A cultivated land non-grain change detection method based on remote sensing image

Abstract

The present application relates to a kind of cultivated land non-grain change detection method based on remote sensing image, comprising: obtaining remote sensing image containing cultivated land, cultivated land changes garden, cultivated land changes forest, cultivated land changes facility farm, cultivated land changes water surface, cultivated land changes abandoned six types of ground objects at different time, carry out feature analysis to remote sensing image, form interpretation feature library;According to the standard of interpretation feature library, construct the sample library containing six types of ground objects and the negative sample of non-six types of ground objects, label and information enhancement processing are carried out to each type of sample;Based on DASNet and PSPNet joint construction cultivated land non-grain detection model, introduce WDMC loss function and use sample library to carry out model training, to be detected image is input into the cultivated land non-grain detection model trained and carries out non-grain change map spot extraction, and carries out precision evaluation.Solved the extraction precision of prior art in mixed planting and other diversified agricultural mode is poor, conventional change detection algorithm is greatly influenced by pseudo change, lack of robustness and other problems.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image recognition technology and relates to a method for detecting changes in cultivated land that are no longer used for grain production based on remote sensing images. Background Technology

[0002] As a crucial component of food production, the use of arable land directly determines food production capacity. With rapid urbanization and economic development, some arable land is being repurposed, such as for industrial use, residential use, agricultural facilities, green spaces, and water bodies. This non-grain conversion has a significant impact on agricultural resources and food security. To effectively strengthen arable land protection and land use control, local governments have introduced numerous monitoring programs for arable land protection, explicitly proposing the comprehensive use of modern information technologies such as satellite remote sensing for regular dynamic monitoring. Utilizing remote sensing imagery to detect changes in land features is of great significance for monitoring non-grain conversion.

[0003] Currently, existing technologies typically use manual visual methods or general change detection algorithms to interpret and analyze remote sensing images from different time phases. However, existing technologies lack the establishment of sample interpretation feature standards and the construction of samples after changes in cultivated land, resulting in the extraction of many false changes based on very limited samples, and the robustness of the model is poor. Summary of the Invention

[0004] The technical problem solved by this invention is to overcome the shortcomings of the prior art and propose a method for detecting changes in cultivated land that are no longer used for grain production based on remote sensing images. This method addresses the problems of poor extraction accuracy under diverse agricultural patterns such as mixed planting, the significant impact of spurious changes on existing conventional change detection algorithms, and the lack of robustness.

[0005] The solution of this invention is: a method for detecting changes in cultivated land conversion to non-grain crops based on remote sensing imagery, comprising the following steps:

[0006] Acquire remote sensing images of six types of land features at different times, including cultivated land, cultivated land converted to orchards, cultivated land converted to forest land, cultivated land converted to facility agriculture, cultivated land converted to ponds or reservoirs, and cultivated land converted to abandoned land. Perform feature analysis on the remote sensing images to form an interpretation feature database.

[0007] Based on the standards of the interpretation feature library, a sample library containing positive samples of six types of land features and negative samples of non-six types of land features is constructed, and each type of sample is labeled and information augmented.

[0008] A farmland non-grain conversion detection model is jointly constructed based on the dual-attention fully convolutional Siamese neural network DASNet and the semantic segmentation network PSPNet. The WDMC loss function is introduced and the model is trained using a constructed sample library. The image to be detected is input into the trained farmland non-grain conversion detection model to extract non-grain conversion change patches and evaluate the accuracy.

[0009] Furthermore, the resolution of the acquired remote sensing images is no less than 0.5m.

[0010] Furthermore, the feature analysis of the remote sensing images to form an interpretation feature library includes:

[0011] 1) Characteristics of arable land: Identified by referring to the reflectance spectral characteristics and texture features of the corresponding crops;

[0012] 2) Characteristics of newly added orchards in cultivated land: On four-band remote sensing images, the changes from cultivated land to orchards are identified by the differences in the spectral characteristics of crops and the temporal characteristics of crops.

[0013] 3) Characteristics of newly added forest land in cultivated land: On four-band remote sensing images, the Normalized Difference Vegetation Index (NDVI) is constructed using the near-infrared and red bands. Combined with texture features, forest land in cultivated land is identified. Compared with cultivated land, forest land has a coarser texture. When the newly added forest land coverage in a piece of cultivated land exceeds 40%, it is considered that the cultivated land has changed.

[0014] 4) Characteristics of newly added facility agriculture in cultivated land: When newly added facility agriculture plots appear in cultivated land plots, the standard deviation of the plot spectrum, the standard deviation of the sub-object mean, and the contrast based on the gray-level co-occurrence matrix all increase significantly. The gray level of the plot changes significantly, the texture is coarser and more obvious, and the brightness increases significantly.

[0015] 5) Characteristics of newly added ponds and reservoirs in cultivated land: On four-band remote sensing images, the normalized water index (NDWI) is constructed using the green band and near-infrared band to identify newly added ponds and reservoirs in cultivated land; the NDWI of the water surface objects is greater than 0.3, the brightness is less than 0.5, the water color is less than 0.6, and the ratio of the near-infrared band is less than 0.15.

[0016] 6) Characteristics of newly abandoned farmland: Abandoned plots are identified by combining the NDVI index with the texture of ground features;

[0017] The interpretation feature library includes specific features and example images of six types of land features.

[0018] Furthermore, the annotation of each type of sample includes:

[0019] The location and boundaries of non-grain crop changes are marked in remote sensing images: First, manual interpretation and annotation are performed using ArcGIS software. After quality inspection and manual verification, the annotation results are organized and stored to provide accurate training data for subsequent non-grain crop identification and classification tasks.

[0020] Furthermore, the data sources used for labeling each type of sample include remote sensing images acquired from different time periods, regions, resolutions, and sensors.

[0021] Furthermore, the sample data in the sample library meets the requirements of the "Construction Specification for Intelligent Interpretation Sample Library of Domestic Satellite Remote Sensing Image". The average accuracy of the sample boundary is better than 2 pixels, and the delineated effective range vector does not exceed the effective range of the image itself. The sample attributes are complete, including image source, number of bands, resolution, spatial information, feature description, and sample type information. It supports multi-band samples, the sample spatial distribution is uniform, each sample covers the full frame, and there is no overlap between different samples.

[0022] Furthermore, the information enhancement processing includes geometric transformation, color transformation, and noise addition of the image, applying the selected enhancement method to the original sample data to generate new, enhanced samples; the labels of the enhanced samples remain consistent with those of the original samples.

[0023] Furthermore, the dual-attention fully convolutional Siamese neural network (DASNet) is used to extract variation information between input image pairs. It employs deep learning methods, constructing a deep network structure through a series of convolutional and pooling layers to capture high-level features in the images, specifically including:

[0024] The two input images are subjected to convolution operations to extract features and generate feature maps for each image. The two input images are images of the same scene at different time points.

[0025] After extracting the feature maps of the two images, DASNet combines them using a specific fusion strategy to generate a fused feature map containing change information. The specific fusion strategy includes element-wise subtraction, multiplication, or concatenation.

[0026] DASNet introduces an attention mechanism to learn and emphasize regions related to change in the fused feature maps, suppresses irrelevant information, and highlights the truly changing regions.

[0027] DASNet transforms the attention-weighted fused feature map into a change prediction map through one or more convolutional layers. The change prediction map is represented in binary form, with each pixel indicating whether a change has occurred at the corresponding location.

[0028] Furthermore, the semantic segmentation network PSPNet is used to perform semantic segmentation on the input image, identify and label different objects or regions in the image, specifically including:

[0029] A change prediction map is generated using DASNet, and then compared with the original image or its features. Figure 1 The input is then fed into PSPNet for semantic segmentation;

[0030] PSPNet uses a deep convolutional neural network as its backbone network to extract deep features from the input image. It captures global contextual information through pooling operations at different scales and combines this information with the original feature map to generate an enhanced feature map containing rich contextual information.

[0031] After passing through a series of convolutional layers, the enhanced feature map is converted into a semantic segmentation prediction map. The semantic segmentation prediction map is a multi-channel probability map, where each channel corresponds to a category, and the value of each pixel represents the probability that the pixel belongs to the corresponding category.

[0032] Furthermore, the step of using the constructed sample library for model training includes:

[0033] Data sets for training and testing are collected from the sample library, including pairs of images of the same scene at different time points and corresponding label information, annotations of change regions, and labels for semantic segmentation.

[0034] The collected data is preprocessed, including image scaling, cropping, and normalization, to ensure that the data format and distribution of the input model are consistent. At the same time, the label data is appropriately transformed as needed for comparison with the model output.

[0035] Initialize the weights and parameters of DASNet and PSPNet, including setting the initial weight values ​​of the network layers, selecting the optimization algorithm and learning rate;

[0036] Image pairs from the training dataset are input into the model, and the network output is computed through forward propagation, including the change prediction map generated by DASNet and the semantic segmentation prediction map generated by PSPNet.

[0037] The loss function value is calculated based on the model output and label data: for change detection tasks, binary cross-entropy loss is used; for semantic segmentation tasks, pixel-level cross-entropy loss or Dice loss is used.

[0038] The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the network weights are updated using an optimization algorithm to make the model output closer to the labeled data.

[0039] During training, the model is evaluated periodically using a validation set until the evaluation metrics stabilize and the model training and testing are complete.

[0040] The beneficial effects of this invention compared to the prior art are:

[0041] (1) This invention solves the bottleneck problem of poor extraction accuracy in the existing technology under diversified agricultural models such as mixed planting. By analyzing the differences in the growth and development of different crops in the demonstration area, and combining the phenological information of crops in the study area, the key phenological characteristics are selected to construct a sample library for detecting "non-grain" changes that is adapted to local characteristics.

[0042] (2) In order to overcome the problem that existing conventional change detection algorithms are greatly affected by false changes and lack robustness, this invention uses DAS as the core algorithm for change detection of high-resolution images. Through a dual attention mechanism, it captures key information to obtain more discriminative feature representation, thereby enhancing the extraction and recognition performance in mixed planting agricultural mode. Attached Figure Description

[0043] Figure 1 This is a flowchart of a method for detecting changes in cultivated land conversion to non-grain crops based on remote sensing imagery, according to an embodiment of the present invention.

[0044] Figure 2 This is an image showing the extraction results of the experimental area in an embodiment of the present invention. Detailed Implementation

[0045] The present invention will be further described below with reference to the embodiments.

[0046] Example 1

[0047] To overcome the problem that conventional change detection algorithms are greatly affected by false changes, this embodiment first establishes an interpretation feature library for cultivated land and five major land change categories, and constructs a sample library for six major land categories to enhance the difference features of samples. It then uses DAS to construct a model algorithm for detecting the non-grain conversion of cultivated land, and selects a research area in Gutian County, Fujian Province as an experimental area for technology verification and testing.

[0048] The operating system used in this embodiment is Ubuntu 20.0, the CPU is an Intel(R) Xeon(R) Platinum 8358P CPU @ 2.60GHz, the GPU is an RTX 3090, the memory is 64GB, and the hard disk space is 1TB. The deep learning framework is PyTorch 1.10.0, and the CUDA version is 11.3. This embodiment selects a region in Gutian County, Fujian Province, my country as the experimental area. Image quality was screened for the experimental area, and 74 scenes of sub-meter level domestic high-resolution remote sensing image data were selected to obtain monthly experimental area data.

[0049] like Figure 1 As shown, a method for extracting information on the non-grain conversion of arable land based on remote sensing imagery includes the following steps:

[0050] S1. Construction of the interpretation feature database: Acquire remote sensing images of six types of land features at different times, including cultivated land, cultivated land converted to orchards, cultivated land converted to forest land, cultivated land converted to facility agriculture, cultivated land converted to ponds and water surfaces, and cultivated land converted to abandoned land. Perform feature analysis on the remote sensing images to form an interpretation feature database.

[0051] The resolution of the acquired remote sensing images shall be no less than 0.5m;

[0052] Feature analysis of remote sensing images is performed as follows:

[0053] 1) Characteristics of arable land: Arable land refers to permanent farmland used for growing local staple crops. In Gutian County, Fujian Province, rice is the main crop, but other crops include potatoes, sweet potatoes, and peanuts. Identification can be made primarily by referring to the reflectance spectral characteristics and texture features of the corresponding crops.

[0054] 2) Characteristics of newly added orchards in cultivated land: In four-band (red, green, blue, and near-infrared) remote sensing imagery, we primarily identify the transformation of cultivated land into orchards based on two main characteristics: First, the spectral characteristics of crops often vary depending on crop type, growing season, growth status, and field management. Therefore, scientifically and rationally utilizing these differences in crop spectral characteristics can distinguish between staple crops and orchard crops such as citrus, peaches, and plums. Second, we mainly identify the transformation from cultivated land to orchards based on the temporal characteristics of crops. Distinguishing between crop categories solely based on spectral characteristics is insufficient. Because different crops have specific growth patterns and phenological characteristics, the spectral characteristics of the same crop differ at different growth stages, and even within the same growth period, different crops exhibit spectral differences. Therefore, fully utilizing the typical seasonal rhythms of crops is a key theoretical basis for distinguishing between crops and between crops and other green vegetation.

[0055] 3) Characteristics of newly added forest land in cultivated land: In four-band (red, green, blue, and near-infrared) remote sensing images, vegetation shows strong reflectance in the near-infrared band and weak reflectance in the red band. A Normalized Difference Vegetation Index (NDVI) is constructed using the near-infrared and red bands, and combined with texture features, forest land in cultivated land can be identified. Compared to cultivated land, forest land has a coarser texture. When the newly added forest coverage in a large area of ​​cultivated land exceeds 40%, we consider the cultivated land to have changed, thus ruling out interference from windbreaks in the cultivated land.

[0056] 4) Characteristics of newly added facility agriculture in arable land: Facility agriculture refers to land used directly for planting, breeding, and crop processing, including production facilities and ancillary facilities, generally existing in the form of hardened ground. Once a new facility agriculture type plot appears within an arable land plot, several characteristics such as the plot's spectral standard deviation, the sub-object mean standard deviation, and the contrast based on the gray-level co-occurrence matrix will significantly increase. The plot's gray level will change noticeably, the texture will become coarser and more pronounced, and the brightness will also increase significantly.

[0057] 5) Characteristics of newly added ponds and reservoirs in farmland: In four-band (red, green, blue, and near-infrared) remote sensing images, the overall brightness of water surfaces is relatively low, with strong reflection in the green band and weak reflection in the near-infrared band. Using the green and near-infrared bands to construct the Normalized Difference Water Index (NDWI) can identify newly added ponds and reservoirs in farmland. The NDWI of water surfaces is greater than 0.3, and the brightness is generally less than 0.5. When eutrophication or other pollution occurs, the situation becomes more complex. Using only NDWI and brightness features, water bodies are easily confused with shadows, newly constructed asphalt roads, and farmland covered by shallow water. Compared to shadows and other dark features, water bodies have lower chromaticity, generally less than 0.6, and the near-infrared ratio is less than 0.15.

[0058] 6) Characteristics of newly abandoned farmland: Abandonment refers to the situation where existing farmland is left uncultivated. Abandoned plots generally contain weeds and bare land, and can be identified by combining the NDVI index with the texture of ground features.

[0059] The interpretation feature library includes detailed features and example images of the aforementioned land features.

[0060] S2. Sample library construction and sample enhancement: Based on the standard of the interpretation feature library, a sample library containing positive samples of six types of land features and negative samples of non-six types of land features is constructed, and each type of sample is labeled and information enhancement is performed.

[0061] Among them, the positive samples of the six types of land features include six types: staple food farmland covering crops such as rice, sweet potato, potato, and peanut; farmland converted into orchards; farmland converted into forest land; farmland converted into facility agriculture; farmland converted into ponds or reservoirs; and farmland converted into abandoned land.

[0062] Labeling non-grain samples in processed images involves marking the locations and boundaries of non-grain crop changes in remote sensing images. This provides accurate training data for subsequent non-grain crop identification and classification tasks. First, manual interpretation and labeling are performed using ArcGIS software. After quality control and manual verification, the labeling results are then organized and stored.

[0063] The data sources used for labeling each type of sample include remote sensing images acquired from different time periods, regions, resolutions, and sensors.

[0064] The sample data in the sample library meets the requirements of the "Construction Specifications for Intelligent Interpretation Sample Library of Domestic Satellite Remote Sensing Image". The average accuracy of the sample boundary is better than 2 pixels, and the delineated effective range vector should not exceed the effective range of the image itself as much as possible. The sample attributes are complete, including information such as image source, number of bands, resolution, spatial information, feature description, and sample type. Multi-band samples are supported. Areas with poor image quality should not be selected as sample areas. The sample space is evenly distributed, each sample covers the full frame, and there is no overlap between different samples.

[0065] Information augmentation is performed on each type of sample, including:

[0066] Based on task requirements and data type, select appropriate augmentation methods. These methods can include geometric transformations, color transformations, noise addition, etc., of the image. Apply the selected augmentation method to the original sample data to generate new, modified samples. For supervised learning tasks, ensure that the labels of the augmented samples remain consistent with those of the original samples. Merge the augmented samples with the original samples to form a larger, more diverse dataset.

[0067] This embodiment manually labeled a total of 2800 512*512 samples, including 300 samples of farmland converted to orchards, 1000 samples of farmland converted to forest land, 800 samples of farmland converted to facility agriculture, 200 samples of farmland converted to ponds or reservoirs, and 500 samples of farmland converted to abandoned land. After processing with augmentation methods, the number of samples was expanded to 17000. Each sample data includes a label for the previous time phase and a label for the next time phase. The label type for the previous time phase is farmland, and the label types for the next time phase are orchard, forest land, facility agriculture, pond or reservoir, and abandoned land. The sample labeling format is a mask, and the blank areas are unchanged negative samples.

[0068] S3. Construction and Accuracy Evaluation of Farmland Non-Grain Conversion Detection Model: Based on dual-attentive fully convolutional siamese networks (DASNet) and semantic segmentation network PSPNet (Pyramid Scene Parsing Network), a farmland non-grain conversion detection model is jointly constructed. The WDMC (Weighted Double-Margin Contrastive) loss function is introduced, and the model is trained using the sample library established in step S2. The image to be detected is input into the trained farmland non-grain conversion detection model to extract non-grain conversion change patches, and the accuracy is evaluated.

[0069] The specific model for detecting the conversion of arable land to non-grain crops is as follows:

[0070] DASNet is primarily responsible for extracting change information between pairs of input images. It employs deep learning, constructing a deep network structure through a series of convolutional and pooling layers to capture high-level features in the images. First, feature extraction is performed on two input images (typically images of the same scene at different time points). This is usually achieved through a series of convolutional operations, generating feature maps for each image. After extracting the feature maps from the two images, DASNet combines them using specific fusion strategies (such as element-wise subtraction, multiplication, or concatenation) to generate a fused feature map containing change information. To further improve the accuracy of change detection, DASNet introduces an attention mechanism. The attention module learns and emphasizes change-related regions in the fused feature map, suppressing irrelevant information, thereby highlighting the truly changed regions. Finally, DASNet converts the attention-weighted fused feature map into a change prediction map through one or more convolutional layers. The change prediction map is typically represented in binary form, where each pixel indicates whether a change has occurred at the corresponding location.

[0071] PSPNet is primarily used for semantic segmentation of input images, i.e., identifying and labeling different objects or regions in an image. It generates a change prediction map using DASNet, and then compares this prediction map with the original image or its features. Figure 1 The input image is fed into PSPNet for semantic segmentation. This allows PSPNet to consider variations in the semantic segmentation process. PSPNet uses a deep convolutional neural network (such as ResNet) as its backbone to extract deep features from the input image. Global contextual information is captured through pooling operations at different scales, and this information is combined with the original feature map to generate an enhanced feature map containing rich contextual information. Finally, the enhanced feature map is transformed into a semantic segmentation prediction map after a series of convolutional layers. The semantic segmentation prediction map is typically a multi-channel probability map, where each channel corresponds to a class, and the value of each pixel represents the probability that the pixel belongs to the corresponding class.

[0072] The model is trained using the sample library established in step S2, specifically as follows:

[0073] First, datasets for training and testing are collected from the sample library. These datasets include paired images (such as images of the same scene at different time points) and corresponding label information, annotations of changed regions, and labels for semantic segmentation.

[0074] The collected data undergoes preprocessing, including image scaling, cropping, and normalization, to ensure consistent data format and distribution in the input model. Additionally, the label data is appropriately transformed as needed for comparison with the model output.

[0075] Initialize the weights and parameters of the dual-attention fully convolutional Siamese neural network DASNet and the semantic segmentation network PSPNet. This typically involves setting the initial weight values ​​for the network layers, as well as selecting appropriate optimization algorithms and hyperparameters such as learning rates.

[0076] Image pairs from the training dataset are input into the model, and the network output is computed through forward propagation. This includes a change prediction map generated by DASNet and a semantic segmentation prediction map generated by PSPNet.

[0077] The loss function value is calculated based on the model output and label data. For change detection tasks, binary classification cross-entropy loss, etc., can be used; for semantic segmentation tasks, pixel-level cross-entropy loss or Dice loss, etc., can be used. Depending on the specific combination strategy, a joint loss function may need to be calculated to comprehensively consider the losses of both tasks.

[0078] The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the network weights are updated using an optimization algorithm. This process is repeated iteratively, gradually decreasing the value of the loss function to make the model output closer to the labeled data.

[0079] During training, the model is periodically evaluated using a validation set. This helps monitor model performance and allows for adjustments to hyperparameters or other measures to prevent overfitting if necessary. Model training and testing are complete once the evaluation metrics have stabilized.

[0080] In summary, the farmland non-grain conversion detection model constructed in this invention captures key information through a dual attention mechanism to obtain more discriminative feature representations. Simultaneously, it introduces the WDMC loss function to reduce the network's focus on invariant features and increase attention to changing features, thereby reducing false changes and improving the accuracy of non-grain conversion detection. Based on the algorithm model constructed in this invention, the algorithm achieves a recall rate better than 80% and an accuracy better than 70% in testing.

[0081] This invention provides support for farmland protection and can also provide technical research and scientific reference for the large-scale promotion and application of intelligent monitoring technology for "non-grain conversion", enabling timely detection and evaluation of non-grain conversion processes.

[0082] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

[0083] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A method for detecting changes in cultivated land converted to non-grain crops based on remote sensing imagery, characterized in that, Includes the following steps: Acquire remote sensing images of six types of land features at different times, including cultivated land, cultivated land converted to orchards, cultivated land converted to forest land, cultivated land converted to facility agriculture, cultivated land converted to ponds or reservoirs, and cultivated land converted to abandoned land. Perform feature analysis on the remote sensing images to form an interpretation feature database. Based on the standards of the interpretation feature library, a sample library containing positive samples of six types of land features and negative samples of non-six types of land features is constructed, and each type of sample is labeled and information augmented. Based on the dual-attention fully convolutional Siamese neural network DASNet and the semantic segmentation network PSPNet, a farmland non-grain conversion detection model is jointly constructed. The WDMC loss function is introduced and the model is trained using a constructed sample library. The image to be detected is input into the trained farmland non-grain conversion detection model to extract non-grain conversion change patches and evaluate the accuracy. The dual-attention fully convolutional Siamese neural network (DASNet) is used to extract variation information between input image pairs. It employs deep learning methods, constructing a deep network structure through a series of convolutional and pooling layers to capture high-level features in the images, specifically including: The two input images are subjected to convolution operations to extract features and generate feature maps for each image. The two input images are images of the same scene at different time points. After extracting the feature maps of the two images, DASNet combines them using a specific fusion strategy to generate a fused feature map containing change information. The specific fusion strategy includes element-wise subtraction, multiplication, or concatenation. DASNet introduces an attention mechanism to learn and emphasize regions related to change in the fused feature maps, suppresses irrelevant information, and highlights the truly changing regions. DASNet transforms the attention-weighted fused feature map into a change prediction map through one or more convolutional layers. The change prediction map is represented in binary form, with each pixel indicating whether a change has occurred at the corresponding position. The semantic segmentation network PSPNet is used to perform semantic segmentation on the input image, identify and label different objects or regions in the image, specifically including: A change prediction map is generated using DASNet, and then the change prediction map and the original image are input into PSPNet for semantic segmentation. PSPNet uses a deep convolutional neural network as its backbone network to extract deep features from the input image. It captures global contextual information through pooling operations at different scales and combines this information with the original feature map to generate an enhanced feature map containing rich contextual information. After passing through a series of convolutional layers, the enhanced feature map is converted into a semantic segmentation prediction map. The semantic segmentation prediction map is a multi-channel probability map, where each channel corresponds to a category, and the value of each pixel represents the probability that the pixel belongs to the corresponding category.

2. The method for detecting changes in cultivated land conversion to non-grain crops based on remote sensing imagery according to claim 1, characterized in that, The resolution of the acquired remote sensing images is no less than 0.5m.

3. The method for detecting changes in cultivated land conversion to non-grain crops based on remote sensing imagery according to claim 1, characterized in that, The process of performing feature analysis on remote sensing images to form an interpretation feature library includes: 1) Characteristics of arable land: Identified by referring to the reflectance spectral characteristics and texture features of the corresponding crops; 2) Characteristics of newly added orchards in cultivated land: On four-band remote sensing images, the changes from cultivated land to orchards are identified by the differences in the spectral characteristics of crops and the temporal characteristics of crops. 3) Characteristics of newly added forest land in cultivated land: On four-band remote sensing images, the Normalized Difference Vegetation Index (NDVI) is constructed using the near-infrared and red bands. Combined with texture features, forest land in cultivated land is identified. Compared with cultivated land, forest land has a coarser texture. When the newly added forest land coverage in a piece of cultivated land exceeds 40%, it is considered that the cultivated land has changed. 4) Characteristics of newly added facility agriculture in cultivated land: When newly added facility agriculture plots appear in cultivated land plots, the standard deviation of the plot spectrum, the standard deviation of the sub-object mean, and the contrast based on the gray-level co-occurrence matrix all increase significantly. The gray level of the plot changes significantly, the texture is coarser and more obvious, and the brightness increases significantly. 5) Characteristics of newly added ponds and reservoirs in cultivated land: On four-band remote sensing images, the Normalized Difference Water Index (NDWI) is constructed using the green band and near-infrared band to identify newly added ponds and reservoirs in cultivated land; the NDWI of the water surface objects is greater than 0.3, the brightness is less than 0.5, the water color is less than 0.6, and the ratio of the near-infrared band is less than 0.15; 6) Characteristics of newly abandoned farmland: Abandoned plots are identified by combining the NDVI index with the texture of land features; The interpretation feature library includes specific features and example images of six types of land features.

4. The method for detecting changes in cultivated land conversion to non-grain crops based on remote sensing imagery according to claim 1, characterized in that, The annotation of each type of sample includes: The location and boundaries of non-grain crop changes are marked in remote sensing images: First, manual interpretation and annotation are performed using ArcGIS software. After quality inspection and manual verification, the annotation results are organized and stored to provide accurate training data for subsequent non-grain crop identification and classification tasks.

5. The method for detecting changes in cultivated land conversion to non-grain crops based on remote sensing imagery according to claim 1, characterized in that, The data sources used for labeling each type of sample include remote sensing images acquired from different time periods, regions, resolutions, and sensors.

6. The method for detecting changes in cultivated land conversion to non-grain crops based on remote sensing imagery according to claim 1, characterized in that, The sample library has an average boundary accuracy better than 2 pixels, and the delineated effective range vector does not exceed the effective range of the image itself; the sample attributes are complete, including image source, number of bands, resolution, spatial information, feature description, and sample type information, supports multi-band samples, the sample space is evenly distributed, each sample covers the full frame, and there is no overlap between different samples.

7. The method for detecting changes in cultivated land conversion to non-grain crops based on remote sensing imagery according to claim 1, characterized in that, The information enhancement process includes geometric transformation, color transformation, and noise addition of the image. The selected enhancement method is used on the original sample data to generate new, enhanced samples. The labels of the enhanced samples are consistent with those of the original samples.

8. The method for detecting changes in cultivated land conversion to non-grain crops based on remote sensing imagery according to claim 1, characterized in that, The process of training the model using the constructed sample library includes: Data sets for training and testing are collected from the sample library, including pairs of images of the same scene at different time points and corresponding label information, annotations of change regions, and labels for semantic segmentation. The collected data is preprocessed, including image scaling, cropping, and normalization, to ensure that the data format and distribution of the input model are consistent. At the same time, the label data is appropriately transformed as needed for comparison with the model output. Initialize the weights and parameters of DASNet and PSPNet, including setting the initial weight values ​​of the network layers, selecting the optimization algorithm and learning rate; Image pairs from the training dataset are input into the model, and the network output is computed through forward propagation, including the change prediction map generated by DASNet and the semantic segmentation prediction map generated by PSPNet. The loss function value is calculated based on the model output and label data: for change detection tasks, binary cross-entropy loss is used; for semantic segmentation tasks, pixel-level cross-entropy loss or Dice loss is used. The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the network weights are updated using an optimization algorithm. During training, the model is evaluated periodically using a validation set until the evaluation metrics stabilize and the model training and testing are complete.

Images