of all gray-level combinations in the gray-level co-occurrence matrix, calculate the product of each probability value and its natural logarithm, sum all the product results and take the negative value to obtain the cultivation texture entropy value of the independent spatial patch. Among them, the tillage texture entropy value is used to quantitatively characterize the degree of disorder of the tillage texture of the cultivated land corresponding to the independent spatial patch. The higher the entropy value, the higher the disorder of the texture, and the lower the entropy value, the higher the regularity of the texture. For a single-channel grayscale image corresponding to an independent spatial patch, preset the high threshold parameter and low threshold parameter required by the Canny edge detection algorithm; Based on the above preset threshold parameters, the Canny edge detection algorithm is used to perform pixel-by-pixel traversal recognition of the single-channel grayscale image, and to filter and extract all edge pixels representing the boundaries of ground features within the single-channel grayscale image area. Contour tracking is performed based on the adjacency of pixel spatial positions. The edge pixels that are adjacent to each other in space and have continuous directions are connected sequentially to form several complete and continuous linear contours, which are all the continuous linear contours within the single-channel grayscale image area corresponding to the independent spatial patch. For each continuous linear contour, traverse all pixels on the continuous linear contour, calculate the Euclidean distance between two adjacent pixels in turn, and sum the Euclidean distances of all adjacent pixels to obtain the contour length of the continuous linear contour. Extract the coordinates of the starting and ending pixels on the continuous linear contour, determine the direction of the contour by connecting the two points, calculate the angle between the line connecting the two points and the preset reference direction (such as the horizontal direction), and obtain the direction angle of the continuous linear contour. Preset contour length threshold and angle threshold (the angle threshold is used to determine the degree of fit between the contour direction and the field ridge direction in the land cover interpretation rules); For all continuous linear contours of independent spatial patches, the contour length and orientation angle of each contour are compared with the preset length threshold and preset angle threshold respectively. Continuous linear contours that simultaneously satisfy the following conditions are selected: contour length is greater than the preset length threshold and orientation angle is less than the preset angle threshold. These are recorded as candidate contours for field ridges. For each candidate contour of a field ridge, extract the spatial coordinate values ​​of all pixels on that contour. Using the spatial coordinates (xi, yi) of all pixels on the candidate contour of the field ridge as the fitting input data, based on the principle of least squares, with the goal of minimizing the sum of squared distances from all pixel coordinates to the fitted line, the parameters a (slope) and b (intercept) in the fitted line y=ax+b are iteratively solved until the parameters a and b tend to stabilize (the difference between the parameters in two consecutive iterations is less than the preset convergence threshold), and finally the fitted line equation y=ax+b corresponding to the candidate contour of the field ridge is obtained. Based on the fitted line equation corresponding to the candidate contour of the field ridge, the spatial coordinate value (xi,yi) of each pixel on the contour is substituted into the equation. The vertical distance between each pixel coordinate and the corresponding fitted line is calculated using the Euclidean distance calculation formula. This vertical distance is recorded as the fitting deviation of the pixel. The average fitting deviation of the candidate contour of the field ridge is obtained by arithmetically averaging the fitting deviations of all pixels on the same candidate contour of the field ridge; the linearity of the candidate contour of the field ridge is obtained by taking the reciprocal of the average fitting deviation. The linearity of the field ridges in the independent spatial patch is obtained by arithmetically averaging the linearity of all candidate contours of the field ridges. The linearity of field ridges is used to quantitatively characterize the regularity of the field ridges corresponding to an independent spatial patch: the higher the linearity value, the more regular the field ridges are and the more the arable land morphology conforms to the characteristics of standard arable land; the lower the linearity value, the more disordered the field ridges are and the more the arable land morphology deviates from the characteristics of standard arable land. For each independent spatial patch, taking the independent spatial patch as the center, extract the cultivation texture entropy value and field ridge linearity of the eight surrounding independent spatial patches; calculate the Pearson correlation coefficient of the cultivation texture entropy value and the Pearson correlation coefficient of the field ridge linearity between the independent spatial patch and each of the adjacent independent spatial patches; count the number of correlation coefficients of the two types that are less than the preset correlation threshold. When the number of correlation coefficients below the threshold corresponding to any feature meets the preset judgment condition, it is determined that the independent spatial patch has local perturbation. If there is local disturbance in an independent spatial patch, the cultivation texture entropy value and field ridge linearity corresponding to that independent spatial patch are removed, and the arithmetic mean of the features corresponding to the eight surrounding independent spatial patches is used to replace the cultivation texture entropy value and field ridge linearity of that independent spatial patch. The intensity values ​​of land cover changes, the temporal variation coefficient of vegetation cover, the subset of vegetation growth periodicity features, the entropy value of tillage texture, and the linearity of field ridges of the independent spatial patch are integrated. Normalization and dimensionless processing are performed on the above-mentioned feature data to form a multidimensional core feature dataset corresponding to the independent spatial patch. After all the above processing steps have been completed for all independent spatial patches, the multidimensional core feature dataset corresponding to each independent spatial patch is obtained.

[0021] It should be noted that feature extraction from multiple dimensions, including intensity of land cover changes, vegetation growth cycle, cultivation texture and field ridge linearity, can comprehensively depict the static morphological attributes and dynamic change characteristics of cultivated land. It can accurately reflect the degree of land cover changes and vegetation growth patterns, and also intuitively reflect the cultivation texture and field ridge regularity, providing comprehensive and detailed feature basis for cultivated land status determination and outflow monitoring. For the entropy value of tillage texture and linearity data of field ridges with local disturbances, the feature mean of the surrounding adjacent patches is used for replacement and correction. This can effectively remove the interference caused by local abnormal data, reduce the impact of noise and non-tillage factors on the feature results, and make the extracted feature data more consistent with the real spatial form and attribute characteristics of tillage. After integrating multiple features such as intensity of land cover change, vegetation characteristics, and cultivation texture, and performing normalization and dimensionless processing, a unified and standardized multidimensional core feature dataset is formed. This dataset can be efficiently connected with the early data preprocessing stage and the later cultivated land outflow discrimination module, directly adapting to the usage requirements of model discrimination and monitoring analysis, and ensuring the stable and smooth progress of the entire cultivated land remote sensing monitoring process.

[0022] Farmland outflow monitoring module: Constructs a multi-feature coupled weighted discrimination model, analyzes the feature weights of each independent spatial patch, calculates the farmland outflow discrimination score of each independent spatial patch and completes the initial patch judgment, and organizes and constructs a set of farmland outflow patches to be confirmed; trains a support vector machine model to classify and discriminate the set of farmland outflow patches to be confirmed, determines the final farmland outflow patches, and monitors the farmland outflow status of the entire region. The specific process is as follows: Select cultivated land parcel samples that have been verified in the field and confirmed by land type attributes to construct a model training sample set; The training sample set includes four types of samples: normal cultivated land parcels, conventional crop rotation parcels, cultivated land outflow parcels for non-agricultural use, and cultivated land outflow parcels for non-grain use. The samples cover cultivated land units with different terrains, farming patterns and outflow types throughout the region. Each sample in the training sample set corresponds to a multi-dimensional feature dataset, which consists of normalized land cover change intensity values, vegetation cover temporal variation coefficient, vegetation growth periodicity feature subset, tillage texture entropy value, and field ridge linearity. The multi-dimensional feature dataset of the training sample set and the corresponding land use attributes are input into the random forest discriminant model to construct a binary classification discriminant model with the classification objective of distinguishing between arable land outflow and non-outflow. Among them, normal arable land plots and conventional crop rotation plots are non-outflow samples, while arable land non-agricultural outflow plots and arable land non-grain outflow plots are outflow samples. The out-of-bag error method was used to calculate the overall importance score of each feature to the farmland outflow discrimination result. The formula for calculating the overall importance score of a single feature is as follows: , Where: is the comprehensive importance score of the \(i\)-th type of feature. The higher the score, the greater the influence of this type of feature on the discrimination of cultivated land outflow; \(N\) is the total number of decision trees in the random forest model, and the value can be 50 - 100 trees; is the classification error rate of cultivated land outflow obtained based on out-of-bag data after randomly permuting and perturbing the \(i\)-th type of feature by the \(j\)-th decision tree; is the classification error rate of cultivated land outflow obtained by the \(j\)-th decision tree based on the original out-of-bag data; \(i\) is the feature serial number, with values from 1 to 5, corresponding in sequence to the land cover change intensity value, the temporal variation coefficient of vegetation coverage, the vegetation growth periodicity feature subset, the tillage texture entropy value, and the linearity of the ridge. \(j\) is the decision tree serial number, with values from 1 to \(N\); Normalize the comprehensive importance scores of the five types of features to obtain the comprehensive adaptive weight coefficients corresponding to each type of feature; Construct a multi-feature coupled weighted discrimination model based on the comprehensive adaptive weight coefficients. Substitute the normalized feature data of each independent spatial patch in the whole area of Xishuangbanna Prefecture into the multi-feature coupled weighted discrimination model to calculate the cultivated land outflow discrimination score of the independent spatial patch: , Where: \(S\) is the cultivated land outflow discrimination score of the independent spatial patch; is the normalized feature value of the \(i\)-th type of feature; is the comprehensive adaptive weight coefficient of the \(i\)-th type of feature; Preset the first discrimination threshold \(S1\) and the second discrimination threshold \(S2\), and \(S_{1}<S_{2}\); For all independent spatial patches in the whole area, compare the discrimination score \(S\) of each independent spatial patch with the preset first \(S1\) and the second discrimination threshold \(S2\): If \(S < S_{1}\), it is determined as a normal cultivated land patch; If \(S_{1}\leq S\leq S_{2}\), it is determined as a pseudo-change to-be-determined patch; If \(S > S_{2}\), it is determined as a suspected cultivated land outflow patch; For each patch determined as a pseudo-change to-be-determined, select its corresponding land cover change intensity index \(P1\), temporal change consistency index \(P2\), and spectral feature matching index \(P3\), and calculate the pseudo-change comprehensive determination value. The calculation formula is: \(P = P1\times a1+P2\times a2+P3\times a3\) to obtain the pseudo-change comprehensive determination value \(P\); Where, \(a1\), \(a2\), \(a3\) are preset weight coefficients, and \(a1 + a2 + a3 = 1\); Furthermore, for each map patch identified as a pseudo-change awaiting further evaluation, the process for obtaining the land cover change intensity index P1, the temporal change consistency index P2, and the spectral feature matching index P3 is as follows: Land feature change intensity index: Obtain the land feature change intensity value of the pseudo-change patch to be judged, and compare it with the maximum change intensity threshold of normal cultivated land preset in the land use interpretation rules; If the intensity value of land feature change does not exceed the threshold, the index value is determined according to the ratio of the intensity value of land feature change to the threshold. If the intensity value of land feature change exceeds the threshold, the index value is directly set to 1. Consistency index of temporal change: Taking the pseudo-change undetermined patch as the center, count the number of patches that have undergone land cover changes synchronously with the eight adjacent independent spatial patches around it; Calculate the proportion of synchronously changing patches to the total number of adjacent patches, and compare this proportion with a preset change consistency threshold; If the ratio is lower than the preset threshold, it indicates that the change of the patch is a local isolated change and a non-regional synchronous change, then the index value is set to 1; If the ratio is not lower than the preset threshold, it indicates that the change has regional consistency, and the index value is set to 0. Spectral feature matching index: Extract the average spectral reflectance data of the pseudo-change patch in the image at the time of the change (this data can be obtained by performing routine radiometric calibration and atmospheric correction on the image at the time of the change, cropping the pixel reflectance within the patch area and taking the arithmetic mean); match this average spectral reflectance data with the standard spectral reflectance features of clouds and shadows respectively to obtain the matching degree between the spectrum of the patch and the spectrum of clouds and shadows, and take the highest matching degree as the value of the spectral feature matching index, where the matching degree value is from 0 to 1; A false change judgment threshold is preset. If the false change comprehensive judgment value P of the false change to be judged patch is greater than or equal to the false change judgment threshold, the patch is judged as a false change patch and removed. If the comprehensive value of the pseudo-change of the pseudo-change plot is less than the pseudo-change judgment threshold, the plot will be merged with the suspected farmland outflow plot to form a set of farmland outflow plots to be confirmed. The aforementioned model training sample set was divided into a model training set and a model validation set according to a preset ratio. Using support vector machine as the classification model framework and land use attributes as classification labels, the average optical spectral reflectance, average radar backscattering coefficient, and average hyperspectral feature band reflectance of each sample in the training set were extracted as multimodal classification features for model training based on the multi-source fusion and registration remote sensing images obtained by the multi-source data preprocessing module. The above three types of features are the original pixel values ​​of optical bands, radar bands, and hyperspectral bands in the multi-source fusion and registration remote sensing images, which can be directly obtained through patch boundary cropping and arithmetic mean calculation. Multimodal classification feature data is input into the support vector machine model for iterative training. After the model is trained, the classification accuracy of the model is verified using a model validation set. The model parameters are continuously optimized based on the verification results until the classification accuracy of the model tends to stabilize and meets the preset threshold. Finally, the trained support vector machine classification model is obtained. Based on the land classification results output by the support vector machine classification model, the land parcels that were identified as normal crop rotation were removed from the land parcels to be confirmed as farmland outflows. The two types of land parcels identified as farmland outflows for non-agricultural purposes and farmland outflows for non-grain purposes were retained and used as the final farmland outflow parcels. Extract the relevant attribute information of all final farmland outflow patches, including: spatial boundary coordinates, actual area of ​​the patch, outflow type, time of change and spatial location information. Organize and summarize the above information to form a standardized farmland outflow patch standard attribute dataset. Based on the standard attribute dataset of cultivated land outflow patches, and relying on a unified spatial reference system, a full-area cultivated land outflow monitoring vector layer is generated. Among them, the monitoring vector layer fully carries the spatial geometric information and attribute fields of each final farmland outflow patch; The statistical data of all farmland outflow plots are classified and statistically analyzed by administrative region, outflow type, and time period of change. The distribution quantity, coverage area and regional proportion of each type of plot are calculated, and a standardized and unified statistical report on farmland outflow is compiled. The monitoring vector layers, classification statistical reports, and farmland outflow attribute data are uniformly integrated and archived, and simultaneously pushed to the farmland dynamic monitoring business port to provide data support for regional farmland protection and control, change verification and analysis, and complete the whole-process remote sensing monitoring of farmland outflow across the entire region.

[0023] It should be noted that by quantifying the comprehensive importance of multi-dimensional features and generating adaptive weights based on the random forest model, a multi-feature coupled weighted discrimination model is constructed. This model can reasonably assign weights based on the actual contribution of different features to the discrimination of farmland outflow, avoiding the dominance of a single feature in the discrimination result. This makes the basis for the initial judgment of farmland outflow more scientific and balanced, and more in line with the actual discrimination pattern of farmland outflow. By constructing a triple pseudo-change judgment index based on the intensity of land cover change, consistency of temporal change, and spectral feature matching, it is possible to accurately identify and eliminate pseudo-change patches caused by non-real outflow of cultivated land, such as cloud, shadow, and local isolated changes, which can significantly reduce the probability of misjudgment during the monitoring process and ensure the authenticity and credibility of cultivated land outflow judgment results. The support vector machine model is trained using the multimodal features of multi-source fusion and registration remote sensing images. After iterative optimization and accuracy verification, the model achieves refined classification of the plots to be confirmed. It can accurately distinguish between non-agricultural outflow, non-grain outflow and conventional crop rotation status of cultivated land, refine the determination of cultivated land outflow type, and meet the classification requirements of precise monitoring of cultivated land. Through a two-layer discrimination mechanism of "initial screening and elimination - precise confirmation", the rapid initial screening and false change removal of all map patches in the whole area are completed first, and then high-precision classification and discrimination are carried out. This takes into account both the processing efficiency and discrimination accuracy of large-area farmland monitoring, and is adapted to the actual work needs of rapid monitoring of farmland outflow in the whole area. Integrating all dimensions of information such as spatial boundaries, area, type, and time sequence of farmland outflow patches, a standardized attribute dataset, monitoring vector layer, and classification statistical report are generated, transforming scattered discrimination results into standardized monitoring results. The data format and content are adapted to farmland supervision operations. The monitoring results are simultaneously pushed to the farmland dynamic monitoring business portal, which opens up the data link between remote sensing monitoring and farmland protection and control, allowing monitoring data to directly serve the verification of illegal outflow and regional control decisions, and achieving seamless connection between farmland outflow monitoring and actual supervision work; A complete monitoring closed loop has been established, from sample construction, model training, patch identification, classification confirmation to output results. Each link is smoothly connected and the process is standardized, which can support the normalized and long-term remote sensing monitoring of the outflow of cultivated land across the entire region and improve the standardization level of cultivated land protection monitoring.

[0024] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A remote sensing monitoring platform for farmland protection based on multi-source information, characterized in that, include: Multi-source data preprocessing module: acquires multi-source fusion and registration remote sensing images, digital differential images, land cover interpretation rules and time-series standardized image data of the entire Xishuangbanna Prefecture, integrates them to form a standardized basic dataset for farmland monitoring, and performs independent spatial patch division processing on the multi-source fusion and registration remote sensing images; Multidimensional feature extraction module: Constructs spatial mapping between independent spatial patches and global digital differential imagery, calculates the intensity value of land cover change in independent spatial patches; analyzes vegetation characteristics of independent spatial patches, and constructs a subset of vegetation growth periodicity features; Extract the cultivation texture entropy value and field ridge linearity of independent spatial patches and complete the local perturbation correction; integrate various features to generate a multi-dimensional core feature dataset; Farmland outflow monitoring module: Construct a multi-feature coupled weighted discrimination model, analyze the feature weights of each independent spatial patch, calculate the farmland outflow discrimination score of each independent spatial patch and complete the initial judgment of the patch, organize and construct a set of farmland outflow patches to be confirmed; train a support vector machine model to classify and discriminate the set of farmland outflow patches to be confirmed, determine the final farmland outflow patches, and monitor the farmland outflow status of the entire region.

2. The farmland protection remote sensing monitoring platform based on multi-source information according to claim 1, characterized in that: The specific process of integrating and forming a standardized basic dataset for farmland monitoring, and performing independent spatial patch delineation on multi-source fused and registered remote sensing images, is as follows: The project aims to acquire multi-source fusion registered remote sensing images of the entire Xishuangbanna Prefecture after spatial registration and multi-source fusion processing, and simultaneously acquire digital differential images of the entire prefecture, land cover interpretation rules, and time-series standardized image data, and integrate them to form a standardized basic dataset for farmland monitoring. Existing multi-scale segmentation methods are used to process multi-source fusion and registration remote sensing images, resulting in independent spatial patches with homogeneous spectral and textural features. All independent spatial patches are then numbered according to their spatial coordinates.

3. The farmland protection remote sensing monitoring platform based on multi-source information according to claim 2, characterized in that: The specific process for constructing a spatial mapping between independent spatial patches and the global digital differential image, and calculating the intensity values ​​of land cover changes in independent spatial patches, is as follows: For each independent spatial patch, the corresponding spatial boundary coordinates are obtained. Based on a unified spatial reference system, the spatial boundary coordinates of the patch are projected and matched with the pixel coordinates of the global digital differential image to establish a one-to-one spatial mapping relationship. Traverse the entire digital differential image to extract all pixels within the patch area and form a basic pixel set of ground feature changes; Extract the radiance values ​​of each pixel in the time-normalized images of two adjacent periods, and calculate the absolute value of the difference to obtain the time-series change value of a single pixel; Calculate the initial mean and standard deviation of the temporal change values ​​within the patch, iteratively remove outliers and recalculate until the mean converges, and take the mean of the temporal change values ​​of the final effective pixels to obtain the ground feature change intensity value.

4. The farmland protection remote sensing monitoring platform based on multi-source information according to claim 3, characterized in that: The specific process of analyzing the vegetation characteristics of independent spatial patches and constructing a subset of vegetation growth periodicity features is as follows: Based on a unified spatial reference system, the coordinates of independent spatial patches and time-normalized images are matched, and the image regions corresponding to the patches in each time phase are extracted. The normalized vegetation index of the pixels in the region is calculated and converted into vegetation cover to obtain the multi-time phase vegetation cover sequence of the patches. The mean and standard deviation of vegetation cover in multiple time phases are calculated to obtain the time series variation coefficient. The peak and trough values ​​of the sequence are extracted and the duration of the peak values ​​is statistically analyzed. The above parameters are integrated to construct a subset of vegetation growth periodicity features.

5. The farmland protection remote sensing monitoring platform based on multi-source information according to claim 4, characterized in that: The specific process for extracting the cultivation texture entropy value and field ridge linearity of independent spatial patches is as follows: Based on the land cover interpretation rules, the optical band image region corresponding to the independent spatial patch is extracted from the multi-source fusion and registration remote sensing image. The optical band image region is then processed into grayscale to obtain the single-channel grayscale image corresponding to the independent spatial patch. Based on the analysis of single-channel grayscale images using the gray-level co-occurrence matrix, the cultivation texture entropy value of independent spatial patches is calculated. An edge detection algorithm is used to extract the linear contours of a single-channel grayscale image, and candidate contours of field ridges are obtained by screening. The candidate contours of field ridges are fitted with straight lines by the least squares method and the linearity is calculated. The arithmetic mean of the linearity of all candidate contours of field ridges is taken to obtain the linearity of field ridges of independent spatial patches.

6. The farmland protection remote sensing monitoring platform based on multi-source information according to claim 5, characterized in that: The specific process of completing local perturbation correction and generating a multidimensional core feature dataset is as follows: Centered on each independent spatial patch, the cultivation texture entropy value and field ridge linearity of the surrounding adjacent independent spatial patches are extracted; Calculate the correlation coefficients of the tillage texture entropy values ​​and the correlation coefficients of the field ridge linearity between the independent spatial patch and its adjacent independent spatial patches, respectively. The number of correlation coefficients of tillage texture entropy and field ridge linearity that are less than the preset correlation threshold is counted separately. When the preset judgment condition is met, it is determined that the independent spatial patch has local disturbance. For map patches with local disturbances, the arithmetic mean of the corresponding features of the surrounding adjacent map patches is used for replacement; the intensity value of land cover change, the temporal variation coefficient of vegetation coverage, the subset of vegetation growth periodic features, the entropy value of tillage texture and the linearity of field ridges are integrated and normalized to generate a multidimensional core feature dataset.

7. A remote sensing monitoring platform for farmland protection based on multi-source information as described in claim 6, characterized in that: The specific process of constructing a multi-feature coupled weighted discrimination model and analyzing the feature weights of each independent spatial patch is as follows: A training sample set was constructed by selecting farmland patch samples that had been verified in the field and confirmed by land type attributes. The samples included four types of farmland patches and corresponding multi-dimensional feature datasets. The multi-dimensional feature datasets and corresponding land type attributes were input into a random forest model to construct a binary classification model for farmland outflow discrimination. The out-of-bag error method is used to calculate the overall importance score of each feature. The overall importance score is then normalized to obtain the overall adaptive weight coefficients for each feature.

8. A remote sensing monitoring platform for farmland protection based on multi-source information as described in claim 7, characterized in that: The specific process of calculating the farmland outflow discrimination score for each independent spatial patch, completing the initial patch judgment, and organizing and constructing the set of farmland outflow patches to be confirmed is as follows: A multi-feature coupled weighted discrimination model is constructed based on comprehensive adaptive weight coefficients. The normalized feature data of each independent spatial patch are substituted into the model to calculate the farmland outflow discrimination score. Based on the preset discrimination threshold, the patches are divided into normal farmland patches, pseudo-change patches to be judged, and suspected farmland outflow patches. For pseudo-change patches to be determined, the intensity index of land cover change, the consistency index of temporal change, and the spectral feature matching index are obtained, and the comprehensive determination value of pseudo-change is calculated. Pseudo-change patches are removed according to the determination threshold. Pseudo-change patches that are not determined as pseudo-changes are merged with suspected farmland outflow patches to form a set of farmland outflow patches to be confirmed.

9. A remote sensing monitoring platform for farmland protection based on multi-source information as described in claim 8, characterized in that: The specific process of calculating the farmland outflow discrimination score for each independent spatial patch, completing the initial patch judgment, and organizing and constructing the set of farmland outflow patches to be confirmed is as follows: The model training sample set is divided into a model training set and a model validation set according to a preset ratio. Multimodal classification features are extracted based on multi-source fusion and registration remote sensing images. The feature data is input into the support vector machine model to complete training and accuracy verification, and an optimized support vector machine classification model is obtained. The model is used to classify and identify the farmland outflow plots to be confirmed, and the final farmland outflow plots are determined after removing normal crop rotation plots; Extract the attribute information of the final farmland outflow patches to form a standard attribute dataset of farmland outflow patches, generate a full-area farmland outflow monitoring vector layer and complete classification and statistics, integrate relevant data and push it to the farmland dynamic monitoring business port to complete the full-area farmland outflow remote sensing monitoring.

Images