A method and system for monitoring and identifying non-grain cultivation of farmland

Abstract

The present application belongs to the technical field of agricultural remote sensing monitoring and cultivated land use identification, and discloses a cultivated land non-grain monitoring and identification method and system. The method obtains multi-temporal remote sensing images of a target area, grain function area vector boundary data and sample annotation data, generates annual multi-band feature images, constructs a random forest multi-classification model and a support vector machine binary classification model, outputs double-model non-grain identification results and confidence, rasterizes the grain function area vector boundary into a mask, spatially constrains the double-model results, takes the union of non-grain patches within the mask range, combines random forest confidence, support vector machine confidence, model consistency markers and band similarity penalty items to calculate cross-validation scores, and divides and confirms patches, to-be-reviewed patches and excluded patches according to the scores. The present application is suitable for remote sensing monitoring, patch review and statistics of cultivated land non-grain in a grain function area.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing monitoring and farmland protection information technology, specifically to a method and system for monitoring and identifying farmland conversion to non-grain crops. Background Technology

[0002] Arable land is a crucial spatial foundation for grain production. As key areas for ensuring stable grain production, grain functional zones require continuous monitoring of land use status for arable land protection, agricultural supervision, and grassroots verification. With the development of multi-temporal remote sensing imagery, geographic information systems, and machine learning classification technologies, using remote sensing imagery to identify changes in arable land planting structure, non-grain crop planting, abandonment of farmland, and aquaculture has become an important technical approach for arable land use monitoring.

[0003] Chinese invention patent application CN118608977A discloses a method for detecting non-grain conversion of arable land based on remote sensing imagery. This method acquires remote sensing images from different time phases, constructs a sample library including types of conversions such as arable land to orchards, arable land to forest land, arable land to facility agriculture land, arable land to ponds or reservoirs, and arable land to abandoned land. A non-grain conversion detection model is jointly constructed based on DASNet and PSPNet to extract non-grain conversion change patches. This technology can provide an automated identification method for detecting non-grain conversion changes under diverse agricultural patterns.

[0004] However, existing remote sensing technologies for identifying non-grain crops mostly focus on extracting changed patches or land cover categories from the image itself, typically outputting the model's identification results directly as candidate results. In actual regulatory scenarios, non-grain crop determination is not simply an image classification problem; it is also related to the boundaries of grain functional zones, the scope of cultivated land use control, and the needs of grassroots verification. If only the remote sensing classification results of the entire area are used for identification, it is easy to include legally existing orchards, seedlings, or other economic crop areas within non-grain functional zones in the non-grain crop output, resulting in inaccurate spatial applicability. At the same time, rice, dryland crops, orchards, seedlings, and some perennial economic crops have similar characteristics in multiple time-phase bands throughout the year. Using only a single model or a single classification result can easily lead to misjudgment or omission. Existing methods lack a patch-level cross-validation mechanism that combines grain functional zone mask constraints, random forest multi-classification results, support vector machine binary classification results, model consistency judgment, and similarity penalty for easily confused categories. This would allow for the simultaneous output of confirmed patches and the generation of patches to be verified and removed, meeting the hierarchical verification needs in grain functional zone supervision. Summary of the Invention

[0005] The technical objective of this invention is to provide a monitoring and identification method and system capable of performing dual-model cross-validation and diversion output on candidate non-grain land patches within the grain functional zone.

[0006] To achieve the above-mentioned technical objectives, the present invention provides the following technical solutions.

[0007] In a first aspect, the present invention provides a method for monitoring and identifying non-grain conversion of arable land, based on grain functional zone masking and dual-model cross-validation, comprising the following steps:

[0008] S1. Acquire multi-temporal remote sensing images of the target area, vector boundary data of grain functional zones and sample annotation data, and perform cloud cover filtering, cloud masking, image mosaicking, target area cropping and non-zero pixel annual mean synthesis on the multi-temporal remote sensing images to obtain annual multi-band feature images.

[0009] S2. Based on the sample annotation data, extract sample band features from the annual multi-band feature image to construct a first training sample set for multi-land cover classification and a second training sample set for non-grain binary classification.

[0010] S3. Train the random forest multi-classification model and the support vector machine binary classification model respectively, and output the first classification result, the second classification result and the corresponding declassified confidence score.

[0011] S4. Rasterize the vector boundary data of the grain functional area into a grain functional area mask that is consistent with the space of the annual multi-band feature image, and use the grain functional area mask to constrain the first classification result and the second classification result.

[0012] S5. Within the masked area of ​​the grain functional zone, take the union of the non-grain patches in the first classification result and the non-grain patches in the second classification result to generate candidate non-grain patches. Calculate the cross-validation score based on the random forest confidence, support vector machine confidence, model consistency marker, and the band similarity penalty term between the candidate non-grain patches and the center vectors of grain crop samples and orchard samples.

[0013] S6. Based on the cross-validation score, the candidate non-grain-related patches are divided into confirmed patches, patches to be reviewed, and patches to be removed, and the monitoring and identification results including patch boundaries, patch area, the number of the grain functional area to which it belongs, model consistency marker, and review priority are output.

[0014] Specifically, the multi-temporal remote sensing images are publicly acquired or authorized multispectral remote sensing images, and include at least the blue light band, green light band, red light band, and near-infrared band; the sample annotation data are jointly determined by Tianditu multi-temporal images, field survey data, and land use vector data; the sample annotation data includes food crop samples, non-food samples, and non-arable land background samples; the food crop samples include rice samples and dryland crop samples; the non-food samples include abandoned farmland samples, pond aquaculture samples, perennial cash crop samples, and seedling and garden crop samples; the non-arable land background samples include one or more of the following: orchard samples, water body samples, building samples, and forest land samples.

[0015] Specifically, in step S1, cloud cover filtering includes removing images with cloud coverage rates higher than a preset cloud cover threshold based on the cloud cover percentage in the remote sensing image metadata; cloud masking includes identifying cloud pixels based on a blue light band threshold and setting the cloud pixels to null values; non-zero pixel annual average synthesis includes averaging the non-zero valid observations of the same pixel location in multiple time phases, so that cloud pixels, invalid pixels, and zero-value pixels do not participate in the annual average calculation.

[0016] Specifically, in step S2, extracting sample band features includes reading the blue light band value, green light band value, red light band value, and near-infrared band value of the corresponding pixels of the sample annotation points or sample annotation patches, forming a sample feature vector. The sample feature vector satisfy:

[0017]

[0018] In the formula, For the first The sample feature vector of each sample; Values ​​for the blue light band; Values ​​for the green light band; Values ​​for the red light band; This is the value in the near-infrared band.

[0019] Specifically, in step S2, the first training sample set is labeled with multiple categories according to rice, dryland crops, non-grain crops, orchards, water bodies, buildings, and forest land; the second training sample set is obtained by reclassifying the first training sample set, wherein the non-grain category is labeled as having non-grain categories, and rice, dryland crops, orchards, water bodies, buildings, and forest land are labeled as not having non-grain categories; before training the random forest multi-classification model and the support vector machine binary classification model, sample balancing is performed on the first training sample set and / or the second training sample set.

[0020] Specifically, in step S3, the random forest multi-classification model outputs the first confidence score of each pixel or patch belonging to the non-grain category. The support vector machine binary classification model outputs the second confidence score for each pixel or patch belonging to the non-grain-differentiated category. The first confidence level The second confidence level is determined based on the number of decision trees in the random forest that vote for the non-food category. Determined based on the classification interval or probability calibration results of the support vector machine.

[0021] Specifically, in step S4, rasterizing the vector boundary data of the grain functional area into a grain functional area mask includes:

[0022] Transform the vector boundary data of grain functional zones into the coordinate system of annual multi-band feature images;

[0023] A binary mask is generated based on the pixel size, row and column range, and affine transformation parameters of the annual multi-band feature image.

[0024] When the center point of a pixel falls within the vector boundary of the grain functional area, or when a portion of the pixel area exceeding a preset proportion falls within the vector boundary of the grain functional area, the pixel is marked as a grain functional area pixel.

[0025] A binary mask is used to remove non-grain functional area pixels, so that non-grain functional area pixels do not participate in the non-grain conversion confirmation output.

[0026] Specifically, in step S5, the cross-validation score of candidate non-grain-based patches... satisfy:

[0027]

[0028] In the formula, For the first Cross-validation scores of candidate non-grain-based patches; For the first Random forest confidence of candidate non-grained patches; For the first Support vector machine confidence of each candidate de-grained patch; A model consistency marker is used when both the first and second classification results indicate that the model is consistent. When a candidate non-grain-type patch appears non-grain-type... Take the first value, otherwise take the second value; For easily confused penalty items; , , , These are the weighting coefficients.

[0029] More specifically, the aforementioned easily confused penalty item The similarity between the annual multi-band mean vector of candidate non-grain patches and the center vectors of rice, dryland crops, and orchards is determined, and the following conditions must be met:

[0030]

[0031] In the formula, For the first The easily confused penalty item for each candidate non-grain-based patch; For the first The annual multi-band mean vector of each candidate non-grained patch; The center vector of the rice sample; The center vector of the dryland crop sample; The center vector of the orchard sample; This is a vector similarity function;

[0032] Specifically, in step S6, when the cross-validation score is greater than or equal to the first scoring threshold, the corresponding candidate non-grain-based plot is classified as a confirmed plot; when the cross-validation score is less than the first scoring threshold but greater than or equal to the second scoring threshold, the corresponding candidate non-grain-based plot is classified as a plot to be reviewed; when the cross-validation score is less than the second scoring threshold, the corresponding candidate non-grain-based plot is classified as a plot to be removed; the review priority of the plot to be reviewed is determined based on the cross-validation score, the plot area, the proportion of non-grain-based area within the grain functional area, and the model consistency marker.

[0033] In a second aspect, the present invention also provides a monitoring and identification system for non-grain conversion of arable land, used to implement the monitoring and identification method for non-grain conversion of arable land described in the first aspect, the system comprising:

[0034] The data acquisition module is used to acquire multi-temporal remote sensing images of the target area, vector boundary data of grain functional zones, and sample annotation data;

[0035] The image preprocessing module is used to perform cloud cover filtering, cloud masking, image mosaicking, target area cropping, and annual mean synthesis of non-zero pixels on multi-temporal remote sensing images to obtain annual multi-band feature images.

[0036] The sample construction module is used to extract sample band features from annual multi-band feature images and construct the first training sample set and the second training sample set.

[0037] The dual-model recognition module is used to train a random forest multi-classification model and a support vector machine binary classification model, and outputs the first classification result, the second classification result, and the corresponding declassified confidence score.

[0038] The mask constraint module is used to rasterize the vector boundary data of grain functional areas into a grain functional area mask, and to constrain the first classification result and the second classification result using the grain functional area mask;

[0039] The cross-validation module is used to generate candidate de-grained patches and calculate cross-validation scores based on random forest confidence, support vector machine confidence, model consistency markers, and band similarity penalty terms.

[0040] The results output module is used to classify confirmed patches, patches to be reviewed, and patches to be removed based on cross-validation scores, and output the monitoring and identification results.

[0041] This invention synthesizes annual multi-band features from multi-temporal remote sensing images and constructs a random forest multi-classification model and a support vector machine binary classification model, enabling parallel output of multi-feature classification results and non-grain crop classification results. By rasterizing the vector boundaries of grain functional zones into masks, spatial constraints are applied to the results of the two models, ensuring that non-grain crop areas within non-grain functional zones are excluded from the non-grain crop classification confirmation output. Furthermore, this invention takes the union of the non-grain crop patches output by the two models within the mask area and calculates a cross-validation score based on the random forest confidence score, support vector machine confidence score, model consistency marker, and a penalty term for band similarity between candidate patches and the center vectors of rice, dryland crops, and orchard samples. This classifies candidate non-grain crop patches into confirmed patches, patches awaiting verification, and patches to be removed. This processing method reduces the risk of direct misjudgment caused by easily confused categories and provides a traceable data foundation for patch verification, area statistics, and priority ranking in grassroots supervision. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the process for monitoring and identifying the non-grain conversion of arable land according to the present invention.

[0043] Figure 2 This is a schematic diagram of the structure of a monitoring and identification system for non-grain conversion of arable land according to the present invention.

[0044] Figure 3 This is a schematic diagram illustrating the grain functional area mask constraint, dual-model parallel recognition, and cross-validation scoring diversion in this invention. Detailed Implementation

[0045] The embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrating the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the technical features in the following embodiments can be combined with each other. Based on the content described in this specification, those skilled in the art can make adaptive adjustments to the remote sensing data sources, band combinations, model parameters, scoring thresholds, statistical units, and output formats without any creative effort.

[0046] This invention provides a method and system for monitoring and identifying farmland non-grain conversion. The method takes multi-temporal remote sensing images, vector boundary data of grain functional zones, and sample annotation data as inputs to generate annual multi-band feature images; it constructs a random forest multi-classification model and a support vector machine binary classification model respectively; it then limits the output spatial range using a grain functional zone mask; finally, it performs union extraction and cross-validation scoring on the non-grain conversion patches output by the two models, classifying candidate non-grain conversion patches into confirmed patches, patches to be reviewed, and patches to be removed.

[0047] In this specification, a grain functional zone refers to a spatial area designated in farmland protection and grain production management to ensure grain production functions. Non-grain cropping refers to the planting of non-grain crops, seedlings and garden crops, perennial cash crops, abandonment of farmland, and aquaculture within a grain functional zone. A patch refers to a spatial object composed of one or more spatially continuous pixels with the same identification attribute. A candidate non-grain crop patch refers to a patch within the masked area of ​​a grain functional zone that is identified as non-grain crop or exhibits non-grain crop characteristics by a random forest multi-classification model or a support vector machine binary classification model. The confusion penalty term is a calculated term used to characterize the band similarity between candidate non-grain crop patches and the center vectors of grain crop and orchard samples.

[0048] I. Method Implementation

[0049] like Figure 1 As shown, the present invention provides a method for monitoring and identifying non-grain conversion of arable land based on grain functional zone masking and dual-model cross-validation, comprising the following steps.

[0050] S1. Acquire data and generate annual multi-band feature images.

[0051] Multi-temporal remote sensing images of the target area, vector boundary data of grain functional zones, and sample annotation data are acquired. Cloud cover filtering, cloud masking, image mosaicking, target area cropping, and annual average of non-zero pixels are performed on the multi-temporal remote sensing images to obtain annual multi-band feature images.

[0052] In one implementation, the multi-temporal remote sensing imagery utilizes publicly available Sentinel-2A / B satellite multispectral remote sensing imagery. Sentinel-2A / B remote sensing imagery comprises multispectral surface observation data acquired by the Sentinel-2A and Sentinel-2B satellites. This remote sensing imagery includes at least blue, green, red, and near-infrared bands and can be used for crop identification, land cover classification, and monitoring of farmland use changes. The target area can be a county, township, administrative village, or a continuous area containing a grain functional zone. The vector boundary data of the grain functional zone can be in shp, geojson, or gdb format to represent the spatial extent of the grain functional zone. Sample annotation data can be determined jointly using Tianditu multi-temporal imagery, field survey data, land use vector data, and manual interpretation results.

[0053] In other embodiments, multi-temporal remote sensing images may also be other publicly available or authorized multispectral remote sensing images, as long as they include at least the blue, green, red and near-infrared bands and can be spatially registered with the vector boundary data of grain functional zones.

[0054] Cloud cover filtering involves filtering based on the percentage of cloud cover in the remote sensing image metadata, removing images with cloud coverage exceeding a preset cloud cover threshold. The preset cloud cover threshold can be 20%, or adjusted to 10% to 30% depending on the available cloud quantity. Cloud masking involves identifying cloud pixels based on a blue light band threshold and setting them to null or zero values. For target areas spanning multiple image grids, geometric correction and image mosaicking are performed on images from the same or adjacent dates to ensure full coverage of the target area. Subsequently, the images are cropped according to the administrative or research boundaries of the target area, forming a multi-temporal image set for the target area.

[0055] Annual multi-band characteristic images are obtained by synthesizing the annual average values ​​of non-zero pixels. Specifically, for the same pixel location and the same band, valid observations from multiple temporal phases are read, and cloud pixels, invalid pixels, and zero-value pixels are removed before calculating the average value. For the p-th pixel and the b-th band, its annual average value is... satisfy:

[0056]

[0057] In the formula, Let be the annual average value of the p-th pixel in the b-th band; This represents the observation value of the p-th pixel at the t-th time phase and in the b-th band; The effective observation marker is set to 1 when the corresponding observation value is a non-zero effective value, and 0 otherwise; T is the number of time phases involved in the synthesis.

[0058] Through the above processing, effective observation information from multiple remote sensing images is synthesized into an annual multi-band feature image, which is used for subsequent sample feature extraction and model prediction. This step is mainly used to provide spatially consistent, band-consistent model input data with fewer missing values.

[0059] S2. Construct the first training sample set and the second training sample set.

[0060] Based on the sample annotation data, sample band features are extracted from annual multi-band feature images to construct a first training sample set for multi-land cover classification and a second training sample set for non-grain binary classification.

[0061] In one implementation, the sample labeling data includes food crop samples, non-food crop samples, and non-arable land background samples. Food crop samples include rice samples and dryland crop samples; non-food crop samples include abandoned farmland samples, pond aquaculture samples, perennial cash crop samples, and seedling and garden crop samples; non-arable land background samples include one or more of the following: orchard samples, water body samples, building samples, and forest land samples.

[0062] like Figure 1 and Figure 3 As shown, during sample band feature extraction, the coordinate system of the sample annotation points or sample annotation patches is transformed to the coordinate system of the annual multi-band feature image. The row and column numbers of the sample are determined based on the affine transformation parameters of the annual multi-band feature image. The blue, green, red, and near-infrared band values ​​of the effective pixels within that pixel or sample patch are then read to form the sample feature vector. :

[0063]

[0064] In the formula, Let be the feature vector of the k-th sample; This represents the blue light band value corresponding to the k-th sample. This represents the green band value corresponding to the k-th sample; This represents the red band value corresponding to the k-th sample; This represents the near-infrared band value corresponding to the k-th sample.

[0065] The first training sample set is labeled with multiple categories, including rice, dryland crops, non-grain crops, orchards, water bodies, buildings, and woodlands. This first training sample set is used to train a random forest multi-class classification model, enabling the model to output multiple land cover categories while preserving the differences between the non-grain category and easily confused categories such as rice, dryland crops, and orchards.

[0066] The second training sample set is obtained by reclassifying the first training sample set. Specifically, non-grain crops are labeled as those that have appeared, while rice, dryland crops, orchards, water bodies, buildings, and forest land are labeled as those that have not appeared. The second training sample set is used to train the support vector machine binary classification model, enabling the model to determine whether a pixel or patch belongs to a non-grain state.

[0067] Before training, sample equalization can be performed on the first and / or second training sample sets. Sample equalization can employ random undersampling, random oversampling, or the synthetic minority class oversampling technique (SMOTE). For cases where the number of samples in different classes differs significantly, sample equalization can reduce the impact of majority class samples on the model's discrimination boundary.

[0068] In this step, the first training sample set is used to retain the ability to distinguish multiple land cover categories, and the second training sample set is used to enhance the ability to detect non-grain binary classification. The two training sample sets correspond to two recognition models, providing an input basis for subsequent dual-model parallel recognition and patch-level cross-validation.

[0069] S3. Training the random forest multi-class classification model and the support vector machine binary classification model.

[0070] Train a random forest multi-class classification model and a support vector machine binary classification model respectively, and output the first classification result, the second classification result and the corresponding declassified confidence score.

[0071] 1. Random Forest Multi-Classification Model

[0072] The random forest multi-classification model takes the first training sample set as input, with input features being blue light band, green light band, red light band, and near-infrared band, and output categories being rice, dryland crops, non-grain crops, orchards, water bodies, buildings, and forest land.

[0073] In one implementation, the random forest multi-classification model includes multiple decision trees. Each decision tree is trained based on a subset of samples formed by sampling with replacement, and randomly selects some band features as candidate features when splitting nodes. When the model outputs, each decision tree votes on the category of the target pixel or patch, and the category with the most votes is taken as the first classification result.

[0074] For the i-th candidate, the random forest multi-classification model outputs the first confidence score that it belongs to the non-grain category. The first confidence level can be determined based on the ratio of the number of decision trees that voted for the non-food category to the total number of decision trees:

[0075]

[0076] In the formula, The confidence level of the random forest de-graining of the i-th object; Let be the number of decision trees in the random forest that classify the i-th object as a non-food category; The total number of decision trees in the random forest.

[0077] In one implementation, the number of decision trees in the random forest multi-classification model can be set to 150, the random seed can be set to 15, and the ratio of training set to test set can be set to 80% and 20%. The above parameters are only one feasible example, and those skilled in the art can adjust them according to the number of samples, the size of the region, and computing resources.

[0078] 2. Support Vector Machine Binary Classification Model

[0079] The support vector machine binary classification model takes the second training sample set as input, with input features being blue light band, green light band, red light band, and near-infrared band, and output categories being the category where non-grain-related features appear and the category where non-grain-related features do not appear.

[0080] In one implementation, the support vector machine (SVM) binary classification model employs a radial basis function kernel, with the kernel width using an adaptive scaling parameter, and the regularization parameter can be set to 1.0. The SVM model distinguishes between non-non ...

[0081] For the i-th object, the support vector machine binary classification model outputs the second confidence score that it belongs to the category of non-grained occurrence. The second confidence level can be determined based on the normalized results of the support vector machine classification margin, or by converting the classification margin into a probability value through probability calibration.

[0082] 3. Model Prediction and Block Processing

[0083] When the target area imagery is large, a block-based prediction approach can be used. The annual multi-band feature imagery is divided into several image blocks according to a preset block size. Band data is read block by block and converted into a feature matrix, which is then input into a random forest multi-classification model and a support vector machine binary classification model, respectively, to obtain the first and second classification results. The prediction results are written back according to the spatial location of the original image blocks, forming a classification raster spatially consistent with the annual multi-band feature imagery.

[0084] In this step, the random forest multi-class classification model and the support vector machine binary classification model are not mutually exclusive. The random forest multi-class classification model preserves the differences between multiple land cover categories, while the support vector machine binary classification model enhances the binary classification response for non-grained states. The outputs of both are retained and incorporated into the subsequent grain functional zone mask constraint and patch-level cross-validation processes.

[0085] S4. Generate a mask for the functional regions of grain and constrain the model output.

[0086] The vector boundary data of grain functional zones are rasterized into a grain functional zone mask that is consistent with the spatial characteristics of the annual multi-band feature image, and the grain functional zone mask is used to constrain the first and second classification results.

[0087] like Figure 3 As shown, the generation of the grain functional area mask includes the following process: converting the vector boundary data of the grain functional area to the coordinate system of the annual multi-band feature image; generating a binary mask according to the pixel size, row and column range and affine transformation parameters of the annual multi-band feature image; when the center point of a pixel falls within the vector boundary of the grain functional area, or when a portion of the pixel area exceeding a preset proportion falls within the vector boundary of the grain functional area, the pixel is marked as a grain functional area pixel; otherwise, it is marked as a non-grain functional area pixel.

[0088] Grain functional zone mask It can be represented as:

[0089]

[0090] In the formula, Let be the mask value of the grain functional area for the p-th pixel.

[0091] When using the grain functional zone mask to constrain the first and second classification results, pixels with a mask value of 0 are removed from the non-grain-related confirmation output, preventing them from participating in the generation of candidate non-grain-related patches. This process limits the spatial applicability of the monitoring and identification results, avoiding the output of non-grain crop areas within non-grain functional zones as non-grain-related patches within grain functional zones.

[0092] S5. Generate candidate non-grain-producing patches and calculate cross-validation scores.

[0093] Within the masked area of ​​the grain functional zone, the union of the non-grain-related patches in the first classification result and the non-grain-related patches in the second classification result is taken to generate candidate non-grain-related patches. Based on the confidence of random forest, the confidence of support vector machine, the model consistency label, and the band similarity penalty term between the candidate non-grain-related patches and the center vectors of grain crop samples and orchard samples, the cross-validation score is calculated.

[0094] 1. Generation of candidate non-grain-type patches

[0095] Within the masked area of ​​the grain functional zone, continuous pixels classified as non-grain-related in the first classification result are aggregated into a first set of non-grain-related patches, and continuous pixels classified as non-grain-related in the second classification result are aggregated into a second set of non-grain-related patches. Then, the union of the first and second sets of non-grain-related patches is taken to obtain a candidate set of non-grain-related patches.

[0096] If the first classification result identifies a region as non-grain land, but the second classification result does not identify it as non-grain land, the region will still be included in the candidate non-grain land set, and the model difference will be reflected in the subsequent scoring through model consistency markers. If the second classification result identifies a region as non-grain land, but the first classification result identifies it as an orchard, woodland, or other category, the region will also be included in the candidate non-grain land set, and the score will be adjusted through confounding penalty terms and model consistency markers.

[0097] 2. Summary of Feature Level at the Map Level

[0098] For each candidate de-grained patch, calculate the annual multi-band mean vector of all valid pixels within it. :

[0099]

[0100] In the formula, Let be the annual multi-band mean vector of the i-th candidate degrazing patch; The mean value of the blue light band of the effective pixels within the i-th candidate non-grained patch; The average value is for the green light band. This is the average value in the red light band; This is the average value in the near-infrared band.

[0101] Simultaneously, the rice sample center vector is calculated based on the training samples. Sample center vector of dryland crops and orchard sample center vector The sample center vector is obtained by averaging the feature vectors of the training samples of the corresponding class. For example, the sample center vector of rice. satisfy:

[0102]

[0103] In the formula, The center vector of the rice sample; This refers to the number of rice samples. Let be the feature vector of the k-th rice sample.

[0104] 3. Easily confused penalty items

[0105] Because rice, dryland crops, orchards, seedlings, and perennial cash crops may have similar vegetation spectral characteristics at certain time points, the category output of a single model may cause confusion. This invention includes a confusion penalty term. This is used to characterize the similarity between candidate non-grain patches and the center vectors of rice, dryland crops, and orchard samples:

[0106]

[0107] In the formula, The confusion penalty term for the i-th candidate non-grained patch; Let be the annual multi-band mean vector of the i-th candidate degrazing patch; The center vector of the rice sample; The center vector of the dryland crop sample; Let S be the center vector of the orchard sample; S(·) is the vector similarity function.

[0108] Cosine similarity can be used as a vector similarity function:

[0109]

[0110] In the formula, S(A,B) is the similarity between vectors A and B; A·B is the vector dot product; |A| and |B| are the vector norms, respectively.

[0111] When the band similarity between a candidate non-grain patch and the center vector of a rice sample, a dryland crop sample, or an orchard sample is high, the patch may belong to an easily confused region. Its cross-validation score is adjusted by the easily confused penalty item to make it more likely to enter the review or elimination process.

[0112] 4. Cross-validation scoring

[0113] Cross-validation score of candidate non-grain-based patches satisfy:

[0114]

[0115] In the formula, The cross-validation score for the i-th candidate non-grain-based patch; The random forest confidence score for the i-th candidate de-fed patch; The support vector machine confidence score for the i-th candidate de-grained patch; Mark the model as consistent; α is a penalty term for items that are easily confused; α, β, γ, and δ are weighting coefficients.

[0116] Model Consistency Marker It can be set as follows: when both the first classification result and the second classification result indicate that the i-th candidate non-grain-based patch has become non-grain-based, Take the first value; otherwise, take the second value. The first value can be 1, and the second value can be 0, or it can be set to a continuous value between 0 and 1 depending on the model consistency level.

[0117] The weighting coefficients can be determined through validation samples or set according to the recognition requirements of different regions. For example, in applications that reduce false negatives, β can be increased; in applications that reduce false positives, δ can be increased. This invention does not limit the specific value of the weighting coefficients, as long as they are used to combine the random forest confidence, support vector machine confidence, model consistency label, and confusion penalty term for scoring candidate de-furnished patches.

[0118] S6, Diversion Output Monitoring and Identification Results

[0119] Based on the cross-validation score, candidate non-grain-related patches are divided into confirmed patches, patches to be reviewed, and patches to be removed. The monitoring and identification results are output, including the patch boundary, patch area, the grain functional zone number to which it belongs, the model consistency marker, and the review priority.

[0120] In one implementation, when the cross-validation score is greater than or equal to a first scoring threshold, the corresponding candidate non-grain-based patch is classified as a confirmed patch; when the cross-validation score is less than the first scoring threshold but greater than or equal to a second scoring threshold, the corresponding candidate non-grain-based patch is classified as a patch to be reviewed; when the cross-validation score is less than the second scoring threshold, the corresponding candidate non-grain-based patch is classified as a rejected patch. The first scoring threshold is greater than the second scoring threshold.

[0121] The review priority of map patches can be determined based on cross-validation scores, patch area, the proportion of non-grain area within the corresponding grain functional zone, and model consistency markers. For patches with larger areas, located in concentrated areas of grain functional zones, and exhibiting inconsistent model outputs but high confidence levels regarding non-grain areas, the review priority can be increased. It can be determined in the following way:

[0122]

[0123] In the formula, Let i be the review priority of the i-th map patch to be reviewed; Scoring for cross-validation; This is the normalized value of the area of ​​the i-th patch; The percentage of non-grain area in the grain functional zone to which the i-th map patch belongs; Mark the model as consistent; , , , These are the weighting coefficients.

[0124] Monitoring and identification results can be output in the form of vector plots, raster layers, statistical tables, or review lists. Each output plot can be associated with one or more of the following information: plot number, plot boundary, plot area, grain functional zone number, random forest classification result, support vector machine classification result, model consistency marker, easily confused penalty item, cross-validation score, diversion type, and review priority. The above output information is used to support plot review, area statistics, and subsequent processing, without limiting specific field names or data table structures. Among them, plots to be reviewed can be summarized according to review priority, plot area, grain functional zone number, or administrative division information to form a review list or priority review list for subsequent manual interpretation, field verification, or regulatory task allocation.

[0125] Furthermore, non-grain-related statistical results can be generated according to grain functional zones, administrative villages, or townships. These results include the total area of ​​grain functional zones, the confirmed non-grain-related area, the non-grain-related area awaiting verification, and the proportion of non-grain-related area. For the m-th statistical unit, the proportion of non-grain-related area... satisfy:

[0126]

[0127] In the formula, The percentage of non-grain-producing area in the m-th statistical unit; The area of ​​the confirmed patch within the m-th statistical unit; Let m be the area of ​​the patch to be reviewed within the m-th statistical unit; Let m be the total area of ​​the grain functional zone within the m-th statistical unit.

[0128] II. System Implementation

[0129] like Figure 2 As shown, the present invention also provides a monitoring and identification system for the non-grain conversion of arable land, used to implement the above-mentioned method. The system includes a data acquisition module, an image preprocessing module, a sample construction module, a dual-model recognition module, a mask constraint module, a cross-validation module, and a result output module.

[0130] The data acquisition module is used to acquire multi-temporal remote sensing images of the target area, vector boundary data of grain functional zones, and sample annotation data. The data acquisition module can access data through remote sensing image data interfaces, vector data import interfaces, and field survey data import interfaces.

[0131] The image preprocessing module is used to perform cloud cover filtering, cloud masking, image mosaicking, target area cropping, and annual mean synthesis of non-zero pixels on multi-temporal remote sensing images to obtain annual multi-band feature images.

[0132] The sample construction module is used to extract sample band features from annual multi-band feature images and construct a first training sample set and a second training sample set. This module can also perform coordinate transformation, sample reclassification, and sample equalization.

[0133] The dual-model recognition module is used to train a random forest multi-classification model and a support vector machine binary classification model, and outputs the first classification result, the second classification result, and the corresponding denormalized confidence scores. This module may include a model training unit, a model storage unit, a block prediction unit, and a classification result concatenation unit.

[0134] The mask constraint module is used to rasterize the vector boundary data of grain functional areas into a grain functional area mask, and to constrain the first and second classification results using the grain functional area mask.

[0135] The cross-validation module is used to generate candidate de-grained patches and calculates the cross-validation score based on the random forest confidence, support vector machine confidence, model consistency label, and band similarity penalty term.

[0136] The results output module is used to classify confirmed patches, patches to be reviewed, and patches to be removed based on cross-validation scores, and outputs the monitoring and identification results. The results output module can also be used to generate non-grain statistical results and review lists according to grain functional area blocks, administrative villages, or townships.

[0137] The aforementioned systems can be deployed as software on general-purpose computers, servers, cloud platforms, or grassroots monitoring terminals, or they can be used in conjunction with geographic information system platforms, remote sensing image processing platforms, or agricultural monitoring platforms.

[0138] III. Examples

[0139] The following embodiments use a grain functional zone in a target area of ​​a county as an example to illustrate the method of the present invention. The data, parameters, and test results used in this embodiment are used to illustrate the possible implementations of the present invention and are not intended to limit the scope of protection of the present invention.

[0140] 1. Example 1: Non-grain monitoring and identification based on annual multi-band feature images

[0141] 1.1 Data Acquisition

[0142] This embodiment selects a target area within a county as the study area. This target area contains various land cover types, including grain functional zones, general cultivated land, forest land, orchards, water bodies, and construction land. The multi-temporal remote sensing imagery used is Sentinel-2A / B, covering the year 2021. Images were selected based on a cloud cover rate of less than 20%, and 24 usable image periods were obtained through cloud masking, mosaicking, and cropping. The vector boundary data for grain functional zones uses the boundaries of grain functional zone blocks provided by the local management department. Sample annotation data was determined jointly using multi-temporal imagery from Tianditu (a Chinese online map platform), land use vector data, and field survey records. The main data are shown in Table 1.

[0143] Table 1 shows the main data used in Example 1.

[0144]

[0145] 1.2 Sample Setup

[0146] A total of 6080 samples were collected in this embodiment. The first training sample set was labeled according to 7 categories: rice, dryland crops, non-grain crops, orchards, water bodies, buildings, and forest land. The number of samples in each category is shown in Table 2.

[0147] Table 2. Categories and Quantities of the First Training Sample Set

[0148]

[0149] The second training sample set was obtained by reclassifying the first training sample set. The non-grain-related samples represent categories where non-grain-related items appeared, while the remaining categories represent categories where non-grain-related items did not appear. The number of samples after reclassification is shown in Table 3.

[0150] Table 3. Categories and Quantities of the Second Training Sample Set

[0151]

[0152] 1.3 Band Feature Statistics

[0153] The blue, green, red, and near-infrared band values ​​of corresponding pixels in the annual multi-band feature image were read to form the sample feature vector. The mean values ​​of the sample bands for each category are shown in Table 4, where Band 1 is the blue band, Band 2 is the green band, Band 3 is the red band, and Band 4 is the near-infrared band.

[0154] Table 4. Band mean values ​​of each category of samples in the first training sample set.

[0155]

[0156] As shown in Table 4, rice, non-grain crops, and orchards exhibit similarities across multiple bands. Therefore, this embodiment includes a confusion penalty in the cross-validation scoring to reduce the probability that candidate patches with high similarity to the center vectors of rice, dryland crops, or orchard samples are directly identified as non-grain crops.

[0157] 1.4 Model Structure and Training Parameters

[0158] The structured configurations of the random forest multi-class classification model and the support vector machine binary classification model are shown in Table 5.

[0159] Table 5 Model Structure and Training Parameters

[0160]

[0161] During the training phase, sample equalization is performed on the second training sample set to mitigate the difference in sample numbers between categories without and with non-grain categories. After model training is complete, the model file is saved, and block predictions are performed on the annual multi-band feature image of the target area. The block size can be set to 2000×2000 pixels; for areas with insufficient edges to form complete blocks, data is read according to the actual remaining row and column range, and the prediction results are written back.

[0162] 1.4 Single-model test results

[0163] The test results of the random forest multi-class classification model are shown in Table 6.

[0164] Table 6. Test results of the random forest multi-classification model

[0165]

[0166] The test results of the support vector machine binary classification model are shown in Table 7.

[0167] Table 7 Test results of the Support Vector Machine binary classification model

[0168]

[0169] The test results above illustrate that the two models have different output features. The random forest multi-classification model can preserve the classification relationship between non-grained land cover and multiple land cover types; the support vector machine binary classification model can provide binary classification judgments for non-grained land cover categories. The outputs of both models are then used in the subsequent cross-validation scoring process.

[0170] 1.6 Results of Mask Constraint for Grain Functional Zones

[0171] A grain functional zone mask constraint is applied to the global prediction results of the single model. Without the grain functional zone mask, the model may output areas of legally existing cash crops, orchards, or other non-grain crops within general cultivated land as non-grain areas. After applying the grain functional zone mask, only candidate non-grain areas within the grain functional zones are retained.

[0172] In this embodiment, after cropping using a grain functional zone mask, the random forest multi-classification model predicts that the non-grain area within the grain functional zone is approximately 557.70 hectares, accounting for 7.99% of the total grain functional zone area. The support vector machine binary classification model, after cropping using a grain functional zone mask, predicts a larger non-grain area than the random forest model. These results indicate that different models differ in their non-grain detection range, making further patch-level cross-validation suitable.

[0173] 1.7 Cross-validation scoring and triage results

[0174] In this embodiment, α=0.35, β=0.35, γ=0.20, and δ=0.10 are set, and the model consistency flag is used. A score of 1 is assigned when both models indicate degrainization, and 0 otherwise. The first scoring threshold is set to 0.65, and the second scoring threshold is set to 0.45. For each candidate degrainized patch, the following calculation is performed: And divert according to the rules in Table 8.

[0175] Table 8. Diversion Rules for Candidate Non-Grain-Growing Plots

[0176]

[0177] To illustrate the cross-validation scoring process, Table 9 lists some scoring examples for candidate patches.

[0178] Table 9 Scoring of Candidate Non-Grain Plots

[0179]

[0180] As shown in Table 9, when both models give high confidence scores for non-grain crops and the candidate patches have low similarity to the center vectors of grain crops or orchard samples, the cross-validation score is high, and the corresponding patches are classified as confirmed patches. When the model results are inconsistent, or when the candidate patches have high similarity to the bands of rice, dryland crops, or orchard categories, the cross-validation score decreases, and the corresponding patches enter the process of pending review or elimination.

[0181] 1.8 Comparison of different processing methods

[0182] To illustrate the processing effect of this invention, three processing methods are compared. Method A uses only a random forest multi-classification model to output non-grain-specific patches; Method B uses only a support vector machine binary classification model to output non-grain-specific patches; Method C is the grain functional area masking and dual-model cross-validation processing method of this invention.

[0183] Table 10 lists the comparison results of the output features of the three methods.

[0184] Table 10 Comparison of outputs from different processing methods

[0185]

[0186] Table 11 lists examples of results from the three methods on the verification samples. The verification samples are formed by a combination of manual interpretation and field sampling, and are used to determine whether the output map features conform to the non-grain status within the grain functional area.

[0187] Table 11 Example of Review Sample Results

[0188]

[0189] Method C in Table 11 transfers some intermediate-state patches to a pending review list, instead of directly outputting them as confirmed non-grain crops, and places them in the pending review or removal process. Therefore, among the directly output confirmed patches, the proportion of those not confirmed during review is lower than with the single-model method. This result indicates that using model consistency markers and easily confused penalty terms for patch-level cross-validation can reduce the number of patches similar to easily confused categories such as grain crops and orchards being directly output as confirmed patches.

[0190] 1.9 Statistical Output at the Administrative Village Scale

[0191] After the map features are distributed, the confirmed map features and the map features to be verified are spatially overlaid with the boundaries of administrative villages and grain functional zones, and the proportion of non-grain areas within each administrative village or grain functional zone is calculated. Table 12 shows the output of some statistical units.

[0192] Table 12 Example of Non-Grain Area Percentage in Statistical Units

[0193]

[0194] The risk level can be set into multiple tiers based on the proportion of non-grain acreage, or it can be set into low-risk, medium-risk, and high-risk levels according to local regulatory needs. For statistical units with a large number of plots to be reviewed or a high proportion of non-grain acreage, the system can generate a priority review list.

[0195] In this embodiment of the invention, annual multi-band feature images are generated through cloud cover filtering, cloud masking, and synthesis of annual average values ​​of non-zero pixels, which can reduce the impact of clouds and invalid pixels on the model input. The first and second training sample sets serve multi-feature classification and non-grain-based binary classification, respectively, giving the two model outputs different discriminative emphases. The grain functional zone mask is used to limit the output spatial range, ensuring that non-grain crop areas within non-grain functional zones do not participate in the non-grain-based confirmation output. Candidate non-grain-based patches are generated by taking the union of the results from the two models, which helps to retain suspicious patches; the cross-validation scoring incorporates random forest confidence, support vector machine confidence, model consistency markers, and easily confused penalty terms, which helps to distinguish between confirmed patches, patches to be reviewed, and patches to be removed; the review priority output can provide a data basis for subsequent manual interpretation, field verification, or regulatory task allocation.

[0196] The above embodiments illustrate that the present invention can form a processing flow within a grain functional area, from remote sensing image input, sample construction, dual-model recognition, mask constraint, candidate patch generation, cross-validation scoring to result output. Those skilled in the art can implement the technical solution of the present invention according to the data structure, model training method, scoring formula, and diversion rules disclosed in this specification.

[0197] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention should not be construed as limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

Claims

1. A method for monitoring and identifying the conversion of arable land to non-grain crops, characterized in that, Includes the following steps: S1. Acquire multi-temporal remote sensing images of the target area, vector boundary data of grain functional zones and sample annotation data, and perform cloud cover filtering, cloud masking, image mosaicking, target area cropping and non-zero pixel annual mean synthesis on the multi-temporal remote sensing images to obtain annual multi-band feature images. S2. Based on the sample annotation data, extract sample band features from the annual multi-band feature image to construct a first training sample set for multi-land cover classification and a second training sample set for non-grain binary classification. S3. Train the random forest multi-classification model and the support vector machine binary classification model respectively, and output the first classification result, the second classification result and the corresponding declassified confidence score. S4. Rasterize the vector boundary data of the grain functional area into a grain functional area mask that is consistent with the space of the annual multi-band feature image, and use the grain functional area mask to constrain the first classification result and the second classification result. S5. Within the masked area of ​​the grain functional zone, take the union of the non-grain patches in the first classification result and the non-grain patches in the second classification result to generate candidate non-grain patches. Calculate the cross-validation score based on the random forest confidence, support vector machine confidence, model consistency marker, and the band similarity penalty term between the candidate non-grain patches and the center vectors of grain crop samples and orchard samples. S6. Based on the cross-validation score, the candidate non-grain-related patches are divided into confirmed patches, patches to be reviewed, and patches to be removed, and the monitoring and identification results including patch boundaries, patch area, the number of the grain functional area to which it belongs, model consistency marker, and review priority are output.

2. The method for monitoring and identifying the conversion of arable land to non-grain crops according to claim 1, characterized in that, The multi-temporal remote sensing images are publicly acquired or authorized multispectral remote sensing images, and include at least the blue light band, green light band, red light band, and near-infrared band; the sample annotation data are jointly determined by Tianditu multi-temporal images, field survey data, and land use vector data; the sample annotation data include food crop samples, non-food samples, and non-arable land background samples; the food crop samples include rice samples and dryland crop samples; the non-food samples include abandoned farmland samples, pond aquaculture samples, perennial cash crop samples, and seedling and garden crop samples; the non-arable land background samples include one or more of the following: orchard samples, water body samples, building samples, and forest land samples.

3. The method for monitoring and identifying the conversion of arable land to non-grain crops according to claim 1, characterized in that, In step S1, cloud cover filtering includes removing images with cloud coverage rates higher than a preset cloud cover threshold based on the cloud cover percentage in the remote sensing image metadata; cloud masking includes identifying cloud pixels based on a blue light band threshold and setting cloud pixels to null values; non-zero pixel annual mean synthesis includes averaging the non-zero valid observations of the same pixel location in multiple time phases, so that cloud pixels, invalid pixels, and zero-value pixels do not participate in the annual mean calculation.

4. The method for monitoring and identifying the conversion of arable land to non-grain crops according to claim 1, characterized in that, In step S2, extracting sample band features includes reading the blue light band value, green light band value, red light band value, and near-infrared band value of the corresponding pixels of the sample annotation points or sample annotation patches, forming a sample feature vector. The sample feature vector satisfy: ； In the formula, For the first The sample feature vector of each sample; Values ​​for the blue light band; Values ​​for the green light band; Values ​​for the red light band; This is the value in the near-infrared band.

5. The method for monitoring and identifying the conversion of arable land to non-grain crops according to claim 1, characterized in that, In step S2, the first training sample set is labeled with multiple categories: rice, dryland crops, non-grain crops, orchards, water bodies, buildings, and forest land. The second training sample set is obtained by reclassifying the first training sample set, wherein the non-grain category is labeled as having a non-grain category, and rice, dryland crops, orchards, water bodies, buildings, and forest land are labeled as not having a non-grain category. Before training the random forest multi-classification model and the support vector machine binary classification model, sample balancing is performed on the first training sample set and / or the second training sample set.

6. The method for monitoring and identifying the conversion of arable land to non-grain crops according to claim 1, characterized in that, In step S3, the random forest multi-class classification model outputs the first confidence score of each pixel or patch belonging to the non-grain category. The support vector machine binary classification model outputs the second confidence score for each pixel or patch belonging to the non-grain-differentiated category. The first confidence level The second confidence level is determined based on the number of decision trees in the random forest that vote for the non-food category. Determined based on the classification interval or probability calibration results of the support vector machine.

7. The method for monitoring and identifying the conversion of arable land to non-grain crops according to claim 1, characterized in that, In step S4, rasterizing the vector boundary data of the grain functional area into a grain functional area mask includes: Transform the vector boundary data of grain functional zones into the coordinate system of annual multi-band feature images; A binary mask is generated based on the pixel size, row and column range, and affine transformation parameters of the annual multi-band feature image. When the center point of a pixel falls within the vector boundary of the grain functional area, or when a portion of the pixel area exceeding a preset proportion falls within the vector boundary of the grain functional area, the pixel is marked as a grain functional area pixel. A binary mask is used to remove non-grain functional area pixels, so that non-grain functional area pixels do not participate in the non-grain conversion confirmation output.

8. The method for monitoring and identifying the conversion of arable land to non-grain crops according to claim 1, characterized in that, In step S5, the cross-validation score of candidate non-grain-based patches is calculated. satisfy: ； In the formula, For the first Cross-validation scores of candidate non-grain-based patches; For the first Random forest confidence of candidate non-grained patches; For the first Support vector machine confidence of each candidate de-grained patch; A model consistency marker is used when both the first and second classification results indicate that the model is consistent. When a candidate non-grain-type patch appears non-grain-type... Take the first value, otherwise take the second value; For easily confused penalty items; , , , These are the weighting coefficients.

9. The method for monitoring and identifying the conversion of arable land to non-grain crops according to claim 8, characterized in that, The easily confused penalty item The similarity between the annual multi-band mean vector of candidate non-grain patches and the center vectors of rice, dryland crops, and orchards is determined, and the following conditions must be met: ； In the formula, For the first The easily confused penalty item for each candidate non-grain-based patch; For the first The annual multi-band mean vector of each candidate non-grained patch; The center vector of the rice sample; The center vector of the dryland crop sample; The center vector of the orchard sample; This is a vector similarity function; In step S6, when the cross-validation score is greater than or equal to the first scoring threshold, the corresponding candidate non-grain-based plot is classified as a confirmed plot; when the cross-validation score is less than the first scoring threshold but greater than or equal to the second scoring threshold, the corresponding candidate non-grain-based plot is classified as a plot to be reviewed; when the cross-validation score is less than the second scoring threshold, the corresponding candidate non-grain-based plot is classified as a plot to be removed; the review priority of the plot to be reviewed is determined based on the cross-validation score, the plot area, the proportion of non-grain-based area within the grain functional zone, and the model consistency marker.

10. A monitoring and identification system for the conversion of arable land to non-grain crops, characterized in that, The system for implementing the method for monitoring and identifying non-grain conversion of arable land according to any one of claims 1 to 9, the system comprising: The data acquisition module is used to acquire multi-temporal remote sensing images of the target area, vector boundary data of grain functional zones, and sample annotation data; The image preprocessing module is used to perform cloud cover filtering, cloud masking, image mosaicking, target area cropping, and annual mean synthesis of non-zero pixels on multi-temporal remote sensing images to obtain annual multi-band feature images. The sample construction module is used to extract sample band features from annual multi-band feature images and construct the first training sample set and the second training sample set. The dual-model recognition module is used to train a random forest multi-classification model and a support vector machine binary classification model, and outputs the first classification result, the second classification result, and the corresponding declassified confidence score. The mask constraint module is used to rasterize the vector boundary data of grain functional areas into a grain functional area mask, and to constrain the first classification result and the second classification result using the grain functional area mask; The cross-validation module is used to generate candidate de-grained patches and calculate cross-validation scores based on random forest confidence, support vector machine confidence, model consistency markers, and band similarity penalty terms. The results output module is used to classify confirmed patches, patches to be reviewed, and patches to be removed based on cross-validation scores, and output the monitoring and identification results.

Images