A visible light remote sensing image water body extraction method based on deep learning

Abstract

The application provides a visible light remote sensing image water body extraction method based on deep learning, belongs to the field of deep learning, and aims to solve problems of variable water body region scale, fuzzy boundary, complex background and the like in a remote sensing image, and improve water body extraction precision through multi-module collaborative optimization. First, at the encoder end, an EWET module is used to replace a ResNet encoder of an original U-Net3+, and a three-branch collaborative mechanism is used to strengthen water body feature extraction capability; second, at a feature fusion end, an AHC-ASPP module is introduced, a hierarchical hollow convolution and a deformable convolution are used to cooperatively expand a receptive field, and precise perception of a multi-scale water body is realized; and finally, an EMAM module is embedded at a jump connection position, a combination of spatial attention and channel attention is used in series, background interference is effectively inhibited, and water body boundary perception is enhanced.

Description

Technical Field

[0001] This invention relates to a method for extracting water bodies from visible light remote sensing images based on deep learning, belonging to the technical field of remote sensing image semantic segmentation and water body information extraction. Background Technology

[0002] Satellite remote sensing technology, as a comprehensive Earth observation method, has significant advantages over traditional observation methods, such as wide coverage and the ability to conduct periodic and repeated monitoring. It has become a primary technology for global and regional surface water monitoring. Visible light remote sensing image water extraction technology aims to build models using visible light remote sensing data to automatically identify water bodies such as rivers, lakes, and oceans, thereby obtaining accurate water resource distribution information. This has significant application value in water resource surveys, flood monitoring, environmental protection, and disaster prevention and mitigation.

[0003] Due to the inherent characteristics of water bodies and the features of the remote sensing imaging environment, water body extraction from visible light remote sensing images presents several challenges: Background features such as building shadows, mountain shadows, and vegetation cover can cause spectral confusion with water targets, making effective separation of water from the background difficult. The scale of water bodies in remote sensing images varies greatly, ranging from small tributaries a few meters wide to large lakes several kilometers wide. This scale diversity adds further difficulty to the receptive field design and multi-scale feature perception of target detection networks. Furthermore, the boundary regions between water bodies and the background typically have low contrast and blurred edges, and the irregular shape of water bodies makes precise boundary localization more complex. Traditional semantic segmentation methods struggle to simultaneously consider both the integrity of the water body's interior and the fineness of its boundaries, leading to problems such as jagged edges, holes, or missed detections in the extraction results.

[0004] This study proposes an improved water extraction algorithm based on U-Net3+, taking into account the characteristics of water bodies in visible light remote sensing images. The aim is to enhance the accuracy and robustness of water body segmentation in complex scenes, achieving more precise water region identification and boundary localization. This method not only optimizes the model's ability to perceive water bodies at multiple scales but also enhances its adaptability to complex background interference such as shadows and vegetation, which is of significant value in promoting the development of water body extraction technology in remote sensing images. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a deep learning-based method for water body extraction from visible light remote sensing images. This method can achieve more accurate water body extraction results when dealing with the characteristics of water body regions in visible light remote sensing images, such as varying scales, complex backgrounds, and blurred boundaries.

[0006] To address the shortcomings of existing technologies, this application proposes an improved method for water body extraction from visible light remote sensing images, based on U-Net3+ optimization. The method includes the following steps: First, acquire the visible light remote sensing image dataset GID and preprocess it, including image cropping, blur removal, and duplicate image removal, to improve data quality. Then, classify the preprocessed dataset according to a classification standard, extracting water body categories as targets and other categories as background, and proportionally dividing it into training, testing, and validation sets. Simultaneously, image data augmentation is performed to expand sample diversity.

[0007] After preparing the dataset, U-Net3+ was improved by adjusting hyperparameters according to task requirements and iteratively optimizing model weights using the training set to minimize prediction error. During training, model performance was evaluated in real time using the validation set to prevent overfitting. Finally, the model's IoU, Precision, Recall, and F1-Score were evaluated using the test set. The RSWE-Net network obtained after optimizing U-Net3+ has significant advantages in water body extraction from visible light remote sensing images: it improves water body extraction accuracy, reduces missed and false detections, adapts to different resolutions and complex environments, enhances the ability to identify small water body boundaries, and provides reliable data support for subsequent hydrological analysis.

[0008] At the encoder end, the U-Net3+ original ResNet encoder is replaced with the EWET module. The EWET module achieves progressive feature extraction through four hierarchical feature processing stages: Stage 1 extracts primary water body edge features at full resolution using standard convolution; Stage 2 uses depthwise separable convolution to enhance local water body geometric features; Stage 3 fuses deformable convolutional pathways and global attention pathways to collaboratively model complex water body boundaries; Stage 4 performs downsampling and semantic feature integration through stride convolution. The core bottleneck unit of EWET adopts a three-branch parallel architecture: the feature recalibration branch generates channel attention weights through global average pooling and sigmoid; the backbone enhancement branch uses depthwise convolution combined with multi-level residual structures to deeply mine the spatial morphology of the water body; the Enhanced-PT attention branch is responsible for global context modeling; the three branches are adaptively aggregated by a gated fusion module to generate high-quality multi-scale water body feature representations.

[0009] The EWET module performs multi-stage feature enhancement on the attention path: It abandons the feature downsampling mechanism of the original module, replacing fixed-size pooling with an adaptive average pooling layer to dynamically match the context regions of water bodies at different scales. A 1×1 convolutional layer is used for nonlinear transformation and feature renormalization, and a group normalization layer standardizes the feature distribution. Key features use 2×2 max pooling to extract local salient responses, helping to capture abrupt changes at water body edges. Query features use 3×3 average pooling with a stride of 1 to maintain boundary continuity, ensuring no resolution loss during attention calculation and avoiding the loss of water body boundary details due to downsampling.

[0010] The AHC-ASPP module is introduced into the decoding path. This module makes key reconstructions based on the traditional ASPP: after the input features are compressed into channel dimensions by 1×1 convolution, they are fed in parallel into four specialized branches: Global Average Pooling (GAP), Depthwise Separable Convolution with a dilation rate r=6, Depthwise Separable Convolution with a dilation rate r=18, and 3×3 Deformable Convolution. Its core innovation lies in hierarchical connections: the dilation rate r=18 branch directly receives the output features from the r=6 branch as input, forming a feature transfer link that progresses from local details to the global context. This breaks the limitation of independent parallelism of branches in the standard ASPP, significantly improving the ability to perceive the overall morphology of large water bodies. The deformable convolution branch enables the convolution kernel to adaptively fit the complex and varied boundary contours of water bodies, greatly improving the boundary segmentation accuracy of irregular small water bodies.

[0011] The EMAM hybrid attention mechanism module consists of a spatial attention module and a feature channel attention module connected in series, embedded at the skip connections between the encoder and decoder in each layer of U-Net3+. After receiving features from the encoder, the SAM module adaptively perceives spatial saliency differences through dynamic parameter pooling, generating a spatial attention weight map. Simultaneously, it extracts feature maps containing local details through a detail information extraction branch. The two are then weighted and fused to obtain a spatial feature mask M. The FCA module receives the spatial feature mask M output by SAM, models long-range semantic dependencies between channels through matrix multiplication, and obtains the channel attention weight matrix through a Softmax operation, achieving precise enhancement of key water body information channels and effective suppression of background interference channels.

[0012] The deformable convolution module can automatically adjust the sampling point offset according to the shape of the water body target, improving the network's adaptability to irregular water body contours. In scenarios such as small tributaries and lake edges, deformable convolution can more accurately fit the water body boundary, thereby improving the model's overall detection performance and robustness for complex water body shapes.

[0013] As can be seen from the above, the advantages of this invention are as follows: This invention provides a deep learning-based method for water body extraction from visible light remote sensing images. This method enhances multi-scale feature extraction of water bodies through the EWET encoder, expands the receptive field to cover water bodies of various scales, from small tributaries to large lakes, through the AHC-ASPP module, and accurately locates water body boundaries and suppresses background interference through the EMAM attention module. The three modules work synergistically to significantly improve the accuracy of water body extraction and the quality of boundary segmentation. This effectively solves the problem of insufficient extraction accuracy in traditional methods when facing challenges such as varying water body scales, complex backgrounds, and blurred boundaries in visible light remote sensing images. It achieves high-precision automated water body information extraction and has significant practical application value. Attached Figure Description

[0014] Figure 1 A schematic diagram of the overall research process for a deep learning-based water body extraction method from visible light remote sensing images provided by this invention; Figure 2 This is a schematic diagram of the overall structure of the RSWE-Net network provided by the present invention; Figure 3 This is a schematic diagram of the structure of the EWET Bottleneck module provided by the present invention; Figure 4 This is a schematic diagram of the structure of the AHC-ASPP module provided by the present invention; Figure 5 This is a schematic diagram of the EMAM hybrid attention mechanism module provided by the present invention; Detailed Implementation The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so as to more clearly illustrate the advantages and features of the present invention and enable those skilled in the art to more easily understand the essence of the present invention. The description of specific embodiments is intended to further clarify the scope of protection of the present invention and provide a basis for defining the claims.

[0015] It should be understood that the term "comprising" as used in this specification and the appended claims is intended to indicate the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the possibility of the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or combinations thereof.

[0016] The technical solutions will now be clearly and completely described with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only a part of this application, and not all embodiments.

[0017] Example Dataset: The GID dataset was used as the experimental dataset for water body extraction from visible light remote sensing images in this invention. The GID dataset contains 150 labeled large-scale Gaofen-2 (GF-2) satellite images, each 6800×7200 pixels in size, covering more than 60 different cities in China, and including five categories: buildings, farmland, forests, grasslands, and water bodies. This invention extracts the water body category from the dataset, setting other categories as the background. Large-scale images containing water bodies were traversed and cropped to 512×512 pixels, resulting in 8953 images. These images were then divided into training, testing, and validation sets in a 7:2:1 ratio.

[0018] Use Labelimg to annotate the acquired images, including the category and location information of the targets.

[0019] The acquired images undergo denoising and contrast enhancement to improve image quality. Images are further enhanced through random cropping, rotation, and scaling to increase the diversity of training samples. The dataset is divided into training, validation, and test sets for model training, tuning, and evaluation. The final set is divided into training, validation, and test sets in a 7:2:1 ratio for model training, tuning, and evaluation.

[0020] In this embodiment, U-Net3+ is used as the baseline model, and its structure is optimized and improved to enhance the ability to extract water targets in visible light remote sensing images, forming the RSWE-Net network. RSWE-Net uses EWET Bottleneck as the encoder network, introduces the AHC-ASPP module in the decoding path, and integrates the EMM module at the skip connections.

[0021] On the encoder side: The original ResNet encoder of U-Net3+ is replaced with the EWET module to construct a four-level hierarchical feature processing stage. Stage 1 extracts primary edge responses at full resolution through 3×3 standard convolutions, combined with batch normalization and ReLU activation, outputting a high-resolution feature map of 512×512, fully preserving the geometric details of small-scale water bodies. Stage 2 adopts a depthwise separable convolutional architecture, using 3×3 depthwise convolutions combined with 1×1 pointwise convolutions to enhance the representation of local geometric features of the water body, and compresses the feature map to 256×256 through max pooling with a stride of 2, expanding the channels to 128 dimensions. Stage 3 innovatively integrates a dual-path processing mechanism of deformable convolutional pathways and global attention pathways to collaboratively model complex water body boundaries and large-scale spatial relationships, outputting a 128×128×256 feature map through max pooling. Stage 4 jointly performs feature extraction and downsampling through 3×3 depthwise separable convolutions, outputting a 64×64×512 feature map, encoding high-level semantic information, and providing global water body distribution guidance for the decoder.

[0022] Feature fusion stage: An AHC-ASPP module is introduced into the decoding path to achieve hierarchical extraction and adaptive fusion of multi-scale water features. The module's input features are first compressed in channel dimension by a 1×1 convolution, then input into four functional branches in parallel: a global average pooling branch performs global semantic compression; a depthwise separable convolution branch with a dilatancy r=6 focuses on extracting medium-scale water structure information; a depthwise separable convolution branch with a dilatancy r=18 receives the output of the r=6 branch in a hierarchical manner, constructing a feature transfer link from local details to the global context, significantly expanding the perception range of large water bodies; and a 3×3 deformable convolution branch adaptively fits irregular water body contours by learning sampling point offsets. In the cross-scale feature fusion stage, the features from the r=6 and r=18 branches are initially fused and then concatenated with the output of the deformable convolution branch. Finally, a spatial weight map is generated through a 1×1 convolution and a sigmoid function to enhance the response of water-sensitive features and suppress background interference such as building shadows and vegetation noise.

[0023] Attention-wise: An EMAM module is integrated at the full-scale skip connections of U-Net3+ to achieve synergistic optimization of spatial detail enhancement and channel feature selection. EMAM is composed of SAM and FCA concatenated. The SAM module first uses dynamic parameter pooling to adaptively perceive spatial saliency differences, and generates a spatial attention map A through a combination of Conv1×1+BN+ReLU; the detail information extraction branch extracts a feature map B containing detail information through a combination of two Conv3×3+BN+ReLU convolutions; the two are weighted and fused to obtain a spatial feature mask M, which directly and powerfully enhances the model's ability to represent water boundary regions.

[0024] The FCA module receives the spatial feature mask M output by SAM and reshapes the feature M into... Then, matrix multiplication is performed between M and its transpose. Finally, the channel attention obtained by applying the softmax layer can be expressed by the following formula: middle This represents the influence weights between channels. Subsequently, matrix multiplication is performed between the transpose of the input matrix X and the weight matrix M. The result is then reshaped to form a new matrix. Next, multiply this matrix by the scale parameter. Then, element-wise summation is performed using the feature matrix M. The final output is represented by the following formula: The experimental hardware environment in this embodiment is as follows: CPU processor: Intel Core i9; RAM: 64GB; GPU: NVIDIA GeForce RTX 3090 (24GB VRAM); Storage: 1TB NVMe SSD. The experimental software environment is as follows: Operating system: Ubuntu 20.04; Programming software: Python 3.8; Model framework: PyTorch 2.0.1; Deep learning acceleration library: CUDA 11.7. The training epochs are 100, the optimizer is Adam, and the learning rate decay strategy uses cosine annealing scheduling.

[0025] In summary, this application proposes a deep learning-based method for water body extraction from visible light remote sensing images. Addressing the problems of missed extraction, incorrect extraction, and inaccurate boundary segmentation caused by factors such as variable water body scale, blurred boundaries, and complex backgrounds in visible light remote sensing images, a novel improved network, RSWE-Net, is designed. This network provides an innovative water body extraction scheme that significantly improves the accuracy and boundary localization capability of water body extraction from visible light remote sensing images, possessing high practical application value, especially suitable for large-scale automated remote sensing water body information extraction tasks.

[0026] The above embodiments illustrate only one specific implementation of the present invention. Although described in detail, they do not limit the scope of patent protection of the present invention. It should be particularly noted that those skilled in the art can make various modifications and improvements without departing from the basic concept of the present invention, and all such modifications and improvements should be considered within the scope of protection of the present invention.

Claims

1. A method for extracting water bodies from visible light remote sensing images based on deep learning, characterized in that, Includes the following steps: S1: Construct a visible light remote sensing image water body extraction dataset. Crop and preprocess the original large-scale satellite images in the GID dataset, extract water body category labeling information, and divide them into training and test sets according to the proportion. At the same time, perform data augmentation on the images to expand the sample diversity. S2: Construct the visible light remote sensing image water body extraction network RSWE-Net, replace the U-Net3+ original ResNet encoder with the EWET encoder, introduce the AHC-ASPP module in the decoding path for multi-scale feature extraction, and embed the hybrid attention mechanism module EAM at the full-scale skip connection. S3: Input the preprocessed training set into RSWE-Net for training, and use multiple rounds of iteration to make the model converge to the optimal state; S4: Input the test set into the trained RSWE-Net for inference and output pixel-level water body segmentation results to achieve accurate extraction of water body regions in visible light remote sensing images.

2. The method for extracting water bodies from visible light remote sensing images based on deep learning as described in claim 1, characterized in that, The data augmentation operations in step S1 include random flipping, random rotation, color jittering, and random cropping to enhance the model's ability to generalize to water targets in complex backgrounds; the image cropping uses a sliding window method to divide the large-scale image into sub-images of size 512×512.

3. The method for extracting water bodies from visible light remote sensing images based on deep learning as described in claim 1, characterized in that, The EWET encoder module includes four hierarchical feature processing stages, with the following processing units for each stage: Stage 1 maintains full resolution and extracts primary edge features through standard convolution; Stage 2 uses depthwise separable convolution to enhance local features; Stage 3 fuses deformable convolutional pathways and global attention pathways to collaboratively model water body boundaries; Stage 4 uses depthwise separable convolution to jointly implement feature extraction and downsampling, outputting high-level semantic features.

4. The method for extracting water bodies from visible light remote sensing images based on deep learning as described in claim 3, characterized in that, The EWET module adopts a three-branch parallel architecture: the feature recalibration branch generates channel attention weights through global average pooling and Sigmoid activation; the main enhancement branch uses 3×3 depth convolution combined with multi-level residual structure to deeply mine the spatial morphological features of water bodies. The Enhanced-PT attention branch achieves cross-regional water body association representation through global context modeling; the outputs of the three branches are adaptively aggregated by the gated fusion module to obtain a unified multi-scale water body feature representation.

5. The method for extracting water bodies from visible light remote sensing images based on deep learning as described in claim 1, characterized in that, The AHC-ASPP module includes: input features are compressed through a 1×1 convolution channel and then fed into four branches in parallel: a global average pooling branch to extract global context information; a depthwise separable convolution branch with a dilatancy r=6 to extract medium-scale water features; a depthwise separable convolution branch with a dilatancy r=18 to receive the output of the previous branch in a hierarchical manner, expanding the perception range for large water bodies; a 3×3 deformable convolution branch to dynamically fit irregular water body contours; and the outputs of the four branches are fused across scales and then a spatial weight map is generated by a 1×1 convolution and a sigmoid function to achieve adaptive enhancement of water-sensitive features.

6. The method for extracting water bodies from visible light remote sensing images based on deep learning as described in claim 5, characterized in that, The hierarchical connection method in the AHC-ASPP module breaks the independent and parallel structure of each branch in the traditional ASPP. The output features of the branch with a void ratio of r=6 serve as the input of the branch with a void ratio of r=18, forming a progressive feature transmission link from local details to global context, which significantly improves the ability to grasp the overall morphology of large-scale water bodies.

7. The method for extracting water bodies from visible light remote sensing images based on deep learning as described in claim 1, characterized in that, The EMAM module consists of a spatial attention module and a feature channel attention module connected in series: SAM receives features from the encoder, generates a spatial attention weight map through dynamic parameter pooling, and extracts local features of branches by fusing detail information, outputting an enhanced spatial feature mask; FCA receives the spatial feature mask output by SAM, constructs an inter-channel influence weight matrix through matrix multiplication and Softmax operation, realizes adaptive feature recalibration of the channel dimension, and finally outputs a highly discriminative feature map that takes into account both spatial details and channel features.

8. The method for extracting water bodies from visible light remote sensing images based on deep learning as described in claim 7, characterized in that, The EMM module is embedded at the encoder-decoder jump connection of each layer of U-Net3+, which can simultaneously highlight important water-related features in both the spatial and channel domains, weaken background interference information, and effectively improve the problem of blurred water boundaries caused by feature transmission loss.

9. The method for extracting water bodies from visible light remote sensing images based on deep learning as described in claim 1, characterized in that... The data-augmented training set is input into the network model for training, and the Adam optimizer is used for multiple rounds of iterative optimization. Overfitting is prevented by monitoring the IoU evaluation metric on the validation set, and finally a converged network model is obtained.

Images