A water body segmentation method and system based on a SAR remote sensing image
Patent Information
- Application Number
- CN202410583466.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-11
- Publication Date
- 2026-10-20
- Estimated Expiration
- 2044-05-11
AI Technical Summary
[0003]但是，相比于普通场景的图像，遥感图像中水体在地表呈现出各种不同形态，从大型湖泊和主要河流到小型池塘和支流，它们的尺寸变化范围非常广泛，这种多样性使得在SAR图像中对水体进行准确分割变得非常具有挑战性；
[0025]本发明首先获取SAR遥感图像数据，对其进行预处理后划分为训练集和测试集，将测试集中的图像送入到模型进行测试，得到分割结果，相比其他图像分割方法，本发明有效提高了SAR遥感图像中细小水体及支流的分割精度。
Smart Images
Figure CN118334347B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water body segmentation methods, and in particular to a water body segmentation method and system based on SAR remote sensing images. Background Technology
[0002] In recent years, floods and water pollution caused by surface water have become hot issues of concern to human society. Therefore, timely detection of river health is crucial for flood early warning and post-disaster assessment. Synthetic Aperture Radar (SAR) satellites use microwave bands for Earth observation, have strong penetration capabilities, and are suitable for all-weather, all-day operation, making them ideal for Earth observation in cloudy and rainy areas. With the development of technology, the role of remote sensing images in water body segmentation is becoming increasingly apparent. Compared with field surveys, it saves manpower and resources and avoids various risks associated with field surveys. Water body segmentation algorithms based on SAR remote sensing images play an important role in water resource surveys, ecological assessments, and flood disaster response.
[0003] However, compared to images of ordinary scenes, water bodies in remote sensing images present a variety of different forms on the Earth's surface, ranging from large lakes and major rivers to small ponds and tributaries. Their size varies greatly, and this diversity makes it very challenging to accurately segment water bodies in SAR images.
[0004] However, in existing technologies, especially for the identification of small water bodies and tributaries, there is often a tendency for loss and misjudgment. Therefore, this invention proposes a water body segmentation method and system based on SAR remote sensing images to solve the above problems. Summary of the Invention
[0005] Given the technical problems in the background technology, such as the tendency for small water bodies and tributaries to be lost or misjudged, this invention proposes a water body segmentation method and system based on SAR remote sensing images.
[0006] This invention proposes a water body segmentation method and system based on SAR remote sensing images, comprising the following steps:
[0007] S1: Acquire SAR remote sensing images, preprocess them, and divide them into training and test sets;
[0008] S2: Improve the DeepLabV3+ network, and denote the improved model as DAAG-DeepLabV3+;
[0009] S3: Use the training set to train the model and obtain an improved model;
[0010] S4: Feed the test set into the model for testing to obtain the final segmentation result.
[0011] Preferably, in step S1, acquiring SAR remote sensing images, preprocessing them, and dividing them into training and testing sets includes:
[0012] Images from Gaofen-3 and Qilu-1 satellites, which are rich in water information, were selected from different times and environments. Before use, they were preprocessed. The preprocessing process was as follows: data import, complex data conversion, filtering, and geocoding.
[0013] The preprocessed SAR remote sensing images were uniformly cropped to 1024×1024 pixels, with a total of 675 images. Then, 80% of the images were used to form the training set, and 20% were used to form the test set.
[0014] Preferably, in step S2, the DeepLabV3+ network is improved, and the improved model is denoted as DAAG-DeepLabV3+, including:
[0015] The model includes an encoding module and a decoding module. The encoding module uses a lightweight backbone network, a densely hollow spatial pyramid pooling (DenseASPP) module, and a bar pooling module. The decoding module introduces an attention gate module.
[0016] The encoder module uses MobileNetV2 as the backbone network, replacing the original Xception backbone network. After processing by the backbone network, the image is divided into two key feature information: high-level feature information P1 and low-level feature information P2. P1 contains the abstract semantic information of the image, while P2 contains the details and texture information of the image. Next, P1 is fed into the improved DenseASPP module, which uses convolutions with different dilation rates to extract receptive field information at different scales and uses bar pooling branches to integrate global and local contextual information. Finally, to ensure the consistency of the number of channels of the features, a 1×1 convolution is used to adjust the features to obtain the final feature layer P3. Through this process, the model can extract rich semantic information while preserving image details.
[0017] The decoder module first upsamples feature layer P3 by 4 times to obtain feature layer P4. Then, it combines the low-level feature information P2, which has been adjusted for the number of channels by a 1×1 convolution, with P4 and inputs it into the Attention gate module. The design of this module helps the network to better focus on water features. Under the action of the Attention gate module, the output feature layer P5 is obtained. Next, feature layers P4 and P5 are superimposed and fused to make full use of information from different levels. Finally, through a 3×3 convolution and a 4x upsampling operation, the final water segmentation result is obtained. Through this series of steps, the decoder module can effectively extract and integrate information from features at different levels, thereby producing accurate water segmentation results.
[0018] Preferably, in step S3, training the model using the training set to obtain an improved model includes:
[0019] The lightweight backbone network MobileNetV2 used in the encoder already has a pre-trained model on the VOC dataset, which can be directly transferred and used.
[0020] The training set is input into the DAAG-DeepLabV3+ network in batches to obtain the image segmentation prediction results;
[0021] The Dice loss function is used to calculate the error between the predicted and actual results to obtain the loss value. Then, based on the loss value, Adam is used as the optimizer for backpropagation to update the weight parameters of the DAAG-DeepLabV3+ model. Finally, through continuous iteration, the optimal model for water segmentation is obtained.
[0022] Preferably, in step S4, the test set is fed into the model for testing to obtain the final segmentation result, including:
[0023] Load the optimal segmentation model and evaluate the IoU, Recall, Precision, and F1-Score values of the DAAG-DeepLabV3+ model using the test set; input the model with images from the test set to obtain visualized water segmentation results.
[0024] The beneficial effects of this invention are:
[0025] This invention first acquires SAR remote sensing image data, preprocesses it, and divides it into training and testing sets. Images from the testing set are then fed into the model for testing to obtain segmentation results. Compared with other image segmentation methods, this invention effectively improves the segmentation accuracy of small water bodies and tributaries in SAR remote sensing images. Attached Figure Description
[0026] Figure 1This is a flowchart of the present invention;
[0027] Figure 2 This is a schematic diagram of the DAAG-DeepLabV3+ structure in an embodiment of the present invention;
[0028] Figure 3 This is a schematic diagram of the improved DenseASPP structure in an embodiment of the present invention;
[0029] Figure 4 This is a schematic diagram of the strip pooling module structure in an embodiment of the present invention;
[0030] Figure 5 This is a schematic diagram of the Attention gate structure in an embodiment of the present invention;
[0031] Figure 6 This is a graph showing the test results of the present invention. Detailed Implementation
[0032] The present invention will be further explained below with reference to specific embodiments.
[0033] Reference Figure 1-6 Example 1
[0034] This embodiment proposes a water body segmentation method and system based on SAR remote sensing images, including the following steps:
[0035] S1: Acquire SAR remote sensing images, preprocess them, and divide them into training and test sets. This includes:
[0036] Images from Gaofen-3 and Qilu-1 satellites, which are rich in water information, were selected from different times and environments. Before use, they were preprocessed. The preprocessing process was as follows: data import, complex data conversion, filtering, and geocoding.
[0037] The preprocessed SAR remote sensing images were uniformly cropped to 1024×1024 pixels, with a total of 675 images. Then, 80% of the images were used to form the training set, and 20% were used to form the test set.
[0038] S2: Improve the DeepLabV3+ network, denoted as DAAG-DeepLabV3+, including:
[0039] The model includes an encoding module and a decoding module. The encoding module uses a lightweight backbone network, a densely hollow spatial pyramid pooling (DenseASPP) module, and a bar pooling module. The decoding module introduces an attention gate module.
[0040] The encoder module uses MobileNetV2 as the backbone network, replacing the original Xception backbone network. After processing by the backbone network, the image is divided into two key feature information: high-level feature information P1 and low-level feature information P2. P1 contains the abstract semantic information of the image, while P2 contains the details and texture information of the image. Next, P1 is fed into the improved DenseASPP module, which uses convolutions with different dilation rates to extract receptive field information at different scales and uses bar pooling branches to integrate global and local contextual information. Finally, to ensure the consistency of the number of channels of the features, a 1×1 convolution is used to adjust the features to obtain the final feature layer P3. Through this process, the model can extract rich semantic information while preserving image details.
[0041] The decoder module first upsamples feature layer P3 by 4 times to obtain feature layer P4. Then, it combines the low-level feature information P2, which has been adjusted for the number of channels by 1×1 convolution, with P4 and inputs it into the Attention gate module. The design of this module helps the network to better focus on water features. Under the action of the Attention gate module, the output feature layer P5 is obtained. Next, feature layers P4 and P5 are superimposed and fused to make full use of information from different levels. Finally, through 3×3 convolution and 4x upsampling, the final water segmentation result is obtained. Through this series of steps, the decoder module can effectively extract and integrate information from features at different levels, thereby producing accurate water segmentation results.
[0042] S3: Train the model using the training set to obtain an improved model, including:
[0043] The lightweight backbone network MobileNetV2 used in the encoder already has a pre-trained model on the VOC dataset, which can be directly transferred and used.
[0044] The training set is input into the DAAG-DeepLabV3+ network in batches to obtain the image segmentation prediction results;
[0045] The Dice loss function is used to calculate the error between the predicted and actual results to obtain the loss value. Then, based on the loss value, Adam is used as the optimizer for backpropagation to update the weight parameters of the DAAG-DeepLabV3+ model. Finally, through continuous iteration, the optimal model for water segmentation is obtained.
[0046] S4: Feed the test set into the model for testing to obtain the final segmentation result, including:
[0047] Load the optimal segmentation model and evaluate the IoU, Recall, Precision, and F1-Score values of the DAAG-DeepLabV3+ model using the test set; input the model with images from the test set to obtain visualized water segmentation results.
[0048] Simulation experiment:
[0049] 1. Simulation experimental conditions:
[0050] Simulation experimental conditions for this invention: Server GPU: GeForce RTX2080 Ti, 12G video memory;
[0051] The software platform for the simulation experiment of this invention is: Ubuntu 18.04 system, Python 3.8, PyTorch-GPU 1.19.0;
[0052] 2. Simulation content and analysis of experimental results:
[0053] The simulation experiment of this invention uses this invention and four existing technologies (SegNet segmentation method, Unet segmentation method, PSPNet segmentation method, and DeepLabV3+ segmentation method) to segment the water body on the test set and obtain the segmentation results.
[0054] The input images used in the simulation experiment of this invention are all the images in the test set, which are from the Gaofen-3 satellite and the Qilu-1 satellite, and contain a total of 675 images.
[0055] Simulation Experiment 1 involves conducting the method of the present invention under the above simulation conditions to obtain the detection results;
[0056] Simulation Experiment 2 uses the SegNet method from existing technologies under the above simulation conditions to conduct experiments and obtain detection results;
[0057] Simulation Experiment 3 uses the Unet method from existing technology under the above simulation conditions to conduct the experiment and obtain the detection results;
[0058] Simulation Experiment 4 uses the PSPNet method from the existing technology under the above simulation conditions to conduct an experiment and obtain detection results;
[0059] Simulation Experiment 5 uses the DeepLabV3+ method from existing technology to conduct experiments under the above simulation conditions and obtains detection results;
[0060] To test the water segmentation effect of the present invention, four evaluation indicators (IoU, Recall, Precision and F1-Score) were used to evaluate the segmentation results of the four methods respectively. The higher the value of these four indicators, the more accurate the water segmentation result. The calculation results are shown in Table 1.
[0061] Comparison table of simulation experiment results of this invention
[0062]
[0063]
[0064] As can be seen from the table above, compared with the existing SegNet, Unet, PSPNet and DeepLabV3+ methods, the IoU of the present invention is 87.06%, Recall is 92.29%, Precision is 94.36%, and F1-Score is 93.31%. The values of all four indicators are higher than those of the existing methods, proving that the present invention can achieve higher water segmentation accuracy.
[0065] The following is combined Figure 6 The effects of this simulation experiment will be further elaborated;
[0066] Figure 6 The yellow circles represent regions that were missed in segmentation, and the blue squares represent regions that were incorrectly segmented. The first column shows the original images from the test set; the second column shows the ground truth images corresponding to the original images; the third column shows the prediction results from SegNet; the fourth column shows the prediction results from the Unet model; the fifth column shows the prediction results from the PSPNet model; the sixth column shows the prediction results from the DeepLabV3+ model; and the seventh column shows the prediction results from this invention. Figure 6 As we can see, the present invention is superior to other segmentation models in terms of the accuracy of segmenting small water bodies and tributaries, and is closest to the water body region in the ground truth image.
[0067] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A water body segmentation method and system based on SAR remote sensing images, characterized in that, Includes the following steps: S1: Acquire SAR remote sensing images, preprocess them, and divide them into training and test sets; S2: Improve the DeepLabV3+ network, and denote the improved model as DAAG-DeepLabV3+; S3: Use the training set to train the model and obtain an improved model; S4: Feed the test set into the model for testing to obtain the final segmentation result; In step S2, the DeepLabV3+ network is improved, and the improved model is denoted as DAAG-DeepLabV3+, including: The model includes an encoding module and a decoding module. The encoding module uses a lightweight backbone network, a densely populated spatial pyramid pooling (DenseASPP) module, and a bar pooling module. The decoding module introduces an attention gate module. The encoder module uses MobileNetV2 as the backbone network, replacing the original Xception backbone network. After processing by the backbone network, the image is divided into two key feature information: high-level feature information P1 and low-level feature information P2. P1 contains the abstract semantic information of the image, while P2 contains the details and texture information of the image. Next, P1 is fed into the improved DenseASPP module, which uses convolutions with different dilation rates to extract receptive field information at different scales and uses strip pooling branches to integrate global and local contextual information. Finally, to ensure the consistency of the number of channels of the features, a 1×1 convolution is used to adjust the features to obtain the final feature layer P3. The decoder module first upsamples feature layer P3 by 4 times to obtain feature layer P4. Then, it combines the low-level feature information P2 after adjusting the number of channels through 1×1 convolution with P4 and inputs it into the Attention gate module. Under the action of the Attention gate module, the output feature layer P5 is obtained. Next, feature layers P4 and P5 are superimposed and fused to make full use of information from different levels. Finally, the final water body segmentation result is obtained through 3×3 convolution and 4x upsampling operation.
2. The water body segmentation method and system based on SAR remote sensing images according to claim 1, characterized in that, In step S1, SAR remote sensing images are acquired, preprocessed, and divided into training and testing sets, including: Images from Gaofen-3 and Qilu-1 satellites, which are rich in water information, were selected from different times and environments. Before use, they were preprocessed. The preprocessing process was as follows: data import, complex data conversion, filtering, and geocoding. The preprocessed SAR remote sensing images were uniformly cropped to 1024×1024 pixels, with a total of 675 images. Then, 80% of the images were used to form the training set, and 20% were used to form the test set.
3. The water body segmentation method and system based on SAR remote sensing images according to claim 1, characterized in that, In step S3, the model is trained using the training set to obtain an improved model, including: The lightweight backbone network MobileNetV2 used in the encoder already has a pre-trained model on the VOC dataset, which can be directly transferred and used. The training set is input into the DAAG-DeepLabV3+ network in batches to obtain the image segmentation prediction results; The Dice loss function is used to calculate the error between the predicted and actual results to obtain the loss value. Then, based on the loss value, Adam is used as the optimizer for backpropagation to update the weight parameters of the DAAG-DeepLabV3+ model. Finally, through continuous iteration, the optimal model for water segmentation is obtained.
4. The water body segmentation method and system based on SAR remote sensing images according to claim 1, characterized in that, In step S4, the test set is fed into the model for testing to obtain the final segmentation result, including: loading the optimal segmentation model, using the test set to evaluate the IoU, Recall, Precision and F1-Score values of the DAAG-DeepLabV3+ model; using the images in the test set, inputting them into the model to obtain the visualized water body segmentation result.