Cultivated land change detection method based on multi-task dynamic detail enhancement

Abstract

The invention provides a cultivated land change detection method based on multi-task dynamic detail enhancement. The method comprises the following steps: inputting preprocessed remote sensing image data; using an encoder to extract deep features of the input image data; using an extrusion excitation type cavity space convolution pooling pyramid module to enhance features; performing up-sampling and feature fusion on the features by using a decoder; optimizing the features by using an adaptive graph channel attention module; predicting a mask image, an equal-height image and an edge image by using the optimized features; and combining prediction results of the three, optimizing by using a multi-task loss function to obtain a trained multi-task cultivated land change detection model, and accurately segmenting a cultivated land change area. According to the method, through dynamic space detail enhancement and edge feature extraction technologies and in combination with the optimized multi-task loss function, the cultivated land change is rapidly and accurately detected from the high-resolution remote sensing image, the agricultural management efficiency is improved, and the precision agricultural development is promoted.

Description

Technical Field

[0001] The present invention relates to the field of agricultural remote sensing image processing, and specifically to a cultivated land change detection method based on multi-task dynamic detail enhancement. Background Technique

[0002] With the continuous growth of the global population, the acceleration of urbanization, and the impact of climate change, cultivated land resources are facing many challenges such as decreasing quantity, deteriorating quality, and changing utilization patterns. Cultivated land, as a strategic resource for the survival and development of human society, is an important carrier of food security. Therefore, accurately detecting changes in cultivated land is of great significance for ensuring food security, optimizing land resource allocation, and coping with global changes. Cultivated land change detection is one of the core foundations of precision agriculture and land resource management, providing key data support for crop yield prediction, agricultural resource allocation, land planning, and agricultural economic planning.

[0003] However, traditional cultivated land change detection mainly relies on on-site investigation and manual visual interpretation of remote sensing images. Although this method has relatively high accuracy, it has problems such as time-consuming, laborious, and low efficiency, and it is difficult to meet the actual needs of large-scale agricultural production and rapid change monitoring. In recent years, the application of remote sensing technology in cultivated land change detection has become increasingly widespread. High-resolution remote sensing images can provide rich land use information, enabling researchers to quickly obtain high-definition images of large areas of cultivated land. However, simply relying on manual extraction of cultivated land change information from a large number of remote sensing images still faces many challenges, not only with low efficiency, but also with unstable extraction accuracy due to subjective factors.

[0004] With the rise of deep learning technology, its powerful capabilities in the field of image recognition and segmentation have brought new breakthroughs to cultivated land change detection. Convolutional neural networks (CNNs) and their variants have been widely applied to semantic segmentation tasks of remote sensing images, capable of automatically learning high-level semantic features in images, thereby realizing rapid recognition and extraction of cultivated land changes. However, existing deep learning models still have some limitations when dealing with cultivated land change detection. For example, traditional models often ignore the edge feature information of cultivated land change areas, resulting in a decline in segmentation accuracy in cases where the boundaries are blurred or the shapes are irregular. In addition, the spectral features, texture features, and seasonal variations of cultivated land vary greatly in different regions, restricting the generalization ability of the model. At the same time, existing methods are difficult to effectively model the many-to-many relationship between pixel spectra and ground object types when dealing with complex backgrounds and irregular plots, easily leading to false detections and missed detections. Summary of the Invention

[0005] In view of the above problems, the present invention proposes a cultivated land change detection method based on multi-task dynamic detail enhancement. By introducing a dynamic spatial detail enhancement module, an edge feature extraction network, and an adaptive multi-task loss function, this method significantly improves the model's ability to capture edge information of cultivated land change areas and its adaptability to complex scenarios. Compared with traditional methods, the present invention can not only efficiently process high-resolution and medium-resolution remote sensing images, but also maintain high segmentation accuracy and robustness on datasets in different regions and with different complexities. The technical solution provided by the present invention is as follows:

[0006] In a first aspect, a cultivated land change detection method based on multi-task dynamic detail enhancement, characterized by comprising the following steps: Step 1, input preprocessed remote sensing image data.

[0007] Step 2, use an encoder to extract deep features of the input image data.

[0008] Step 3, use a squeeze-and-excitation atrous spatial pyramid pooling (SE-ASSP) module at the end of the encoder to enhance the features.

[0009] Step 4, use a decoder to upsample and fuse the features.

[0010] Step 5, use an adaptive graph channel attention (CA) module at the end of the decoder to optimize the features; the adaptive graph channel attention (CA) module is implemented through three stages: feature aggregation, attention calculation, and feature restoration: First, use point convolution to aggregate the channels of the input feature map to obtain channel-compressed features; then, perform spatial aggregation on the features through grouped convolution to form spatially compressed features; then, calculate channel attention and spatial attention for the channel-compressed features and the spatially compressed features respectively; finally, restore the features to the original resolution through spatial restoration operations and channel restoration operations to obtain optimized features.

[0011] Step 6, use the optimized features to predict the mask map, elevation map, and edge map.

[0012] Step 7, combine the prediction results of the mask map, elevation map, and edge map, and use a multi-task loss function for optimization to obtain a trained multi-task cultivated land change detection model for accurately segmenting cultivated land change areas.

[0013] Preferably, the preprocessing method in Step 1 includes radiometric calibration, atmospheric correction, and orthorectification, which are used to eliminate the influence of sensor characteristics, atmospheric conditions, and terrain relief factors on the image data during acquisition.

[0014] Preferably, the encoder described in step 2 uses the VGG16 network as the backbone architecture. The VGG16 network extracts local features of the image data through multiple convolutional blocks and reduces the feature dimension through max-pooling operations.

[0015] Preferably, a Res residual module is introduced into the convolutional blocks in the VGG16 network, and shallow features are transmitted to deep layers through skip connections.

[0016] Preferably, the SE-ASSP module described in step 3 captures features X1, X2, X3, and X4 of the input feature map at different scales through four convolutional branches with different dilation rates, processes the global feature X5 through a global feature branch, and after all branches are processed, the feature maps are fused in the channel dimension to obtain the fused feature Y1; the fused feature is reweighted at the channel level, the importance weights of each channel are learned through global average pooling and a fully connected layer, and they are normalized to the interval using the Sigmoid function, and finally, a 1×1 convolutional layer is used to obtain the final feature Y2.

[0017] Preferably, the decoder described in step 4 consists of multiple modules, and the upsampling of the feature map is achieved through transposed convolutional operations to gradually restore the spatial resolution of the image; the upsampled features are fused with the skip connection features from the encoder.

[0018] Preferably, for the mask map prediction in step 6, the class labels of each pixel are generated through deep semantic features to achieve semantic segmentation of agricultural planting plots; the contour map prediction enhances the model's perception of the spatial relationship of the plots by predicting the height information of the plots; the edge map prediction generates the edge map of cultivated land changes through a specially designed edge feature extraction network to enhance the model's recognition ability of the cultivated land change boundary.

[0019] Preferably, the multi-task loss function described in step 7 is specifically: ; ; , where is the output of the model under the input x and weights W, and are the regression output and classification output of the model respectively, is 's Euclidean loss, is 's cross-entropy loss, the noise parameters and are used to balance the and relative weights of the losses of the input data, is the edge loss.

[0020] In a second aspect, a cultivated land change detection device based on multi-task dynamic detail enhancement includes: a data preprocessing module, configured to obtain high-resolution remote sensing image data and perform preprocessing to construct an agricultural planting plot extraction dataset; a model construction module, configured to build a cultivated land change detection network model with multi-task dynamic detail enhancement and perform training; a cultivated land change detection module, configured to obtain a remote sensing image to be processed, call the trained model for semantic segmentation, and output a mask map, a contour map, and an edge map; and a result analysis and application module, configured to analyze the extraction result, evaluate the segmentation accuracy, and apply the result to an agricultural resource management system.

[0021] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: By introducing a dynamic spatial detail enhancement module, an edge feature extraction network, and an adaptive multi-task loss function, the ability of the model to capture edge information of cultivated land change areas and its adaptability to complex scenarios are significantly improved. In the model training stage, only a small number of hyperparameters need to be adjusted, and the deep learning model can automatically learn the cultivated land change image features under different regions, different seasons, and different spectral characteristics, and automatically save the optimal model parameters. After training is completed, the model can be directly used for actual detection tasks to achieve real-time and efficient detection of cultivated land changes, which is convenient for users to operate and apply.

[0022] By preprocessing remote sensing images, such as operations like cropping, flipping, and brightness adjustment, the diversity of data is increased, thereby enhancing the generalization ability of the model. In the feature extraction stage, an encoder-decoder structure combined with a multi-scale feature extraction module is used to fully capture the local and global features of cultivated land change areas. By continuously training and optimizing the model, the finally obtained model can accurately identify cultivated land change areas and generate detailed mask maps, contour maps, and edge maps, providing intuitive and accurate monitoring results for agricultural management departments.

[0023] The multi-task loss function of the present invention not only considers the uncertainty between tasks but also adaptively adjusts the weights of different losses, significantly improving the performance and generalization ability of the model in multi-task semantic segmentation. In this way, the model can not only adaptively adjust the weights of different tasks but also improve its performance and generalization ability in multi-task semantic segmentation. Finally, a trained multi-task cultivated land change detection model is obtained, which can accurately segment cultivated land change areas.

[0024] In practical applications, the present invention can provide a fast and accurate cultivated land change monitoring solution for agricultural management departments, providing strong technical support for the implementation of precision agriculture, the rational utilization of land resources, and ecological environment protection. At the same time, this method can also be extended to other fields, such as urban planning, ecological environment monitoring, etc., and has broad application prospects and important social and economic values. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings are used to provide a further understanding of the present invention and form a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the accompanying drawings: Figure 1 is a schematic flow chart of the method of the present invention.

[0026] Figure 2 is a schematic diagram of the extrusion excitation type cavity space convolution pooling pyramid SE-ASSP module of the present invention.

[0027] Figure 3 is a schematic diagram of the adaptive graph channel attention CA module of the present invention.

[0028] Figure 4 is a schematic diagram of the device of the present invention. Detailed implementation manners

[0029] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0030] To make the above objects, features and effects of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0031] Embodiment 1: A cultivated land change detection method based on multi-task dynamic detail enhancement, as Figure 1 shown, includes the following steps: Step 1, first obtain the agricultural planting area image data containing the panchromatic band (0.8-meter resolution) and the multi-spectral band (3.2-meter resolution) from a high-resolution remote sensing image data source (such as the GF-2 satellite), and at the same time collect the corresponding agricultural plot annotation information for model training and verification. To ensure data quality, the remote sensing image is successively subjected to radiometric calibration, atmospheric correction and orthorectification to eliminate the errors caused by sensor characteristics, atmospheric influence and terrain undulation. Subsequently, the preprocessed image is cropped into image blocks of 256×256 pixels, and the corresponding annotation information is retained. For auxiliary data (such as digital elevation model DEM and vegetation index, etc.), it is resampled to the same spatial resolution as the remote sensing image by the bilinear interpolation method. Finally, a data set containing 4018 images is constructed and randomly divided into a training set and a verification set according to a ratio of 7:3 to support the training and verification process of the model and ensure the diversity and representativeness of the data distribution.

[0032] Step 2, in the encoder feature extraction stage, use the pre-trained VGG16 network as the backbone network, which consists of 5 convolutional blocks. Each convolutional block contains 3 convolutional layers of 3×3. The convolution operation is implemented by the following formula: , where is the output feature map of the i-th convolutional layer, Conv3×3 represents the 3×3 convolution operation, and are the convolution kernel weights and bias terms respectively, and ReLU is the activation function used to introduce non-linearity.

[0033] After each convolutional block, a 2×2 max pooling operation is performed to reduce the spatial dimension of the feature map while retaining important information.

[0034] In addition, to avoid the problem of gradient vanishing or explosion, a Res residual module is introduced in each convolutional block. The residual module directly passes the shallow features to the deep layer through a skip connection, enhancing the network's ability to capture image details. The output of the residual module is expressed as: , where is the output feature map of the residual module, is the output feature map of the convolutional layer, is the feature map of the skip connection. Through this design, the VGG16 network can effectively extract the deep features of the input image, capture rich local and global context information, and provide high-quality feature representations for subsequent feature enhancement and multi-task prediction.

[0035] Step 3, at the end of the encoder, use the squeeze-and-excitation atrous spatial pyramid pooling (SE-ASSP) module to enhance features, so as to enhance the ability to capture multi-scale features and global information. This module is as shown in Figure 2 , and contains four convolutional branches with different dilation rates , which expand the receptive field through atrous convolution to capture local features at different scales; and a global feature branch , which extracts global information through global average pooling. The feature maps of all branches are concatenated in the channel dimension to obtain the fused feature , expressed as: .

[0036] Subsequently, the SE-ASSP module performs channel-level reweighting on the fused feature Y1. It learns the importance weights of each channel through global average pooling and fully connected layers, normalizes them to the [0,1] interval using the Sigmoid function, and finally fuses and refines the reweighted features through a 1×1 convolutional layer to obtain the final feature Y2, expressed as: Y2 = Conv(Sigmoid(Conv(ReLU(Conv(AVG(Y1)))))) ⊙ Y1 Among them, Conv represents a two-dimensional 1×1 convolutional layer, AVG represents global average pooling, and ⊙ represents broadcast element multiplication.

[0037] This process not only enhances the network's ability to capture features of different scales but also highlights important features through channel reweighting, providing high-quality feature representations for subsequent tasks.

[0038] Step 4, in the decoder feature fusion stage, perform feature upsampling and fusion to gradually restore the spatial resolution of the image and enhance the detailed information of the features.

[0039] First, the input feature X is upsampled through a transposed convolution operation to restore its spatial resolution. The specific formula is: .

[0040] Among them, ConvT2×2 represents a 2×2 transposed convolution operation, and are the transposed convolution weight and bias term respectively.

[0041] Then, the upsampled feature is fused with the skip connection feature from the encoder to enhance the detailed information of the features. The fused feature is further optimized through two 3×3 convolutions and the ReLU activation function, and finally the optimized feature is obtained. The formula is: , where , and , are the weights and bias terms of the two convolutional layers respectively.

[0042] Step 5, introduce an adaptive graph channel attention CA module at the end of the decoder to further optimize the features. This module is as Figure 3 shown and is implemented through three stages: feature aggregation, attention calculation, and feature restoration. First, use point convolution to aggregate the channels of the input feature map to obtain the channel-compressed feature : , where represents the point convolution operation.

[0043] Then, perform spatial aggregation on the feature through grouped convolution to form the spatially compressed feature : , where represents the grouped convolution operation. In the attention calculation stage, perform channel attention and spatial attention calculations on the channel-compressed feature and the spatially compressed feature respectively to capture the dependencies between different channels and spatial positions.

[0044] Feature restoration is the last part of this module. In the feature restoration process, the resolution of the feature is first restored from H’×W’ to H×W through a spatial restoration operation, using 1x1 convolution and pixel rearrangement techniques, and its expression is ; subsequently, through a channel restoration operation, the number of channels is restored from C’ to C using point convolution, and its expression is represented as ; the entire feature restoration process can be expressed as: , where represents the feature after attention calculation. The finally restored feature map has the same dimensions as the original input feature map and contains enhanced global dependency information.

[0045] Step 6: Use the optimized features to predict three main tasks: mask map prediction, contour map prediction, and edge map prediction.

[0046] Mask map prediction is the core task of the network. It generates the class label of each pixel through deep semantic features to achieve semantic segmentation. Mask map prediction generates the class label of each pixel through deep semantic features to achieve semantic segmentation of agricultural planting plots. The specific formula is as follows: , where M is the predicted mask map, representing the class probability of each pixel; The mask map prediction network consists of a convolutional layer and an activation function; σ is the Sigmoid activation function, which is used to map the output value to interval, representing the probability that the pixel belongs to the target class.

[0047] Contour map prediction, as an auxiliary task, aims to learn the spatial geometric information of the cultivated land change area. By predicting the height information of the plot, it enhances the model's perception ability of the spatial relationship of the plot. The output D of the contour map prediction represents the height information of each pixel. The specific formula is as follows: , where is the contour map prediction network, which consists of a convolutional layer and a regression layer, and is used to predict the height value of each pixel.

[0048] Edge map prediction, as an auxiliary task, generates an edge map of cultivated land change through a specially designed edge feature extraction network to enhance the model's recognition ability of the cultivated land change boundary. The specific formula is as follows: , where E is the predicted edge map, representing the probability that each pixel belongs to the edge; is the edge map prediction network, which consists of a convolutional layer and an activation function; σ is the Sigmoid activation function, which is used to map the output value to interval, representing the probability that the pixel belongs to the edge.

[0049] Step 7. Traditional multi-task loss fusion methods usually linearly combine the losses of each task, but it is very difficult to set appropriate weights for different tasks and often requires trial and error, which is time-consuming and inefficient. Therefore, the present invention proposes a multi-task loss function that adaptively adjusts the weights of the losses of each task, and realizes weighting by modeling the uncertainty of task dependence (homogeneous uncertainty), which is constant for the input data but varies among different tasks.

[0050] The present invention formulates the uncertainty of task dependence as the probability of the network output. Let be the output of the network under the input x and the weight W, be the model likelihood. For regression tasks, the model likelihood is defined as a Gaussian distribution with a mean of the model output and a variance of the observation noise σ; for classification tasks, the Softmax function is used to compress the output into probabilities, following the Boltzmann distribution. If the model has a regression output and a classification output , the joint probability is: .

[0051] Convert the inference of the joint probability into a minimization objective, and define the loss function as: ; is the Euclidean loss of is the cross-entropy loss of . c is the entity of the vector and . Consider the noise parameters and as the relative weights for balancing the (or ) and (or ) losses of the given input data. Increasing the value of will cause the weight of and to decrease. In addition, is used to replace as a regularization term to prevent the noise parameters

[0052] from increasing too much during model training. During the training process, is added to the uncertainty-weighted mask and distance loss for edge prediction tasks because directly using uncertainty weighting may lead to too small weights due to the additional noise introduced by fusing multi-level edge cues. Thus, the total multi-task loss function is obtained: . Combine the prediction results of the mask map, contour map, and edge map, and use the multi-task loss function for optimization to obtain a trained multi-task cultivated land change detection model for precise segmentation of cultivated land change areas.

[0053] The multi-task loss function of the present invention not only considers the uncertainty between tasks, but also adaptively adjusts the weights of different losses, significantly improving the performance and generalization ability of the model in multi-task semantic segmentation. In this way, the model can not only adaptively adjust the weights of different tasks, but also improve its performance and generalization ability in multi-task semantic segmentation. Finally, a trained multi-task cultivated land change detection model is obtained, which can accurately segment the cultivated land change area.

[0054] Embodiment 2: A cultivated land change detection method based on multi-task dynamic detail enhancement, as Figure 4 shown, the device can be deployed on edge devices or in the cloud to efficiently and accurately detect cultivated land changes from remote sensing images. The specific modules and functions of the device are as follows: a data preprocessing module, which is responsible for obtaining high-resolution remote sensing image data (such as GF-2 satellite images, panchromatic band with a resolution of 0.8 meters, and multi-spectral band with a resolution of 3.2 meters), and performing preprocessing. The preprocessing steps include radiometric calibration, atmospheric correction, orthorectification, etc., to eliminate the influence of sensor characteristics, atmospheric conditions, and terrain undulation on the images. The preprocessed images are cropped into image patches of 256×256 pixels, and corresponding agricultural plot annotation information is generated to construct an agricultural planting plot extraction dataset.

[0055] A model construction module, which is used to construct a cultivated land change detection model based on a multi-task semantic segmentation network. The model includes key components such as an encoder, a decoder, an SE-ASSP module, and a CA module.

[0056] A cultivated land change detection module, which analyzes the extraction results, evaluates the segmentation accuracy, and applies the extracted cultivated land change information to fields such as agricultural resource management, crop yield prediction, and land planning, providing data support for precision agriculture. The module provides a visualization interface to display the extraction results and analysis reports.

[0057] A result analysis and application module, which analyzes the extraction results, evaluates the segmentation accuracy, and applies the extracted agricultural planting plot information to fields such as agricultural resource management, crop yield prediction, and land planning, providing data support for precision agriculture. The module provides a visualization interface to display the extraction results and analysis reports.

[0058] Embodiment 3: The computer-readable storage medium of the present embodiment stores a computer program, which when executed by a processor implements the steps in a cultivated land change detection method based on multi-task dynamic detail enhancement in Embodiment 1.

[0059] The computer-readable storage medium of this embodiment may be an internal storage unit of the terminal, such as the hard disk or memory of the terminal; the computer-readable storage medium of this embodiment may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card, a secure digital card, a flash card, etc. equipped on the terminal; further, the computer-readable storage medium may also include both the internal storage unit and the external storage device of the terminal.

[0060] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium may also be used to temporarily store data that has been output or will be output.

[0061] Embodiment 4: The computer device of this embodiment includes a processor, a memory, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in a method for detecting cultivated land changes based on multi-task dynamic detail enhancement in Embodiment 1.

[0062] In this embodiment, the processor may be a central processing unit, or may also be other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays, or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor, or the processor may also be any conventional processor, etc.; the memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A part of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0063] Those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0064] Example 5: A computer program product for detecting cultivated land changes based on multi-task dynamic detail enhancement. This program product includes a complete set of cultivated land change detection processes, covering four core functions: data preprocessing, model training, cultivated land change extraction, and result analysis. Users can customize key parameters such as input image resolution and model hyperparameters according to actual needs to adapt to cultivated land change detection tasks in different scenarios. The program adopts a modular design, and users can flexibly select and execute specific modules, such as performing data preprocessing or cultivated land change extraction operations separately. At the same time, the program is equipped with a detailed logging function to facilitate real-time tracking of the running status and result output, ensuring the transparency and traceability of operations. In terms of application scenarios, this program product is widely applicable to various user groups such as agricultural research institutions, agricultural management departments, and enterprises engaged in precision agriculture. Users can quickly deploy a cultivated land change detection system on various hardware platforms such as personal computers, servers, or network devices with the help of this program to achieve efficient and accurate cultivated land change detection tasks. To help users better use this program, the product also provides a detailed user manual and online help documents. These materials introduce the various functions and operation steps of the program in an easy-to-understand way, and even beginners can quickly get started and master the usage method of the program, so as to give full play to the technical advantages and application value of this computer program product in the field of cultivated land change detection and provide strong technical support for promoting agricultural modernization and the development of precision agriculture.

[0065] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A cultivated land change detection method based on multi-task dynamic detail enhancement, characterized in that It includes the following steps: Step 1, input the preprocessed remote sensing image data; Step 2, use an encoder to extract the deep features of the input image data; Step 3, use a squeeze-and-excitation atrous spatial pyramid pooling (SE-ASSP) module at the end of the encoder to enhance the features; Step 4, use a decoder to upsample and fuse the features; Step 5, use an adaptive graph channel attention (CA) module at the end of the decoder to optimize the features; The adaptive graph channel attention (CA) module is implemented through three stages: feature aggregation, attention calculation, and feature restoration. First, use point convolution to aggregate the channels of the input feature map to obtain channel-compressed features; then, perform spatial aggregation on the features through grouped convolution to form spatially compressed features; Then, perform channel attention and spatial attention calculations on the channel-compressed features and spatially compressed features respectively; finally, restore the features to the original resolution through spatial restoration operations and channel restoration operations to obtain optimized features; Step 6, use the optimized features to predict the mask map, contour map, and edge map; Step 7, combine the prediction results of the mask map, contour map, and edge map, and use a multi-task loss function for optimization to obtain a trained multi-task cultivated land change detection model, and accurately segment the cultivated land change area.

2. The cultivated land change detection method based on multi-task dynamic detail enhancement according to claim 1, wherein The preprocessing method in Step 1 includes radiometric calibration, atmospheric correction, and orthorectification, which are used to eliminate the influence of sensor characteristics, atmospheric conditions, and terrain relief factors on the image data during acquisition.

3. The cultivated land change detection method based on multi-task dynamic detail enhancement according to claim 1, characterized in that The encoder described in Step 2 uses the VGG16 network as the backbone architecture. The VGG16 network extracts local features of the image data through multiple convolutional blocks and reduces the feature dimension through max-pooling operations.

4. A cultivated land change detection method based on multi-task dynamic detail enhancement according to claim 3, characterized in that A Residual (Res) module is introduced into the convolutional blocks in the VGG16 network, and the shallow features are passed to the deep layer through skip connections.

5. The cultivated land change detection method based on multi-task dynamic detail enhancement according to claim 4, wherein In step 3, the SE-ASSP module captures features X1, X2, X3, and X4 of the input feature map at different scales through four convolutional branches with different dilation rates, processes the global feature X5 through a global feature branch. After all branches are processed, the feature maps are fused in the channel dimension to obtain the fused feature Y1; For the fused feature perform channel-level reweighting, learn the importance weights of each channel through global average pooling and a fully connected layer, and normalize them to the interval, and finally obtain the final feature Y2 through a 1×1 convolutional layer.

6. A cultivated land change detection method based on multi-task dynamic detail enhancement according to claim 4, characterized in that, The decoder described in Step 4 consists of multiple modules, and realizes the upsampling of the feature map through transposed convolution operations, gradually restoring the spatial resolution of the image; fuses the upsampled features with the skip connection features from the encoder.

7. A cultivated land change detection method based on multi-task dynamic detail enhancement according to claim 1, characterized in that, For the mask map prediction in Step 6, generate the class label of each pixel through deep semantic features to achieve semantic segmentation of agricultural planting plots; contour map prediction enhances the model's perception ability of the spatial relationship of the plots by predicting the height information of the plots; edge map prediction generates an edge map of cultivated land change through a specially designed edge feature extraction network to enhance the model's recognition ability of the cultivated land change boundary.

8. A cultivated land change detection method based on multi-task dynamic detail enhancement according to claim 7, characterized in that, The multi-task loss function in step 7 is specifically: ; ; ,in is the output of the model under input x and weight W, and are the regression output and classification output of the model respectively, yes The Euclidean loss, yes The cross entropy loss, noise parameter and To balance the input data and The relative weight of the loss, is the edge loss.

9. A cultivated land change detection device based on multi-task dynamic detail enhancement is used to implement a cultivated land change detection method based on multi-task dynamic detail enhancement as described in any one of claims 1-8, and is characterized in that, It includes: A data preprocessing module, which is used to obtain high-resolution remote sensing image data and perform preprocessing to construct a dataset for agricultural planting plot extraction; A model construction module, which is used to build a cultivated land change detection network model with multi-task dynamic detail enhancement and perform training; a cultivated land change detection module, which is used to obtain the remote sensing image to be processed, call the trained model for semantic segmentation, and output the mask map, contour map, and edge map; a result analysis and application module, which is used to analyze the extraction results, evaluate the segmentation accuracy, and apply the results to the agricultural resource management system.

Images