A fast extraction method of cultivated land blocks by deep learning based on partition theory and knowledge graph
Through the method based on partition theory and knowledge graph, combined with the TST-Unet model and random forest algorithm, the accuracy of arable land plot extraction and deep learning model accuracy in hilly and mountainous areas is solved, and efficient and intelligent extraction of arable land monitoring is achieved. Improvement of model generalization capabilities.
Patent Information
- Application Number
- CN202410618050.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-17
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-05-17
AI Technical Summary
The existing arable land monitoring methods are weak in generalization during large-scale monitoring, especially when plots with high fracture degree and planting area in hilly and mountainous areas, it is difficult to accurately extract arable land information, and the lack of high-quality training samples limits the The accuracy of deep learning models.
Using a method based on partitioning theory and knowledge graph, the topography type classification is carried out through topography DEM and natural geographical zoning to construct a high-quality arable land sample database, and using the TST-Unet model to combine the Transformer structure and random forest algorithm to perform high-resolution remote sensing imaging. deep semantic segmentation of cultivated land plots.
It realizes the accurate extraction of edge information of cultivated land plots in hilly and mountainous areas, improves the accuracy of cultivated land monitoring and the generalization ability of deep learning models, and solves the problems of low edge segmentation accuracy and low efficiency of training data set production in traditional methods. .
Smart Images

Figure CN118799722B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and specifically to a method for fast extraction of cultivated land blocks through deep learning based on partition theory and knowledge graph. Background Art
[0002] Cultivated land is the basic resource and condition for human survival. It has the functions of ensuring national food security, meeting the land demand for industrialization and urbanization, and ecological environment construction. Real-time and accurate monitoring of the distribution and changes of cultivated land resources is an important prerequisite for promoting the modernization of agricultural production. It is also an urgent need for macro-agricultural policy formulation, agricultural production management, agricultural resource protection and green sustainable development. Using remote sensing technologies such as sky-ground multi-source remote sensing data to carry out cultivated land monitoring can meet most of the needs. With the improvement of the resolution and accessibility of remote sensing data sources, the classification and processing technology of remote sensing data has been continuously improved, and the overall accuracy and timeliness of cultivated land remote sensing monitoring have been continuously improved. However, most of the existing cultivated land monitoring methods are applicable to a single spatiotemporal domain. When conducting large-scale cultivated land monitoring, the generalization ability of the models and methods is weak, and the universality is relatively low. In addition, traditional remote sensing information extraction methods, such as maximum likelihood, support vector machine, random forest, etc., are based on pixel classification, and their classification results often show "pepper and salt phenomenon", and the integrity, boundary information and extraction accuracy of the plot are relatively low. In existing studies, cultivated land information extraction has achieved good classification results in plain areas with dense and single planting. However, the characteristics of cultivated land in my country are complex and diverse, especially in hilly and mountainous areas, where the terrain is undulating and the fragmentation of plots is high. In addition, the planting area and the shape and size of the plot boundaries are different, resulting in large spatial heterogeneity and increasing difficulty in extracting cultivated land information. With the continuous improvement of image spatial resolution, more refined spectral and spatial characteristics of objects are presented, which can provide data support for irregular cultivated land information in hilly and mountainous areas. In addition, deep learning methods based on multi-spectral features of images provide technical support for image classification under complex conditions. How to accurately extract cultivated land plots in fragmented areas through deep learning based on high-resolution images? This is the problem solved by this technology. The TST-Unet deep learning model proposed in the present invention can accurately obtain the detailed edge information of fragmented plots, and realize the rapid and intelligent extraction of cultivated land in mountainous areas.
[0003] The quantity and quality of training samples determine the detection accuracy of the monitoring model. In particular, the construction of a change detection model based on deep learning requires a large amount of high-precision remote sensing training sample data as input. Obtaining samples through visual interpretation and field surveys is time-consuming, labor-intensive, and inefficient. At this stage, research often directly uses existing data classification products for model training, but due to the limited accuracy of related data sets, the generalization ability of deep learning is poor. The lack of a sufficient number of high-quality training samples is a major factor limiting the accuracy of deep semantic segmentation networks. Therefore, how to quickly obtain high-quality training samples in large areas for deep learning training? This is an important part of the problem solved by this technology. The present invention proposes a plan for the construction of a cultivated land knowledge graph database based on partitioning theory to provide support and guarantee for deep learning.
[0004] There is currently no targeted, high-quality deep learning training dataset for the extraction of cultivated land from medium- and high-resolution remote sensing data; or the existing training datasets are mostly based on visual interpretation and have low production efficiency; or the segmentation results of existing deep semantic segmentation networks have problems such as low edge segmentation accuracy, which limits the extraction accuracy of remote sensing data.
[0005] In order to address the defects and needs of the above-mentioned traditional methods, such as high complexity and low efficiency, a deep learning rapid extraction method of cultivated land blocks based on partitioning theory and knowledge graph is needed. Summary of the invention
[0006] The purpose of the present invention is to provide a method for fast extraction of cultivated land blocks by deep learning based on partition theory and knowledge graph. The present invention constructs an integrated method involving the production of high-quality training samples in large areas and deep semantic segmentation learning and extraction of cultivated land blocks. Based on this method, it helps to accelerate the generalization application and migration ability of deep learning technology in the field of agricultural remote sensing.
[0007] The present invention is implemented in this way. The present invention provides a method for fast extraction of cultivated land blocks by deep learning based on partition theory and knowledge graph, which is specifically performed in the following steps:
[0008] S1: Based on the zoning and stratification theory, with terrain DEM as the basis, combined with natural geographical divisions and landform types, the terrain of regional remote sensing images is visually judged and the landform types are classified; when judging the regional remote sensing images, if plains dominate the terrain, the area is determined to be a single terrain area; when mountains dominate the terrain in the regional remote sensing images, the area is determined to be a complex terrain area.
[0009] S2: Collect and form a basic database of cultivated land samples, as well as sample point data collected on the spot and sample point data visually interpreted based on Google Earth high-resolution images, and create a database based on the ARCGIS software platform using the two basic formats of TIFF raster data and SHP vector data included in the cultivated land samples;
[0010] The formation of the basic database of cultivated land samples specifically includes the disclosure of data platforms, such as land use type data from platforms such as the National Agricultural Science Data Center, the Resources and Environmental Science Data Center of the Chinese Academy of Sciences, and the National Qinghai-Tibet Plateau Science Data Center, and land use data products that have been publicly published by different scientific research institutions, such as Tsinghua University's China 1980-2015 annual 30-meter land cover / land use dataset, Wuhan University's 1990-2022 annual 30-meter land use dataset, etc.
[0011] S3: In the created database, a training sample set is extracted according to the classification of landform types;
[0012] The landform types are classified into large plots of cultivated land in a single area and plots with a single terrain structure. The cultivated land data are extracted based on high spatiotemporal resolution data, and the cultivated land samples in the created database are directly used as the training sample set for intelligent extraction of cultivated land in a single terrain area;
[0013] The sample data of cultivated land in the database created in the area with complex terrain were used to classify cultivated land and non-cultivated land using the random forest method.
[0014] S4: Collect and preprocess high-resolution images, select mountainous areas with severe farmland fragmentation as the test area, and collect high-spatial-resolution remote sensing images; obtain high-spatial-resolution optical remote sensing images of the study area, use remote sensing data processing software to perform geometric precision correction, and stitch them to completely cover the study area based on geographic coordinates;
[0015] The classification of cultivated land and non-cultivated land in areas with complex terrain is carried out according to the following steps:
[0016] S 4.1 : Construct feature factors, including altitude, slope and aspect, shadow features, and select three major features: spectral features, texture features, and terrain features to form the original feature set; the spectral features include spectral features of different bands and related index features calculated by the bands; the texture features are extracted using the grayscale co-occurrence matrix, and the grayscale relationship characteristics between two adjacent pixels at a certain distance in the image are calculated, and the frequency of occurrence of these pixels is counted to generate a symmetrical matrix, as shown in formula (1);
[0017] Formula (1)
[0018] in, f(x,y) represents the image needed to generate the gray-level co-occurrence matrix, (x,y) represents a point in the image, (x + dx, y + dy) for (x,y) The corresponding pixel coordinates after offset, i represent(x,y) The grayscale value of the pixel, j represent (x+ (dx, y + dy) The gray value of δ is the offset distance, θ is the offset direction, express (x,y) At an offset distance of δ , the offset direction is θ The probability of occurrence when
[0019] S 4.2 : Perform feature screening, use the distance based on conditional probability theory as a measure of the degree of separability between different features, extract some channel layers from the original data to form the feature-screened data, and use it as the input classifier for classification; as shown in formula (2);
[0020] Formula (2)
[0021] S 4.3 : Build a decision tree, use the CART algorithm to build a decision tree classifier, set the random forest algorithm tree range to (0, 1200), increase by 5, find the number of trees corresponding to the highest overall classification accuracy through multiple cycles, and use the corresponding classification result as the final classification result;
[0022] S 4.4 : Manual correction is performed based on the temporary recognition results obtained by random forest, with reference to ground survey data, Google Earth and Gaofen-2 satellite high-resolution data, and the classification results are corrected through manual visual interpretation, including removing false positives and correcting missed points. The accuracy of the final results is evaluated using confusion matrix, F1 score, user precision, producer precision and overall accuracy indicators.
[0023] S5: training label preparation and processing, based on the database established in step S3, select high-quality training sample data; perform data enhancement on the training labels;
[0024] Data enhancement includes random cropping, flipping, rotating, color jittering, and image shifting of training label data. New samples are generated while preserving the original features, effectively enhancing the robustness and generalization performance of the model.
[0025] S6: Build and train the TST-Unet model; use the Transformer structure as an encoder in the model to extract features from remote sensing images and capture longer-term features. The Transformer encoder consists of L layers of multi-head self-attention and multi-layer perceptrons. As shown in equations (3) to (5);
[0026] Formula (3)
[0027] Formula (4)
[0028] Formula (5)
[0029] Where LN is the layer normalization algorithm; MLP is the encoded image representation; MSA is multi-head self-attention, is the scaling factor in the attention mechanism;
[0030] Furthermore, a crop encoder branch is added. By training the original image divided into four parts, the result is added to the result of the original encoder after encoding and then input into the decoder, and fused with the features of the original encoder to finally output the result.
[0031] The original image is input into CNN for feature extraction, and three layers of convolution downsampling are performed. The feature map is reduced compared to the original image, becoming 1 / 2, 1 / 4 and 1 / 8 of the original image, and the corresponding feature matrices are obtained respectively. Each downsampling process includes a convolution, regularization, ReLU activation, and maximum pooling layer. The downsampled image is input into the embedded layer, and then into the Transformer layer. The intermediate high-resolution CNN feature layer is cycled 12 times for encoding and hidden feature mining.
[0032] The feature layer is sliced to extract high-level farmland plot features. In the decoding process, the image is upsampled and feature connection is completed, and then a depth-wise separable convolution substitution is performed. The upsampling is repeated 4 times. Finally, a 1×1 convolution layer is generated on the high-level semantic feature map to generate a single feature map of the farmland plot result, completing the remote sensing image segmentation and farmland pixel extraction process.
[0033] S7: Data recognition is performed using the trained TST-Unet model.
[0034] Furthermore, the present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a main controller, the method described in any one of the above is implemented.
[0035] Compared with the prior art, the present invention has the following beneficial effects:
[0036] Construct an integrated method involving the production of high-quality training samples in large areas and deep semantic segmentation learning and extraction of cultivated land plots. This method will help accelerate the generalization and migration capabilities of deep learning technology in the field of agricultural remote sensing.
[0037] It solves the problem that the existing training data sets are mostly based on visual interpretation and have low production efficiency; it solves the problem that the existing deep semantic segmentation network segmentation results have low edge segmentation accuracy and improves the extraction accuracy of remote sensing data. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. The following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.
[0039] Figure 1 is a flow chart of the method of the present invention;
[0040] Figure 2 is a schematic diagram of the TST-Unet model architecture of the present invention;
[0041] Figure 3 This is an example diagram of the deep learning extraction results of fragmented farmland in the present invention. DETAILED DESCRIPTION
[0042] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but is only for selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.
[0043] See also Figures 1 - 3 The present invention provides a method for fast extraction of cultivated land blocks through deep learning based on partition theory and knowledge graph.
[0044] In this embodiment, Figures 1 - 2 ,
[0045] S1: Based on the zoning and stratification theory, with terrain DEM as the basis, combined with natural geographical divisions and landform types, the terrain of regional remote sensing images is visually judged and the landform types are classified; when judging the regional remote sensing images, if plains dominate the terrain, the area is determined to be a single terrain area; when mountains dominate the terrain in the regional remote sensing images, the area is determined to be a complex terrain area.
[0046] S2: Collect and form a basic database of cultivated land samples, as well as sample point data collected on the spot and sample point data visually interpreted based on Google Earth high-resolution images, and create a database based on the ARCGIS software platform using the two basic formats of TIFF raster data and SHP vector data included in the cultivated land samples;
[0047] The basic database of cultivated land samples specifically includes land use type data from public data platforms (such as the National Agricultural Science Data Center, the Resources and Environmental Science Data Center of the Chinese Academy of Sciences, the National Qinghai-Tibet Plateau Science Data Center, etc.), and land use data products that have been publicly published by different scientific research institutions (such as Tsinghua University's China 1980-2015 annual 30-meter land cover / land use dataset, Wuhan University's 1990-2022 annual 30-meter land use dataset, etc.).
[0048] S3: In the created database, a training sample set is extracted according to the classification of landform types;
[0049] The landform types are classified into large plots of cultivated land in a single area and plots with a single terrain structure. The cultivated land data are extracted based on high spatiotemporal resolution data, and the cultivated land samples in the created database are directly used as the training sample set for intelligent extraction of cultivated land in a single terrain area;
[0050] The sample data of cultivated land in the database created in the area with complex terrain were used to classify cultivated land and non-cultivated land using the random forest method.
[0051] S4: Collect and preprocess high-resolution images, select mountainous areas with severe farmland fragmentation as the test area, and collect high-spatial-resolution remote sensing images; obtain high-spatial-resolution optical remote sensing images of the study area, use remote sensing data processing software to perform geometric precision correction, and stitch them to completely cover the study area based on geographic coordinates;
[0052] The classification of cultivated land and non-cultivated land in areas with complex terrain is carried out according to the following steps:
[0053] S 4.1: Construct feature factors, including altitude, slope and aspect, shadow features, and select three major features: spectral features, texture features, and terrain features to form the original feature set; the spectral features include spectral features of different bands and related index features calculated by the bands; the texture features are extracted using the grayscale co-occurrence matrix, and the grayscale relationship characteristics between two adjacent pixels at a certain distance in the image are calculated, and the frequency of occurrence of these pixels is counted to generate a symmetrical matrix, as shown in formula (1);
[0054] Formula (1)
[0055] in, f(x,y) represents the image needed to generate the gray-level co-occurrence matrix, (x,y) represents a point in the image, (x + dx, y + dy) for (x,y) The corresponding pixel coordinates after offset, i represent (x,y) The grayscale value of the pixel, j represent (x+ (dx, y + dy) The gray value of δ is the offset distance, θ is the offset direction, express (x,y) At an offset distance of δ , the offset direction is θ The probability of occurrence when
[0056] S 4.2 : Perform feature screening, use the distance based on conditional probability theory as a measure of the degree of separability between different features, extract some channel layers from the original data to form the feature-screened data, and use it as the input classifier for classification; as shown in formula (2);
[0057] Formula (2)
[0058] S 4.3 : Build a decision tree, use the CART algorithm to build a decision tree classifier, set the random forest algorithm tree range to (0, 1200), increase by 5, find the number of trees corresponding to the highest overall classification accuracy through multiple cycles, and use the corresponding classification result as the final classification result;
[0059] S 4.4 : Manual correction is performed based on the temporary recognition results obtained by random forest, with reference to ground survey data, Google Earth and Gaofen-2 satellite high-resolution data, and the classification results are corrected through manual visual interpretation, including removing false positives and correcting missed points. The accuracy of the final results is evaluated using confusion matrix, F1 score, user precision, producer precision and overall accuracy indicators.
[0060] S5: training label preparation and processing, based on the database established in step S3, select high-quality training sample data; perform data enhancement on the training labels;
[0061] Data enhancement includes random cropping, flipping, rotating, color jittering, and image shifting of training label data. New samples are generated while preserving the original features, effectively enhancing the robustness and generalization performance of the model.
[0062] S6: Figure 2 , TST-Unet model is constructed and trained; in the model, the Transformer structure is used as an encoder to extract features from remote sensing images and capture longer features. The Transformer encoder consists of L layers of multi-head self-attention and multi-layer perceptron. As shown in equations (3)-(5);
[0063] Formula (3)
[0064] Formula (4)
[0065] Formula (5)
[0066] Where LN is the layer normalization algorithm; MLP is the encoded image representation; MSA is multi-head self-attention, is the scaling factor in the attention mechanism;
[0067] In this embodiment, a cropping encoder branch is added, and the original image is divided into four parts for training. After encoding, the result is added to the result of the original encoder and then input into the decoder, and fused with the features of the original encoder to finally output the result.
[0068] The original image is input into CNN for feature extraction, and three layers of convolution downsampling are performed. The feature map is reduced compared to the original image, becoming 1 / 2, 1 / 4 and 1 / 8 of the original image, and the corresponding feature matrices are obtained respectively. Each downsampling process includes a convolution, regularization, ReLU activation, and maximum pooling layer. The downsampled image is input into the embedded layer, and then into the Transformer layer. The intermediate high-resolution CNN feature layer is cycled 12 times for encoding and hidden feature mining.
[0069] The feature layer is sliced to extract high-level farmland plot features. In the decoding process, the image is upsampled and feature connection is completed, and then a depth-wise separable convolution substitution is performed. The upsampling is repeated 4 times. Finally, a 1×1 convolution layer is generated on the high-level semantic feature map to generate a single feature map of the farmland plot result, completing the remote sensing image segmentation and farmland pixel extraction process.
[0070] S7: Figure 3 , data recognition is performed through the trained TST-Unet model.
[0071] In this embodiment, the present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a main controller, the method described in any one of the above is implemented.
[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention has various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A deep learning method for fast extraction of cultivated land blocks based on partition theory and knowledge graph, characterized by: Please follow the steps below: S1: Based on the theory of zoning and stratification, taking terrain DEM as the basis, combined with natural geographical divisions and landform types, the terrain of regional remote sensing images is visually judged and the landform types are classified; S2: Collect and form a basic database of cultivated land samples, as well as sample point data collected on the spot and sample point data visually interpreted based on Google Earth high-resolution images, and create a database based on the ARCGIS software platform using the two basic formats of TIFF raster data and SHP vector data included in the cultivated land samples; S3: In the created database, a training sample set is extracted according to the classification of landform types; S4: Collect and preprocess high-resolution images, select mountainous areas with severe farmland fragmentation as the test area, and collect high-spatial-resolution remote sensing images; obtain high-spatial-resolution optical remote sensing images of the study area, use remote sensing data processing software to perform geometric precision correction, and stitch them to completely cover the study area based on geographic coordinates; S5: training label preparation and processing, based on the database established in step S3, select high-quality training sample data; perform data enhancement on the training labels; S6: Build and train the TST-Unet model; use the Transformer structure as an encoder in the model to extract features from remote sensing images and capture longer-term features. The Transformer encoder consists of L layers of multi-head self-attention and multi-layer perceptrons; Add a crop encoder branch, which is trained by dividing the original image into four parts. After encoding, the result is added to the result of the original encoder and then input into the decoder, and then fused with the features of the original encoder to finally output the result. The original image is input into CNN for feature extraction, and three layers of convolution downsampling are performed. The feature map is reduced compared to the original image, becoming 1 / 2, 1 / 4 and 1 / 8 of the original image, and the corresponding feature matrices are obtained respectively. Each downsampling process includes a convolution, regularization, ReLU activation, and maximum pooling layer. The downsampled image is input into the embedded layer, and then input into the Transformer layer to cycle 12 times for the intermediate high-resolution CNN feature layer for encoding and hidden feature mining. The feature layer is sliced to extract high-level farmland plot features. After completing feature connection in the upsampling of the image during the decoding process, a depth-separable convolution substitution is performed. The upsampling is repeated 4 times. Finally, a 1×1 convolution layer is generated on the high-level semantic feature map to generate a single feature map of the farmland plot result, completing the remote sensing image segmentation and farmland pixel extraction process. S7: Data recognition is performed using the trained TST-Unet model.
2. The method for rapid extraction of cultivated land blocks by deep learning based on partition theory and knowledge graph according to claim 1 is characterized in that: In step S1, the regional remote sensing image is judged. If the plains dominate the terrain, the region is determined to be a simple terrain area; if the mountainous areas dominate the terrain in the regional remote sensing image, the region is determined to be a complex terrain area.
3. The method for rapid extraction of cultivated land blocks by deep learning based on partition theory and knowledge graph according to claim 1 is characterized in that: In step S2, a basic database of cultivated land samples is formed, which specifically includes land use type data disclosed by the data platform and land use data products that have been publicly published by different scientific research institutions.
4. The method for rapid extraction of cultivated land blocks by deep learning based on partition theory and knowledge graph according to any one of claims 1 or 2, characterized in that: In step S3, the landform type classification terrain includes single area cultivated land plots, single terrain structure plots, cultivated land data extraction based on high spatiotemporal resolution data, and the cultivated land samples in the created database are directly used as the training sample set for intelligent extraction of cultivated land in single terrain area; The sample data of cultivated land in the database created in the area with complex terrain were used to classify cultivated land and non-cultivated land using the random forest method.
5. The method for rapid extraction of cultivated land blocks by deep learning based on partition theory and knowledge graph according to claim 4 is characterized in that: The classification of cultivated land and non-cultivated land in areas with complex terrain is carried out according to the following steps: S 4.1 : Construct feature factors, including altitude, slope and aspect, shadow features, and select three major features: spectral features, texture features, and terrain features to form the original feature set; the spectral features include spectral features of different bands and related index features calculated by the bands; the texture features are extracted using the grayscale co-occurrence matrix, and the grayscale relationship characteristics between two adjacent pixels at a certain distance in the image are calculated, and the frequency of occurrence of these pixels is counted to generate a symmetrical matrix, as shown in formula (1); Formula (1) in, f(x,y) represents the image needed to generate the gray-level co-occurrence matrix, (x,y) represents a point in the image, (x +dx,y+dy) for (x,y) The corresponding pixel coordinates after offset, i represent (x,y) The grayscale value of the pixel, j represent (x+dx,y+ dy) The gray value of δ is the offset distance, θ is the offset direction, express (x,y) At an offset distance of δ , the offset direction is θ The probability of occurrence when S 4.2 : Perform feature screening, use the distance based on conditional probability theory as a measure of the degree of separability between different features, extract some channel layers from the original data to form the feature-screened data, and use it as the input classifier for classification; as shown in formula (2); Formula (2) S 4.3 : Build a decision tree, use the CART algorithm to build a decision tree classifier, set the random forest algorithm tree range to (0, 1200), increase by 5, find the number of trees corresponding to the highest overall classification accuracy through multiple cycles, and use the corresponding classification result as the final classification result; S 4.4 : Manual correction is performed based on the temporary recognition results obtained by random forest, with reference to ground survey data, Google Earth and Gaofen-2 satellite high-resolution data, and the classification results are corrected through manual visual interpretation, including removing false positives and correcting missed points. The accuracy of the final results is evaluated using confusion matrix, F1 score, user precision, producer precision and overall accuracy indicators.
6. The method for rapid extraction of cultivated land blocks by deep learning based on partition theory and knowledge graph according to claim 1 is characterized in that: In step S5, data enhancement includes random cropping, flipping, rotating, color jittering, and image shifting of the training label data.
7. The method for rapid extraction of cultivated land blocks by deep learning based on partition theory and knowledge graph according to claim 1 is characterized in that: In step S6, the Transformer structure is used as an encoder in the model to extract features from remote sensing images and capture longer-term features. The Transformer encoder consists of L layers of multi-head self-attention and a multi-layer perceptron.
8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the main controller, the method according to any one of claims 1 to 7 is implemented.
