An end-to-end extraction method for natural field type remote sensing images based on multiple attention channels

Abstract

The application discloses an end-to-end extraction method for natural field type remote sensing images based on multiple attention channels. The application comprises the following steps: S1. inputting a natural field type remote sensing image to generate a preliminary feature map through a backbone network; S2. sequentially inputting the preliminary feature map into at least two residual learning stages of the backbone network to perform deep feature extraction, finally generating a high-dimensional feature map, and reserving a first feature map output by the first residual learning stage; S3. inputting the high-dimensional feature map into an ASPP (Atrous Spatial Pyramid Pooling) module to generate a pooled feature map; S4. splicing the first feature map weighted by a first channel attention module and the pooled feature map weighted by a second channel attention module along a channel dimension to form a combined feature map; and S5. inputting the combined feature map into a decoder module containing a third channel attention module to generate a final refined segmentation map. The application is applied to the technical field of remote sensing images.

Description

Technical Field

[0001] This invention relates to a method for segmenting and extracting natural field patterns from remote sensing images, and particularly to an end-to-end extraction method for natural field pattern remote sensing images based on multiple attention channels. Background Technology

[0002] Natural field topography, as a crucial foundational data source for agriculture, plays an indispensable role in accurately extracting ridge data for monitoring crop growth, predicting crop yield, and managing agricultural risks, thus forming the basis of intelligent agriculture. Research on ridge data extraction mainly includes edge detection, traditional machine learning, and deep learning-based image segmentation methods.

[0003] Early research methods primarily focused on edge detection and traditional machine learning. Edge detection methods directly extract field ridge boundaries by calculating pixel differences using edge detection operators such as Sobel and Canny. However, edge detection operators are only suitable for situations where the contrast between the boundary and the background is significant. When the natural field boundaries in the image are blurred, edge detection methods struggle to obtain closed extraction results and are prone to generating redundant boundaries. Traditional machine learning methods, such as support vector machines, maximum likelihood estimation, and decision trees, perform field detection. They extract field ridge data from remote sensing images based on pixel aggregation and rely on manually selecting features and designing classifiers for classification. With appropriate feature selection, they perform well and are easy to understand and implement for simple classification tasks. However, they are often limited by predefined rules and assumptions. Faced with the high resolution and complex backgrounds of high-resolution remote sensing images, the extraction accuracy often falls short of practical applications.

[0004] In recent years, research on natural field shape detection has mainly focused on high-precision extraction driven by deep learning. Driven by the development of deep learning technology, convolutional neural networks have been widely applied to land cover classification tasks in remote sensing imagery. Models such as UNet, DeepLabV3+, and Segformer have been introduced into field extraction, achieving good results. For example, some researchers have applied the DeepLabv3+ model to semantic segmentation tasks, accurately extracting sugarcane fields in sugarcane-growing areas of Wuming City, Guangxi Province, using GF-2 and BJ-2 satellite imagery. Other researchers have used Worldview-3 satellite imagery, based on three different fully convolutional neural network models, to extract farmland in southern Bangladesh. Deep learning methods provide strong support for agricultural monitoring and land management through automated feature learning and efficient data processing. However, due to the complexity and heterogeneity of the feature space caused by high image resolution, single-task deep learning methods have low classification accuracy in extracting field edges and corners. The extracted field shapes differ from the actual shapes, focusing only on pixel classification within the field ridges while neglecting spatial relationships between ridges and morphological variations in the fields. To address the aforementioned issues, some researchers have combined edge detection technology with deep learning algorithms, utilizing ridge boundary data to construct cascaded multi-task networks and incorporating spatial topological information between ridges to extract data from farmland in Denmark and Chongqing. Other researchers have developed multi-task neural network models tailored to the characteristics of hilly farmland in southern China, extracting data from farmland in Pucheng County, Fujian Province, using ridge boundary data, convolutional neural networks, and attention modules. While multi-task neural network models have improved the accuracy of natural field shape detection, they still suffer from insufficient detail in edge detection and poor consistency in boundary delineation in complex natural scenes.

[0005] In summary, deep learning-based semantic segmentation technology can efficiently identify and extract key features in images, significantly improving segmentation accuracy. However, it still suffers from insufficient detail in detecting field-shaped edges in complex natural scenes and poor continuity in boundary delineation. Furthermore, it does not provide a complete end-to-end remote sensing result generation scheme. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an end-to-end extraction method for natural field patterns in remote sensing images based on multiple attention channels, which can extract natural field patterns in remote sensing images more intelligently, accurately and robustly.

[0007] The technical solution adopted in this invention is as follows: This invention includes the following steps:

[0008] Step 1. Model initialization: Construct the DeepLabV3PlusCA model, select DeepLabV3_ResNet50 as the backbone network, input a natural field-shaped remote sensing image, and perform preliminary feature extraction on the remote sensing image through the pre-trained backbone network to generate a preliminary feature map;

[0009] Step 2. Deep feature extraction and key layer feature preservation: The initial feature map is passed through at least two residual learning stages of the backbone network for deep feature extraction. Finally, a high-dimensional feature map is generated, and the first feature map output by the first residual learning stage is preserved.

[0010] Step 3. Hollow Spatial Pyramid Pooling Processing: Input the high-dimensional feature map into the hollow spatial pyramid pooling ASPP module to generate a pooled feature map. Then, use bilinear interpolation to upsample the pooled feature map to a size that is consistent with the spatial size of the first feature map.

[0011] Step 4. Feature concatenation: Apply the first channel attention module to the first feature map for weighted adjustment, apply the pooled feature map processed by bilinear interpolation to the second channel attention module for weighted adjustment, and concatenate the weighted first feature map and the weighted pooled feature map along the channel dimension to form a combined feature map;

[0012] Step 5. Segmentation map generation: The combined feature map is input into the decoder module. Inside the decoder module, a third channel attention module is applied to further focus on the channel features related to the natural field pattern and enhance the edge detail modeling capability. Then, the preliminary segmentation result is upsampled by bilinear interpolation to restore the resolution to the original input remote sensing image and generate the final refined segmentation map.

[0013] Furthermore, the extraction method also includes step 6. Loading and preprocessing training remote sensing image data and its corresponding label image data, inputting the preprocessed training remote sensing image data and its corresponding label image data into the DeepLabV3PlusCA model, extracting features through the backbone network, and adjusting the feature weights using the first channel attention module, the second channel attention module, and the third channel attention module to perform data augmentation on the training remote sensing image data. Data augmentation includes random flipping, random rotation, and random scaling operations to increase data diversity and improve the model's generalization ability. The loss value is calculated using a combined loss function, the model parameters are updated using the Adam optimizer, and finally the model performance is evaluated, performance indicators are recorded, and the learning curve is output. If the preset training rounds or convergence criteria are reached, the model is saved and training ends; otherwise, iterative training continues.

[0014] Furthermore, in step 6, the enhanced natural field-shaped remote sensing image training data is processed using a combination loss function L. totalTo optimize the model and calculate the loss value, the combined loss function L... total The binary cross-entropy loss L BCE Dice loss L Dice and gradient loss L Grad Composed of three weighted parts:

[0015] L total =a1*L BCE +a2*L Dice +a3*L Grad

[0016] Where a1, a2, and a3 are the weighting coefficients for each loss;

[0017] The binary cross-entropy loss L BCE The formula used to measure the pixel-by-pixel classification accuracy between the predicted and target values ​​is as follows:

[0018]

[0019] Where N is the total number of pixels, y i It is the true label of pixel i, p i It is the probability that the model predicts pixel i to be a natural field shape;

[0020] The Dice loss L Dice Based on the intersection-union ratio (IUGR) calculation, this study focuses on the degree of overlap between the predicted and actual regions to optimize segmentation similarity. The calculation formula is as follows:

[0021]

[0022] in, It is a constant, N is the total number of pixels, and y is... i It is the true label of pixel i, p i It is the probability that the model predicts pixel i to be a natural field shape;

[0023] The gradient loss L Grad The gradient difference between the predicted image and the target image in the horizontal (x) and vertical (y) directions is calculated to ensure spatial continuity and edge smoothness of the output; the calculation formula is as follows:

[0024]

[0025] in and These represent the gradient operators in the x and y directions, respectively, where N is the total number of pixels, and y is the gradient value. i It is the true label of pixel i, p i It is the probability that the model predicts pixel i to be a natural field shape.

[0026] Furthermore, the extraction method also includes step 7. Block sliding window prediction: For large remote sensing images, a block reading and sliding window prediction strategy is adopted to avoid memory overflow while preserving the image's georeferenced information.

[0027] Furthermore, the original input natural field-shaped remote sensing image in step 1 has a size of [3, H, W], where 3 represents 3 channels, H represents height, and W represents width. The remote sensing image undergoes preliminary feature extraction through a pre-trained backbone network. Specifically, the input remote sensing image is processed sequentially through the initial convolutional layer, batch normalization layer, ReLU activation function, and a max pooling layer of the backbone network to generate a preliminary feature map with a size of [64, H / 4, W / 4].

[0028] Furthermore, in step 2, there are four residual learning stages: layer1, layer2, layer3, and layer4. The initial feature map is downsampled step by step in the four residual learning stages to extract deeper features. After layer4, a high-dimensional feature map with a size of [2048, H / 16, W / 16] is generated. At the same time, the first feature map with a size of [256, H / 4, W / 4] after processing by layer1 is retained because the first feature map contains more low-level spatial detail information and is used for subsequent feature fusion.

[0029] Furthermore, in step 3, a high-dimensional feature map of size [2048, H / 16, W / 16] is input into the Aperture Spatial Pyramid Pooling (ASPP) module. The Aperture Spatial Pyramid Pooling (ASPP) module can effectively capture multi-scale contextual information without increasing the number of parameters and computational cost, thereby enhancing the model's ability to perceive targets at different scales. The pooled feature map is upsampled to a size of [256, H / 4, W / 4] using bilinear interpolation, thus maintaining consistency with the spatial size of the first feature map.

[0030] Furthermore, step 4 specifically involves applying the first channel attention module to the first feature map [256, H / 4, W / 4] output by layer 1. Through adaptive average pooling, a dimensionality-reduced fully connected layer, ReLU activation function, and Sigmoid activation function, the various channels of the feature map are weighted and adjusted to enhance key channels containing fine structural information such as natural field-shaped edges. The pooled feature map of size [256, H / 4, W / 4] is then applied to the second channel attention module. Using the same mechanism as the first channel attention module, the channels of the pooled feature map are weighted to optimize the model's attention to natural field-shaped regions and reduce background noise interference. Finally, the weighted first feature map and the weighted pooled feature map are concatenated along the channel dimension to form a combined feature map of size [512, H / 4, W / 4].

[0031] Furthermore, the decoder module in step 5 includes a convolutional layer, a batch normalization layer, and a ReLU activation layer; the combined feature map of size [512, H / 4, W / 4] is processed by the decoder containing a third channel attention module, and then the number of channels of the combined feature map is reduced to 1 by a 1x1 convolutional layer to generate a preliminary extraction result of size [1, H / 4, W / 4]. Finally, the preliminary segmentation result is upsampled by 4x bilinear interpolation to restore it to the original input image resolution [1, H, W], generating the final refined segmentation map.

[0032] Furthermore, let the large remote sensing image in step 7 be I, with a size of H*W. The image is segmented into 256x256 pixel tiles using a sliding window method, matching the input size of the DeepLabV3PlusCA model. The sliding window step size is s. Assume image I is segmented into n tiles, and the i-th tile is represented as I... i i = 1, 2, ..., n; For each tile I i Input the trained model M for inference and generate the corresponding binary segmentation mask M. i That is, M i =M(I i During the splicing process, the average method is used to fuse the predicted values ​​for the overlapping areas in the sliding window process. Let I be two adjacent tiles. j and I j+1 The overlapping region is O, and O is in I j The predicted value in part is P j，0 , in I j+1 The predicted value in part is P j+1，0 Then the fused predicted value P0 is:

[0033]

[0034] Finally, all the processed segmentation masks are concatenated to form a complete raster extraction result R, and a vector file is output.

[0035] The beneficial effects of this invention are:

[0036] 1. Improved the targeting and effectiveness of feature extraction:

[0037] By introducing the first, second, and third channel attention modules at different stages, the model can adaptively learn and enhance its attention to key feature channels related to natural field shapes. For example, the first channel attention module enhances channels containing fine structural information such as field edges, the second channel attention module optimizes attention to field regions and reduces background noise, and the third channel attention module further strengthens edge detail modeling during the decoding stage. This enables the model to capture and distinguish target features more accurately. In addition, retaining and utilizing the first channel attention module enhances the low-level spatial detail information of ResNet's layer 1 output, which is crucial for accurately segmenting field edges and identifying small fields.

[0038] 2. The quality of the extraction results has been optimized:

[0039] A combined loss function L is adopted total Multiple factors were taken into account, especially the gradient loss L. Grad The introduction of this feature aims to ensure spatial continuity and edge smoothness of the output, which is particularly important for extracting objects with complex boundaries, such as natural field shapes, and helps to generate smoother, more natural segmentation boundaries.

[0040] 3. Enhanced the model's generalization ability and practicality:

[0041] This paper provides a complete solution from model design and training to large-scale remote sensing image block prediction and vectorization output, which improves the practicality and automation of the method. By adopting block sliding window prediction, overlapping area average fusion and preservation of georeferenced information, it effectively solves the memory limitation problem in large-scale remote sensing image processing, and ensures the smooth transition of the stitching results and the accuracy of the geographical location, so that the method can be applied to practical large-scale remote sensing monitoring tasks.

[0042] 4. Optimizations specifically for natural field types:

[0043] The entire methodology, from the design of the attention mechanism (focusing on field edges and regions), the selection of the loss function (emphasizing edge smoothness and regional similarity), to the construction of the specific model, reflects the optimization of the specific task of extracting natural field shapes, aiming to improve the accuracy and effectiveness of extraction. Attached Figure Description

[0044] Figure 1 This is the overall flow of an end-to-end extraction method for natural field-shaped remote sensing images based on multiple attention channels provided in one embodiment of this application;

[0045] Figure 2 It is the network architecture of the DeepLabV3PlusCA model;

[0046] Figure 3 This is an example of the loss function variation of a training model provided in one embodiment of this application;

[0047] Figure 4 This is a model iterative training process provided in one embodiment of this application;

[0048] Figure 5 This is a comparison of the natural field pattern extraction results provided in some embodiments of this application;

[0049] Figure 6 This is the original remote sensing image provided in one embodiment of this application;

[0050] Figure 7 This is an end-to-end prediction result provided in one embodiment of this application;

[0051] Figure 8 This is a partial magnification of the extraction results provided in some embodiments of this application, superimposed with the original image; Detailed Implementation

[0052] like Figure 1 As shown, in this embodiment, the present invention includes the following steps:

[0053] Step 1: Model Initialization:

[0054] A deep learning segmentation model, DeepLabV3PlusCA, is constructed based on the DeepLabV3+ architecture and incorporates a channel attention mechanism. DeepLabV3_ResNet50 is selected as the backbone network. The network architecture of the DeepLabV3PlusCA model is as follows: Figure 2 As shown. The input natural field-shaped remote sensing image accepted by the model is a three-channel remote sensing image sample with a size of [3, 256, 256], that is, 3 channels, and a height and width of 256. Initial feature extraction is performed by passing the input image through the initial convolutional layer (7x7 convolution, stride 2), batch normalization layer, ReLU activation function, and a max pooling layer of the ResNet50 backbone network. After these operations, a preliminary feature map is generated with a size of [64, 64, 64], that is, the number of channels becomes 64, and the spatial resolution is reduced to 1 / 4 of the original image.

[0055] Step 2: Deep Feature Extraction and Key Layer Feature Preservation

[0056] The initial feature map [64,64,64] generated in step 1 then passes through the four main residual learning stages of the backbone network, namely layer1, layer2, layer3 and layer4.

[0057] After layer 1, the feature map size is [256, 64, 64].

[0058] After layer 2, the feature map size is [512, 32, 32].

[0059] After layer 3, the feature map size is [1024, 16, 16].

[0060] After layer 4, a high-dimensional feature map is generated with a size of [2048, 16, 16]. These features contain rich semantic information, but some spatial details are lost.

[0061] To facilitate subsequent fusion of detailed information, the first feature map output by layer 1 of the backbone network is specifically retained. This first feature map has a size of [256, 64, 64]. Due to its low downsampling factor, it retains more low-level spatial details, such as edges and textures, which is crucial for accurately segmenting the boundaries of natural field shapes.

[0062] Step 3: Pyramid pooling treatment of void spaces:

[0063] The high-dimensional feature map [2048,16,16] generated in step 2 is used as the input to the Spatial Pyramid Pooling (ASPP) module with dilated spatial pooling. By using convolutional kernels with different dilation rates, the ASPP module can effectively capture contextual information under different receptive fields without increasing the number of parameters and computational cost, thereby enhancing the model's ability to perceive targets at different scales.

[0064] The feature maps output by each parallel ASPP branch are concatenated along the channel dimension, and then a 1x1 convolutional layer is used for channel dimensionality reduction, for example, reducing the number of channels to a uniform 256. At this time, the pooled feature map size output by the dilated spatial pyramid pooling ASPP module is [256, 16, 16].

[0065] In order to match the size of the first feature map retained in step 2 for subsequent fusion, the pooling feature map is upsampled by 4 times using bilinear interpolation to restore its spatial size to [256, 64, 64].

[0066] Step 4: Feature stitching:

[0067] The first feature map with dimensions [256, 64, 64] retained in step 2 is input into the first channel attention module. Global adaptive average pooling is performed on the first feature map to obtain the global spatial information of each channel, and the dimensions become [256, 1, 1]. Then, the non-linear dependencies between channels are learned through two fully connected layers and a ReLU activation function. Finally, the weights of each channel are generated through a Sigmoid activation function. The channel weights generated by the first channel attention module are multiplied by the original first feature map channel by channel, thereby assigning higher weights to key channels containing fine structural information such as natural field-shaped edges and textures.

[0068] The pooled feature map, whose spatial dimensions are restored to [256, 64, 64] in step 3, is input into the second channel attention module. The same channel attention mechanism as the first channel attention module is used to weight the pooled feature map by channel. This aims to optimize the model's attention to channels related to the natural field-shaped region within multi-scale contextual information, suppressing interference from background noise or irrelevant information.

[0069] The first feature map, weighted by the first channel attention module, and the pooled feature map, weighted by the second channel attention module, are concatenated along the channel dimension. Since both have dimensions of [256, 64, 64], the size of the concatenated combined feature map becomes [512, 64, 64]. This combined feature map possesses both rich low-level spatial details and high-level semantic context information, and is optimized through the attention mechanism.

[0070] Step 5: Generation of segmentation map:

[0071] The combined feature map [512,64,64] generated in step 4 is input into the decoder module. The combined feature map is processed by convolutional layers, batch normalization layers, ReLU activation layers, channel attention layers, and finally a 1x1 convolution in the decoder to generate an output [1,64,64]. Finally, it is restored to a refined segmentation map with the original resolution of [1,256,256] through 4x bilinear upsampling, thus completing the segmentation task.

[0072] Steps 1 through 5 above complete the model construction.

[0073] Step 6: Model training iteration:

[0074] Pixel-level ground truth masks containing natural field patterns and their corresponding labeled images were collected. After cropping and deleting invalid samples, 1270 256x256 samples were obtained. These were then divided into training and testing sets at an 8:2 ratio. Data augmentation operations such as rotation, flipping, and deformation were performed to expand the sample size to 8890 256x256 samples. Finally, the training set contained 7712 samples, and the validation set contained 1778 samples.

[0075] The preprocessed and enhanced training remote sensing image data and their corresponding labeled images are input into the DeepLabV3PlusCA model. The model performs forward computation, extracts features through the backbone network, adjusts the feature weights using the first, second, and third channel attention modules, and finally generates a refined segmentation map.

[0076] To further optimize the DeepLabV3PlusCA model, a combined loss function L is designed. total To optimize the model, the loss function consists of binary cross-entropy loss L. BCE Dice loss L Dice and gradient loss L Grad Composed of three weighted parts:

[0077] L total =a1*L BCE +a2*L Dice +a3*L Grad

[0078] Where a1, a2, and a3 are the weighting coefficients for each loss term, adjusted based on experience or experimentation. In this embodiment, they are set to a1=1.0, a2=0.8, and a3=0.5. These weights balance the contribution of various losses to model optimization.

[0079] L BCE This measures the difference between the model's pixel-level classification predictions and the true labels. For each pixel, the cross-entropy between the probability of its prediction as a natural field shape and the true label (0 or 1) is calculated.

[0080]

[0081] Where N is the total number of pixels, y i It is the true label of pixel i, p i It is the probability that the model predicts pixel i to be a natural field shape.

[0082] L Dice By focusing on the degree of overlap between the predicted segmentation region and the actual segmentation region, the class imbalance problem can be effectively handled, and the overall similarity of the segmentation results can be optimized.

[0083]

[0084] in, It is a very small constant, N is the total number of pixels, and y is... i It is the true label of pixel i, p i It is the probability that the model predicts pixel i to be a natural field shape.

[0085] L Grad Calculate the L1 or L2 norm of the gradient difference between the predicted segmentation map and the ground truth label map in the horizontal (x) and vertical (y) directions. This loss term is designed to penalize blurry or discontinuous edges in the prediction results, ensuring that the output results have good spatial continuity and smooth, clear edges.

[0086]

[0087] in and These represent the gradient operators in the x and y directions, respectively, where N is the total number of pixels, and y is the gradient value. i It is the true label of pixel i, p i It is the probability that the model predicts pixel i to be a natural field shape.

[0088] Based on the combination loss function L total Calculate the loss value between the model's predicted result and the true label. The loss function changes as follows during model training: Figure 3 As shown, the Adam optimizer is used to update the model's network parameters using the backpropagation algorithm based on the calculated loss value. The initial learning rate is set to 10⁻⁴, the number of training epochs is 500, and the batch size is 16. Model performance is evaluated on the validation set after each training epoch or after a fixed number of iterations. Commonly used evaluation metrics include intersection-over-union ratio (IoU), Dice coefficient, pixel accuracy, and F1 score.

[0089] If the model performance reaches the preset convergence criterion on the validation set (e.g., performance no longer significantly improves after several consecutive epochs), or reaches the preset maximum number of training epochs, then the current optimal model weights are saved, and training ends. Otherwise, the next iteration of training continues. The model training iteration process is as follows: Figure 4 As shown.

[0090] Step 7: Segmented sliding window prediction:

[0091] Single image inference comparison results are as follows Figure 5 As shown, where Figure 5 (a) represents the original remote sensing image. Figure 5 (b) represents the label image corresponding to the original image. Figure 5(c) represents the predicted acquisition effect map after applying the method of the present invention, i.e., the refined segmentation map. Figure 5 (d) shows the prediction and acquisition results after using the traditional DeepLabV3+ model method. Figure 5 (e) shows the predicted acquisition result after using the traditional UNet model method. For large TIFF format remote sensing images, which are usually much larger than the model input size, directly loading the entire image for prediction will lead to memory overflow. Therefore, a block-based reading and sliding window prediction strategy is adopted. The large TIFF remote sensing image I (size H×W) is divided into blocks with the model input size (e.g., 256x256 pixels) as the tile size. The original remote sensing image is visualized as follows. Figure 6 As shown. Simultaneously, the georeferenced information of the image must be preserved. The image is divided into n tiles I using a sliding window method. i (i=1,2,…,n), when sliding the window, set a step size s of 128 pixels. Each tile I... i The input is fed into the DeepLabV3PlusCA model M trained in step 7 for inference, generating the corresponding binary segmentation mask M. i =M(I i ), in the predicted result M of splicing each tile i At this time, the overlapping areas generated by the sliding window are merged to obtain a smoother and more consistent overall segmentation result. Let two adjacent tiles I... j and I j+1 The overlapping region is O, and O is in I j The predicted value in part is P j，0 , in I j+1 The predicted value in part is P j+1，0 Then the fused predicted value P0 is:

[0092]

[0093] The segmentation mask M after all processing (including merging overlapping regions) i The raster images are precisely stitched together according to their geographical locations in the original large image to form a complete raster extraction result image R. Finally, the raster extraction result R is converted into a vector polygon file, and the end-to-end prediction result is as follows: Figure 7 As shown. To demonstrate the method of this invention, the result is partially magnified and overlaid with a remote sensing image base map, the effect of which is as follows. Figure 8 As shown.

[0094] In summary, the core innovations of this invention include: First, an improved DeepLabV3PlusCA model architecture was constructed. Based on the ResNet50 backbone network, this model significantly enhances the ability to accurately capture and efficiently fuse fine edges and complex contextual features in remote sensing images by strategically introducing three channel attention modules into key nodes of the backbone network and within the decoder. Second, a binary cross-entropy loss L... BCE Dice loss L Dice and gradient loss L Grad The combined loss function L total The model's classification accuracy, regional prediction overlap, and boundary contour smoothness and continuity were optimized collaboratively, ensuring high-precision contour extraction. To address the challenge of end-to-end inference for large-scale remote sensing images, a large-image prediction and vectorization output scheme based on block sliding windows was further proposed. This scheme, combined with a deep learning model, can efficiently process wide-swath remote sensing images and directly output the extraction results in a vector format that is easy to integrate into geographic information systems.

[0095] This invention applies to the field of remote sensing image processing technology.

[0096] Although the embodiments of the present invention are described with reference to actual solutions, they do not constitute a limitation on the meaning of the present invention. Modifications to the embodiments and combinations with other solutions based on this specification will be obvious to those skilled in the art.

Claims

1. An end-to-end extraction method for natural field-shaped remote sensing images based on multi-attention channels, characterized in that, Includes the following steps: Step 1. Construct a DeepLabV3PlusCA model, input a natural field-shaped remote sensing image, and perform preliminary feature extraction through a pre-trained backbone network to generate a preliminary feature map; Step 2. The initial feature map is passed through at least two residual learning stages of the backbone network for deep feature extraction, and finally a high-dimensional feature map is generated, while the first feature map output by the first residual learning stage is retained. Step 3. Input the high-dimensional feature map into the hole spatial pyramid pooling ASPP module to generate a pooled feature map. Then, use bilinear interpolation to upsample the pooled feature map to a size that is consistent with the spatial size of the first feature map. Step 4. Apply the first channel attention module to the first feature map for weighted adjustment, apply the second channel attention module to the pooled feature map processed by bilinear interpolation for weighted adjustment, and concatenate the weighted first feature map and the weighted pooled feature map along the channel dimension to form a combined feature map. Step 5. Input the combined feature map into the decoder module. Apply the third channel attention module inside the decoder module to enhance the edge detail modeling capability. Then, upsample the preliminary segmentation result through bilinear interpolation to restore it to the resolution of the original input remote sensing image and generate the final refined segmentation map. Step 6. Load and preprocess the training remote sensing image data and its corresponding label image data, and input the preprocessed training remote sensing image data and its corresponding label image data into the DeepLabV3PlusCA model; Step 7. Block-based sliding window prediction: For large remote sensing images, a block-based reading and sliding window prediction strategy is adopted to avoid memory overflow while preserving the image's georeferenced information; Let the large remote sensing image in step 7 be I, with a size of H*W, where H represents the height and W represents the width. The image is segmented into 256x256 pixel tiles using a sliding window method, matching the input size of the DeepLabV3PlusCA model. The sliding window step size is s. Assume image I is segmented into n tiles, and the i-th tile is represented as I... i i = 1, 2, ..., n; Each tile I i Input the trained model M for inference and generate the corresponding binary segmentation mask M. i That is, M i =M(I i );exist During the splicing process, the average method is used to fuse the predicted values ​​for the overlapping areas in the sliding window process. Let I be two adjacent tiles. j and I j+1 The overlapping region is O, and O is in I j The predicted value in part is P j，0 , in I j+1 The predicted value in part is P j+1，0 Then the fused predicted value P0 is: ， Finally, all the processed segmentation masks are concatenated to form a complete raster extraction result R, and a vector file is output.

2. The end-to-end extraction method for natural field-shaped remote sensing images based on multi-attention channels according to claim 1, characterized in that: Step 6 further includes extracting features through the backbone network after the data is input into the DeepLabV3PlusCA model, and adjusting the feature weights using the first, second, and third channel attention modules to perform data augmentation on the training remote sensing image data. Data augmentation includes random flipping, random rotation, and random scaling operations to increase data diversity and improve the model's generalization ability. The loss value is calculated using a combined loss function, and the model parameters are updated using the Adam optimizer. Finally, the model performance is evaluated, performance indicators are recorded, and the learning curve is output. If the preset training rounds or convergence criteria are reached, the model is saved and training ends; otherwise, iterative training continues.

3. The end-to-end extraction method for natural field-shaped remote sensing images based on multi-attention channels according to claim 2, characterized in that: In step 6, the enhanced natural field-shaped remote sensing image training data is processed using a combined loss function L. total To optimize the model and calculate the loss value, the combined loss function L... total The binary cross-entropy loss L BCE Dice loss L Dice and gradient loss L Grad Composed of three weighted parts: L total =a1*L BCE +a2*L Dice +a3*L Grad Where a1, a2, and a3 are the weighting coefficients for each loss; The binary cross-entropy loss L BCE The formula used to measure the pixel-by-pixel classification accuracy between the predicted and target values ​​is as follows: ， Where N is the total number of pixels, y i It is the true label of pixel i, p i It is the probability that the model predicts pixel i to be a natural field shape; The Dice loss L Dice Based on the intersection-union ratio (IUGR) calculation, this study focuses on the degree of overlap between the predicted and actual regions to optimize segmentation similarity. The calculation formula is as follows: ， in, It is a constant, N is the total number of pixels, and y is... i It is the true label of pixel i, p i It is the probability that the model predicts pixel i to be a natural field shape; The gradient loss L Grad The gradient difference between the predicted image and the target image in the horizontal (x) and vertical (y) directions is calculated to ensure spatial continuity and edge smoothness of the output; the calculation formula is as follows: ， in and These represent the gradient operators in the x and y directions, respectively, where N is the total number of pixels, and y is the gradient value. i It is the true label of pixel i, p i It is the probability that the model predicts pixel i to be a natural field shape.

4. The end-to-end extraction method for natural field-shaped remote sensing images based on multi-attention channels according to claim 3, characterized in that: In step 1, the backbone network used is DeepLabV3_ResNet50. The original input natural field-shaped remote sensing image has a size of [3,H,W], where 3 represents 3 channels. The remote sensing image is used for preliminary feature extraction through the pre-trained backbone network. Specifically, the input remote sensing image is processed sequentially through the initial convolutional layer, batch normalization layer, ReLU activation function, and a max pooling layer of the backbone network to generate a preliminary feature map with a size of [64,H / 4,W / 4].

5. The end-to-end extraction method for natural field-shaped remote sensing images based on multi-attention channels according to claim 4, characterized in that: In step 2, there are four residual learning stages: layer1, layer2, layer3, and layer4. The initial feature map is downsampled step by step in the four residual learning stages to extract deeper features. After layer4, a high-dimensional feature map with a size of [2048, H / 16, W / 16] is generated, while the first feature map with a size of [256, H / 4, W / 4] after processing by layer1 is retained.

6. The end-to-end extraction method for natural field-shaped remote sensing images based on multi-attention channels according to claim 5, characterized in that: In step 3, a high-dimensional feature map of size [2048, H / 16, W / 16] is input into the Aperture Spatial Pyramid Pooling (ASPP) module. The Aperture Spatial Pyramid Pooling (ASPP) module can effectively capture multi-scale contextual information without increasing the number of parameters and computational cost, thereby enhancing the model's ability to perceive targets at different scales. The pooled feature map is upsampled to a size of [256, H / 4, W / 4] using bilinear interpolation, thus maintaining consistency with the spatial size of the first feature map.

7. The end-to-end extraction method for natural field-shaped remote sensing images based on multi-attention channels according to claim 6, characterized in that: Specifically, step 4 involves applying the first channel attention module to the first feature map [256, H / 4, W / 4] output by layer 1. This module uses adaptive average pooling, a dimensionality-reduced fully connected layer, ReLU activation, and a Sigmoid activation function to weight and adjust the channels of the feature map, enhancing key channels containing information about the fine structures of the natural field-shaped edges. The pooled feature map of size [256, H / 4, W / 4] is then applied to the second channel attention module. Using the same mechanism as the first channel attention module, the pooled feature map channels are weighted to optimize the model's focus on the natural field-shaped region and reduce background noise interference. Finally, the weighted first feature map and the weighted pooled feature map are concatenated along the channel dimensions to form a combined feature map of size [512, H / 4, W / 4].

8. The end-to-end extraction method for natural field-shaped remote sensing images based on multi-attention channels according to claim 7, characterized in that: The decoder module in step 5 includes a convolutional layer, a batch normalization layer, and a ReLU activation layer. The combined feature map with a size of [512, H / 4, W / 4] is processed by the decoder containing a third channel attention module, and then the number of channels of the combined feature map is reduced to 1 by a 1x1 convolutional layer to generate a preliminary extraction result with a size of [1, H / 4, W / 4]. Finally, the preliminary segmentation result is upsampled by a 4x bilinear interpolation method to restore it to the original input image resolution [1, H, W], generating the final refined segmentation map.

Images