An intelligent method for extracting cultivated land outflow patches

Abstract

This invention discloses an intelligent method for extracting cultivated land outflow patches, including: data collection and preprocessing; rapid sample generation based on existing data and manual collection, and storage of sample metadata and file data in database and file formats, respectively; adaptive modification and combination of appropriate semantic segmentation and change detection algorithms, introduction of fully trained general segmentation and general change detection pre-trained models, and training of iterative optimization networks to obtain optimally accurate non-cultivated land identification and cultivated land change detection models; extraction of suspected cultivated land outflow patches based on the optimally accurate models; batch regularization of suspected cultivated land outflow patches to obtain preliminary suspected cultivated land outflow patches; and verification of the correctness of the boundaries and attributes of the preliminary patches to obtain the final suspected cultivated land outflow patches. This invention can monitor cultivated land outflow in mountainous areas and quickly, efficiently, and accurately extract and update change patches and type information.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image interpretation, and in particular to an intelligent method for extracting cultivated land outflow patches. Background Art

[0002] Arable land is a crucial foundation for food production. In recent years, to ensure agricultural production and food security, China has introduced a series of policies and measures to strictly protect arable land. These policies include strengthening arable land protection and use control, adhering to red lines for arable land protection, and strengthening dynamic monitoring. Leveraging satellite imagery monitoring technology, closely aligned with on-site verification efforts, ensures that every step is based on solid facts and current policies and regulations, enabling fair and objective assessments of the legality of arable land changes, thereby enabling precise and effective implementation of arable land protection. In mountainous provinces, where arable land is fragmented and slopes are steep, arable land monitoring and management primarily relies on purely manual visual interpretation of remote sensing imagery. This is a time-consuming and labor-intensive task for large-scale natural resource monitoring and supervision. With the advancement of artificial intelligence (AI), deep learning-based intelligent interpretation of remote sensing imagery has demonstrated significant improvements in accuracy and efficiency in remote sensing image recognition. Semantic segmentation models now have the ability to extract essential elements such as arable land, forest land, water bodies, buildings, and roads, and change detection models are capable of monitoring changes in typical land features. This provides a viable technical approach for efficient natural resource management and supervision. For example: China Patent Publication No. CN 113989652B disclosed a “method and system for detecting cultivated land changes under layered multiple judgment rules” on April 26, 2022. It uses remote sensing images to extract change mask images by fusing mask images of villages, non-cultivated land and abnormal areas, and performs raster vectorization processing, combined with slow feature analysis to detect changes in cultivated land, which mainly solves the problems of high missed detection rate, low recall rate and difficulty in detecting small targets. China Patent Publication No. CN111582043B disclosed a “method for detecting changes in high-resolution remote sensing images based on multi-task learning” on March 15, 2022. By constructing a semantic segmentation model and a change detection model, combined with a multi-task learning mechanism, and processing semantic segmentation and change detection tasks at the same time, it avoids the problem of error accumulation and improves the accuracy of change detection. In particular, by sharing the feature extraction module, parameter redundancy is reduced and operational efficiency is improved. However, the following problems still exist:

[0003] 1. Accuracy and applicability issues in monitoring farmland changes in complex terrain. While CN113989652B improves small target detection capabilities through layered, multi-determination rules, its approach does not explicitly address highly heterogeneous areas such as mountainous terrain and fragmented farmland. CN111582043B: While multi-task learning improves accuracy, it does not address adaptability to complex terrain.

[0004] 2. High cost and low efficiency of sample data preparation. CN113989652B and CN111582043B: Both rely on a large amount of labeled data but do not propose efficient sample preparation methods.

[0005] 3. Error accumulation and insufficient model generalization. CN111582043B avoids error accumulation through multi-task learning, but does not address the model's generalization capabilities for business scenarios.

[0006] In summary, although remote sensing deep learning technology has made great progress in theoretical research and model construction, the extraction of surface cover information is not effective in mountainous areas with complex terrain, diverse planting methods and fragmented cultivated land characteristics, as well as high-intensity spatial heterogeneity. Summary of the Invention

[0007] The purpose of the present invention is to overcome the above problems and provide an intelligent farmland outflow patch extraction method that can accurately monitor the outflow of farmland in mountainous areas and quickly, efficiently and accurately extract and update the change patches and type information.

[0008] To achieve the above objectives, the present invention adopts the following technical solutions.

[0009] The present invention provides an intelligent method for extracting cultivated land outflow patches, comprising the following steps:

[0010] S1 Data Collection and Preprocessing: Based on business needs, remote sensing image data, land change survey data, geographic national conditions monitoring data, and suspected cultivated land outflow monitoring maps and images issued by the state are collected; the collected data and images are preprocessed: the geographic coordinate system of vectors and images is unified to the CGCS2000 national geodetic coordinate system, the image data format is unified to tif, the remote sensing images are color-uniformed, and the image and vector data are framed and renamed according to the "National Basic Scale Topographic Map Framing and Numbering" by selecting the standard framing frame;

[0011] S2 sample dataset production and storage: The interpretation targets are determined based on image resolution, spectral characteristics, and the results of farmland protection business scenario analysis. A classification system for interpretation targets is established by referring to the classification system of land change survey data and geographical national conditions monitoring data. The sample dataset is produced through two complementary methods: one is to extract samples based on existing data: the land change survey database and geographical national conditions monitoring database are sorted out in combination with supporting images to directly extract vectors and images that meet the interpretation target samples; the other is based on manual image collection: for the parts missed or insufficient in the first method, supplementary measures such as manual collection and fine annotation are adopted;

[0012] Convert the labeled interpretation target vector sample data into the corresponding rasterized label image, so that the raster image has the same resolution and size as the corresponding image; crop the image and the corresponding raster label to obtain the slice sample data and store it in the database;

[0013] The main content of the remote sensing interpretation sample library consists of three parts: images, vectors, and samples. Images include image metadata and image files, vectors include vector metadata and vector files, and samples include sample dataset metadata, sample metadata, and sample dataset files after slicing. Metadata is stored in an SQLite database, while image files, vector files, and sample datasets are stored in files.

[0014] S3 model construction and training: Flexible selection and assembly of required sample data from the sample library. Sample augmentation of the slice sample dataset extracted from the sample library is performed using rotation, scaling, flipping, and color transformation methods. The slice sample dataset is then proportionally divided into training, validation, and test sets. Based on the task objectives, appropriate semantic segmentation and change detection algorithms are selected for adaptive transformation and combination. Before model training begins, fully trained general segmentation and general change detection pre-trained models are introduced. Finally, network hyperparameters are set, the network is iteratively optimized through training, and model accuracy is evaluated in the test set. The non-cultivated land recognition model and cultivated land change detection model with the optimal accuracy are saved.

[0015] S4 Extraction of suspected cultivated land outflow patches: Based on the non-cultivated land identification model with optimal accuracy, the latest image is predicted and inferred, and non-cultivated land identification patches are automatically extracted. The non-cultivated land patches are then overlaid with the cultivated land boundary line data in the land change survey data of the previous year, and the patches that have changed within the cultivated land scope based on the land change survey database of the previous year are obtained as "newly discovered" suspected cultivated land outflow patches; Based on the cultivated land change detection model with optimal accuracy, two images of the same area at different times are directly predicted and inferred, and the "newly occurring" suspected cultivated land outflow patches are automatically compared and extracted from the two images; The "newly discovered" suspected cultivated land outflow patches obtained based on the non-cultivated land identification model and the "newly occurring" suspected cultivated land outflow patches obtained based on the cultivated land change detection model are fused and complemented to obtain the suspected cultivated land outflow patches of the target area;

[0016] S5 data post-processing: Based on the extraction type requirements, batch graphic regularization processing is performed on the suspected cultivated land outflow patches in the target area, including graphic topology optimization, square processing, small patch removal, and smoothing processing, to obtain preliminary results of suspected cultivated land outflow patches;

[0017] S6 Manual Verification: Based on the suspected cultivated land outflow monitoring spots and images issued by the state, internal verification is conducted on the correctness of the boundaries and attributes of the preliminary results of suspected cultivated land outflow in the target area; for spots that are still in doubt, further field verification is carried out after marking to obtain the suspected cultivated land outflow results spots after manual verification.

[0018] In the above-mentioned intelligent method for extracting cultivated land outflow patches, the standard framing frame in step S1 is 1:10,000, 1:25,000 or 1:50,000.

[0019] In the above-mentioned intelligent cultivated land outflow patch extraction method, the slice sample data set described in step S3 is divided into a ratio of 6:2:2, 7:2:1 or 8:1:1.

[0020] The above-mentioned intelligent method for extracting cultivated land outflow patches, wherein in step S5 data post-processing, when there is overlap between patches, the boundary relationship is adjusted according to the geographic spatial position, and the overlapping part is deleted to perform graphic topology optimization; when the automatically extracted house boundary is relatively smooth, a rectangular processing is performed on the basis of the automatic extraction; small patches are eliminated by setting the minimum patch area and eliminating patches that do not meet the requirements; smoothing processing is mainly performed on linear features.

[0021] Compared with the existing technology, the present invention has obvious beneficial effects. It can be seen from the above technical solutions: First, it integrates multi-temporal remote sensing image data, land change survey data, geographical national conditions monitoring data, and suspected cultivated land outflow monitoring maps and images issued by the state, and adopts a sample collection method that combines stock data extraction and manual collection. It not only expands the sample source and shortens the sample preparation time, but also meets the migration application of different regions and different application scenarios, and enhances the generalization and practicality of the model in cultivated land protection tasks under complex terrain; second, it uses file format and SQLite database to store and manage sample file data and sample metadata respectively, which facilitates the flexible selection of sample data , assembled on demand; third, a fully trained general segmentation and general change detection pre-trained model is introduced to improve the convergence speed of the model during the training process, and enhance the generalization ability and accuracy of the model; fourth, through a variety of data post-processing (eliminating small spots, topology optimization, etc.) and manual verification (in-house verification and field verification) methods, the accuracy and objectivity of cultivated land outflow monitoring in mountainous areas are further improved to better meet the national requirements for cultivated land change monitoring; fifth, the present invention combines the deep learning method of semantic segmentation and change detection tasks, and fuses and complements the "newly discovered" suspected cultivated land outflow patches with the "newly occurring" suspected cultivated land outflow patches, which can effectively reduce the omission of cultivated land outflow. The present invention can realize the rapid, efficient and accurate extraction and updating of cultivated land outflow patches, and the obtained suspected cultivated land outflow patches can be superimposed with administrative management data for engineering application. Suspected farmland outflow maps can be incorporated into farmland protection scenarios and integrated with natural resource management data (such as the Three Zones and Three Lines, agricultural land conversion approval data, nature reserves, and historical farmland data) for analysis. This reveals suspected illegal occupation of farmland for building structures, road construction, lake construction, and landscaping, illegal land use, and excessive afforestation. Before the competent authorities issue suspected illegal and irregular maps, intelligent processes can advance the timeline for identifying farmland outflows, integrating phased and ad hoc work into daily routines and easing the pressure of investigations. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the technical roadmap of the present invention;

[0023] Figure 2 This is a diagram of the data storage structure of the remote sensing interpretation sample library of the present invention.

[0024] The concept, specific technical methods and technical effects of the present invention will be further described below with reference to the accompanying drawings. DETAILED DESCRIPTION

[0025] Example 1:

[0026] like Figure 1As shown, the present invention provides an intelligent method for extracting cultivated land outflow patches, comprising the following steps:

[0027] S1 Data collection and preprocessing: Based on business needs (taking the conversion of cultivated land into buildings as an example), remote sensing image data, land change survey data, and geographic national conditions monitoring data are collected as data sources for non-cultivated land identification samples and cultivated land change detection samples; national suspected cultivated land outflow monitoring spots and images are collected as reference standards to verify the correctness of the boundaries and attributes of the automatically extracted preliminary suspected cultivated land outflow results spots; the collected data and images are preprocessed: the geographic coordinate system of vectors and images is unified to the CGCS2000 national geodetic coordinate system, and the data format of images is unified to tif, the remote sensing images are color-uniformed, and then the images and vector data are framed and cropped and renamed according to the 1:10,000 standard frame;

[0028] S2 sample dataset production and storage: The interpretation target is determined based on image resolution, spectral characteristics, and the results of farmland protection business scenario analysis. A classification system for interpretation targets is established by referring to the classification system of land change survey data and geographical national conditions monitoring data. The sample dataset is produced through two complementary methods: the first is to extract samples based on existing data: the land change survey database and geographical national conditions monitoring database are sorted out in combination with supporting images to directly extract vectors and images that meet the interpretation target samples, realizing the effective utilization and activation of existing data stocks; the second is based on manual image collection: for parts that may be missed or insufficient in the first method, supplementary measures such as manual collection and fine annotation are adopted;

[0029] Convert the labeled interpretation target vector sample data into the corresponding rasterized label image, so that the raster image has the same resolution and size as the corresponding image; in order to meet the requirements of the deep learning model for the input image size, the image and the corresponding raster label are cropped to obtain the sliced ​​sample data and store it in the database;

[0030] Remote sensing interpretation sample library data storage structure (such as Figure 2 As shown in the figure, the main content of the sample library consists of three parts: images, vectors, and samples. Images include image metadata and image files, vectors include vector metadata and vector files, and samples include sample dataset metadata, sample metadata, and sample dataset files after slicing. The metadata is stored in an SQLite database, while image files, vector files, and sample datasets are stored in files.

[0031] S3 model construction and training: According to the requirements of the interpretation target classification system, sample data of buildings and farmland-to-buildings are selected from the sample library, and the slice sample data set extracted from the sample library is augmented by rotation, scaling, flipping and color transformation methods; the slice sample data set is then divided into training set, validation set and test set in a ratio of 6:2:2 for subsequent building recognition model and farmland-to-building detection model training; according to the task objectives, appropriate semantic segmentation and change detection algorithms are selected for adaptive transformation and combination. Before the model starts training, introducing a fully trained general segmentation and general change detection pre-training model is an efficient method, which can greatly promote the convergence speed of the model during training and also help improve the generalization ability and accuracy of the model; finally, the network hyperparameters are set, the network is trained and iteratively optimized, the model accuracy is evaluated in the test set, and the building recognition model and farmland-to-building detection model with the optimal accuracy are saved;

[0032] S4 Extraction of suspected cultivated land outflow patches: Based on the building recognition model with optimal accuracy, the latest image is predicted and inferred, and building patches are automatically extracted. The building patches are then overlaid with the cultivated land boundary line data in the land change survey data of the previous year to obtain patches where cultivated land has been converted into buildings based on the land change survey database of the previous year, namely, "newly discovered" suspected cultivated land outflow patches to buildings; Based on the cultivated land-to-building detection model with optimal accuracy, two images of the same area at different times are directly predicted and inferred to obtain patches automatically compared and extracted from the two images, namely, "newly occurring" suspected cultivated land outflow patches to buildings; The "newly discovered" suspected cultivated land outflow patches to buildings obtained based on the building recognition model and the "newly occurring" suspected cultivated land outflow patches to buildings obtained based on the cultivated land-to-building detection model are fused and complemented to obtain the suspected cultivated land outflow patches to buildings in the target area, thus achieving efficient data integration and significant quality improvement;

[0033] S5 data post-processing: set the minimum patch area, remove patches that do not meet the requirements, and perform rectangular processing on the patches of suspected farmland flowing to buildings in the target area to obtain preliminary patches of suspected farmland flowing to buildings;

[0034] S6 Manual Verification: Based on the suspected cultivated land outflow monitoring spots and images issued by the state, an internal verification is conducted on the correctness of the boundaries and attributes of the preliminary results of the suspected cultivated land flowing to buildings in the target area; for the spots that are still in doubt, further field verification is carried out after marking to obtain the suspected cultivated land flowing to buildings results after manual verification.

[0035] Example 2:

[0036] like Figure 1 As shown, the present invention provides an intelligent method for extracting cultivated land outflow patches, comprising the following steps:

[0037] S1 Data collection and preprocessing: Based on business needs (taking the conversion of cultivated land to forest land as an example), remote sensing image data, land change survey data, and geographic national conditions monitoring data are collected as data sources for non-cultivated land identification samples and cultivated land change detection samples; national suspected cultivated land outflow monitoring patches and images are collected as reference standards to verify the correctness of the boundaries and attributes of the automatically extracted preliminary suspected cultivated land outflow patches; the collected data and images are preprocessed: the geographic coordinate system of vectors and images is unified to the CGCS2000 national geodetic coordinate system, the data format of images is unified to tif, the remote sensing images are color-uniformed, and then the images and vector data are framed and renamed according to the 1:25,000 standard frame;

[0038] S2 sample dataset production and storage: The interpretation target is determined based on image resolution, spectral characteristics, and the results of farmland protection business scenario analysis. A classification system for interpretation targets is established by referring to the classification system of land change survey data and geographical national conditions monitoring data. The sample dataset is produced through two complementary methods: the first is to extract samples based on existing data: the land change survey database and geographical national conditions monitoring database are sorted out in combination with supporting images to directly extract vectors and images that meet the interpretation target samples, realizing the effective utilization and activation of existing data stocks; the second is based on manual image collection: for parts that may be missed or insufficient in the first method, supplementary measures such as manual collection and fine annotation are adopted;

[0039] Convert the labeled interpretation target vector sample data into the corresponding rasterized label image, so that the raster image has the same resolution and size as the corresponding image; in order to meet the requirements of the deep learning model for the input image size, the image and the corresponding raster label are cropped to obtain the sliced ​​sample data and store it in the database;

[0040] Remote sensing interpretation sample library data storage structure (such as Figure 2 As shown in the figure, the main content of the sample library consists of three parts: images, vectors, and samples. Images include image metadata and image files, vectors include vector metadata and vector files, and samples include sample dataset metadata, sample metadata, and sample dataset files after slicing. The metadata is stored in an SQLite database, while image files, vector files, and sample datasets are stored in files.

[0041] S3 model construction and training: According to the requirements of the interpretation target classification system, sample data of forest land and cultivated land converted to forest land are selected from the sample library, and the slice sample data set extracted from the sample library is augmented by rotation, scaling, flipping and color transformation methods; the slice sample data set is then divided into training set, validation set and test set in a ratio of 7:2:1 for subsequent forest land recognition model and cultivated land converted to forest land detection model training; according to the task objectives, appropriate semantic segmentation and change detection algorithms are selected for adaptive transformation and combination. Before the model starts training, introducing a fully trained general segmentation and general change detection pre-training model is an efficient method, which can greatly promote the convergence speed of the model during the training process and also help improve the generalization ability and accuracy of the model; finally, the network hyperparameters are set, the network is trained and iteratively optimized, the model accuracy is evaluated in the test set, and the forest land recognition model and cultivated land converted to forest land detection model with the optimal accuracy are saved;

[0042] S4 Extraction of suspected cultivated land outflow patches: Based on the forestland recognition model with optimal accuracy, the latest image is predicted and inferred, forest patches are automatically extracted, and the forest patches are matched with the cultivated land boundary line data in the land change survey data of the previous year, and the patches where cultivated land is converted into forest land based on the land change survey database of the previous year are obtained as "newly discovered" suspected cultivated land flowing to forest patches; Based on the cultivated land to forest land detection model with optimal accuracy, two images of the same area at different times are directly predicted and inferred, and the "newly occurring" suspected cultivated land flowing to forest patches are automatically compared and extracted from the two images; The "newly discovered" suspected cultivated land flowing to forest patches obtained based on the forestland recognition model and the "newly occurring" suspected cultivated land flowing to forest patches obtained based on the cultivated land to forest land detection model are fused and complemented to obtain the suspected cultivated land flowing to forest patches in the target area, thus achieving efficient data integration and significant improvement in data quality;

[0043] S5 data post-processing: set the minimum patch area, remove patches that do not meet the requirements, and obtain preliminary patches of suspected farmland-to-forestland migration;

[0044] S6 Manual Verification: Based on the suspected cultivated land outflow monitoring spots and images issued by the state, internal verification is conducted on the correctness of the boundaries and attributes of the preliminary results spots of suspected cultivated land flowing to forest land in the target area; for the spots that are still in doubt, further field verification is carried out after marking to obtain the suspected cultivated land flowing to forest land results spots after manual verification.

[0045] Example 3:

[0046] like Figure 1 As shown, the present invention provides an intelligent method for extracting cultivated land outflow patches, comprising the following steps:

[0047] S1 Data collection and preprocessing: Based on business needs (taking the conversion of cultivated land into water bodies as an example), remote sensing image data, land change survey data, and geographic national conditions monitoring data are collected as data sources for non-cultivated land identification samples and cultivated land change detection samples; national suspected cultivated land outflow monitoring patches and images are collected as reference standards to verify the correctness of the boundaries and attributes of the automatically extracted preliminary suspected cultivated land outflow patches; the collected data and images are preprocessed: the geographic coordinate system of vectors and images is unified to the CGCS2000 national geodetic coordinate system, the data format of images is unified to tif, the remote sensing images are color-uniformed, and then the images and vector data are framed and renamed according to the 1:50,000 standard frame;

[0048] S2 sample dataset production and storage: The interpretation target is determined based on image resolution, spectral characteristics, and the results of farmland protection business scenario analysis. A classification system for interpretation targets is established by referring to the classification system of land change survey data and geographical national conditions monitoring data. The sample dataset is produced through two complementary methods: the first is to extract samples based on existing data: the land change survey database and geographical national conditions monitoring database are sorted out in combination with supporting images to directly extract vectors and images that meet the interpretation target samples, realizing the effective utilization and activation of existing data stocks; the second is based on manual image collection: for parts that may be missed or insufficient in the first method, supplementary measures such as manual collection and fine annotation are adopted;

[0049] Convert the labeled interpretation target vector sample data into the corresponding rasterized label image, so that the raster image has the same resolution and size as the corresponding image. In order to meet the requirements of the deep learning model for the input image size, the image and the corresponding raster label are cropped to obtain the sliced ​​sample data and store it in the database;

[0050] Remote sensing interpretation sample library data storage structure (such as Figure 2 As shown in the figure, the main content of the sample library consists of three parts: images, vectors, and samples. Images include image metadata and image files, vectors include vector metadata and vector files, and samples include sample dataset metadata, sample metadata, and sample dataset files after slicing. The metadata is stored in an SQLite database, while image files, vector files, and sample datasets are stored in files.

[0051] S3 model construction and training: According to the requirements of the interpretation target classification system, sample data of water bodies and cultivated land transformed into water bodies are selected from the sample library, and the slice sample data set extracted from the sample library is augmented by rotation, scaling, flipping and color transformation methods; the slice sample data set is then divided into training set, validation set and test set in a ratio of 8:1:1 for subsequent water body recognition model and cultivated land transformed into water body detection model training; according to the task objectives, appropriate semantic segmentation and change detection algorithms are selected for adaptive transformation and combination. Before the model starts training, introducing a fully trained general segmentation and general change detection pre-training model is an efficient method, which can greatly promote the convergence speed of the model during the training process and also help improve the generalization ability and accuracy of the model; finally, the network hyperparameters are set, the network is trained and iteratively optimized, the model accuracy is evaluated in the test set, and the water body recognition model and cultivated land transformed into water body detection model with the optimal accuracy are saved;

[0052] S4 Extraction of suspected cultivated land outflow patches: Based on the water body recognition model with optimal accuracy, the latest image is predicted and inferred, and water body patches are automatically extracted. The water body patches are then overlaid with the cultivated land boundary line data in the land change survey data of the previous year, and the patches where cultivated land has become water bodies based on the land change survey database of the previous year are obtained as "newly discovered" suspected cultivated land outflow water body patches; Based on the cultivated land change water body detection model with optimal accuracy, two images of the same area at different times are directly predicted and inferred, and the "newly occurring" suspected cultivated land outflow water body patches are automatically compared and extracted from the two images; The "newly discovered" suspected cultivated land outflow water body patches obtained based on the water body recognition model and the "newly occurring" suspected cultivated land outflow water body patches obtained based on the cultivated land change water body detection model are fused and complemented to obtain the suspected cultivated land outflow water body patches in the target area, achieving efficient data integration and significant improvement in data quality;

[0053] S5 data post-processing: set the minimum patch area, remove patches that do not meet the requirements, and smooth the patches of suspected farmland flowing to water bodies in the target area to obtain preliminary patches of suspected farmland flowing to water bodies;

[0054] S6 Manual Verification: Based on the suspected cultivated land outflow monitoring spots and images issued by the state, an internal verification is conducted on the correctness of the boundaries and attributes of the preliminary results spots of suspected cultivated land flowing to water bodies in the target area; for those spots that are still in doubt, further field verification is carried out after marking to obtain the suspected cultivated land flowing to water bodies results spots after manual verification.

[0055] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. An intelligent method for extracting cultivated land outflow patches, comprising the following steps: S1 Data Collection and Preprocessing: Based on business needs, remote sensing image data, land change survey data, geographic national conditions monitoring data, and suspected cultivated land outflow monitoring maps and images issued by the state are collected; the collected data and images are preprocessed: the geographic coordinate system of vectors and images is unified to the CGCS2000 national geodetic coordinate system, the image data format is unified to tif, the remote sensing images are color-uniformed, and the image and vector data are framed and renamed according to the "National Basic Scale Topographic Map Framing and Numbering" by selecting the standard framing frame; S2 sample data set preparation and storage: Determine interpretation targets based on image resolution, spectral characteristics, and farmland protection business scenario analysis results; establish an interpretation target classification system based on land change survey data and geographic national conditions monitoring data classification system; The sample dataset was produced using two complementary methods: one is to extract samples based on existing data: the land change survey database and the geographic national conditions monitoring database are sorted out in combination with supporting images to directly extract vectors and images that meet the interpretation target samples; the other is to manually collect images: for the parts missed or insufficient by the first method, supplementary measures such as manual collection and fine annotation are adopted; Convert the labeled interpretation target vector sample data into the corresponding rasterized label image, so that the raster image has the same resolution and size as the corresponding image; crop the image and the corresponding raster label to obtain the slice sample data and store it in the database; The main content of the remote sensing interpretation sample library consists of three parts: images, vectors, and samples. Images include image metadata and image files, vectors include vector metadata and vector files, and samples include sample dataset metadata, sample metadata, and sample dataset files after slicing. Metadata is stored in an SQLite database, while image files, vector files, and sample datasets are stored in files. S3 model construction and training: Flexible selection and assembly of required sample data from the sample library. Sample augmentation of the slice sample dataset extracted from the sample library is performed using rotation, scaling, flipping, and color transformation methods. The slice sample dataset is then proportionally divided into training, validation, and test sets. Based on the task objectives, appropriate semantic segmentation and change detection algorithms are selected for adaptive transformation and combination. Before model training begins, fully trained general segmentation and general change detection pre-trained models are introduced. Finally, network hyperparameters are set, the network is iteratively optimized through training, and model accuracy is evaluated in the test set. The non-cultivated land recognition model and cultivated land change detection model with the optimal accuracy are saved. S4 Extraction of suspected cultivated land outflow patches: Based on the non-cultivated land identification model with optimal accuracy, the latest image is predicted and inferred, and non-cultivated land identification patches are automatically extracted. The non-cultivated land patches are then overlaid with the cultivated land boundary line data in the previous year's land change survey data, and the patches that have changed within the cultivated land scope based on the previous year's land change survey database are obtained as "newly discovered" suspected cultivated land outflow patches; based on the cultivated land change detection model with optimal accuracy, two images of the same area at different times are directly predicted and inferred, and the "newly occurring" suspected cultivated land outflow patches are automatically extracted by comparing the two images; the "newly discovered" suspected cultivated land outflow patches obtained based on the non-cultivated land identification model and the "newly occurring" suspected cultivated land outflow patches obtained based on the cultivated land change detection model are fused and complemented to obtain the suspected cultivated land outflow patches of the target area; S5 data post-processing: Based on the extraction type requirements, batch graphic regularization processing is performed on the suspected cultivated land outflow patches in the target area, including graphic topology optimization, square processing, small patch removal, and smoothing processing, to obtain preliminary results of suspected cultivated land outflow patches; S6 Manual Verification: Based on the suspected cultivated land outflow monitoring spots and images issued by the state, internal verification is conducted on the correctness of the boundaries and attributes of the preliminary results of suspected cultivated land outflow in the target area; for spots that are still in doubt, further field verification is carried out after marking to obtain the suspected cultivated land outflow results spots after manual verification.

2. The intelligent method for extracting cultivated land outflow patches according to claim 1, wherein the standard framing frame in step S1 is 1:10,000, 1:25,000 or 1:50,000.

3. An intelligent cultivated land outflow patch extraction method as claimed in claim 1, wherein the slice sample data set partition ratio in step S3 is 6:2:2, 7:2:1 or 8:1:

1.

4. An intelligent method for extracting cultivated land outflow patches as described in claim 1, wherein in step S5 data post-processing, when there is overlap between patches, the boundary relationship is adjusted according to the geographic spatial position, and the overlapping part is deleted to perform graphic topology optimization; when the automatically extracted house boundaries are relatively smooth, square processing is performed on the basis of automatic extraction; small patches are eliminated by setting a minimum patch area to eliminate patches that do not meet the requirements; smoothing processing is mainly performed on linear features.

Images