A dual-branch multi-scale fusion SAR flood intelligent detection method

Abstract

The application provides a kind of double-branch multiscale fusion SAR flood intelligent detection method, belongs to remote sensing image processing technical field.The technical scheme is as follows: a kind of double-branch multiscale fusion SAR flood intelligent detection method, comprising the following steps: S1, processing SAR image and constructing SAR flood detection dataset;S2, construct the double-branch multiscale fusion network containing twin encoding and double-branch decoder;S3, using the SAR flood detection dataset trains and verifies network, obtains the flood detection model of training completion;S4, the double-branch SAR image pair to be detected is input into the flood detection model, and the flood detection result graph identical with input image size is output.The beneficial effects of the present application are as follows: the double-branch multiscale fusion network proposed in the present application realizes the accurate identification of large-scale submerged area and small-scale fragmented area by adaptively fusing global and local features and coordinating multiscale change information, significantly improves the flood detection precision.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a dual-branch, multi-scale fusion SAR flood intelligent detection method. Background Technology

[0002] Floods are among the most common natural disasters globally, causing significant impacts on human societies and ecosystems. In the field of flood monitoring, remote sensing technology detects changes by comparing pre- and post-disaster image data, quickly identifying inundated areas and high-risk zones, providing crucial support for accurately delineating disaster-affected areas. Synthetic aperture radar, with its advantages of penetrating clouds and being unaffected by lighting conditions, has become a primary tool for flood monitoring, widely used in disaster assessment, emergency response, and risk prevention.

[0003] Traditional change detection methods can be categorized into three types: those based on transformation, algebraic operations, and classification. The first type enhances change features through techniques such as change vector analysis and principal component analysis. The second type employs arithmetic strategies like ratio calculation, regression analysis, and difference normalization to identify changes. The third type utilizes traditional machine learning models such as support vector machines and K-means clustering for pixel-level classification. However, these methods heavily rely on manual feature design and generally suffer from limited generalization capabilities.

[0004] With the continuous development of technologies such as deep learning and computer vision, change detection is gradually becoming a research hotspot in the interdisciplinary field of remote sensing and computer vision. Unlike traditional methods, deep learning-based methods can integrate feature extraction and change detection into a single network, directly generating detection results in an end-to-end manner. In deep learning change detection models, convolutional neural networks (CNNs) have become the mainstream architecture due to their powerful image processing capabilities. Frameworks such as FC-EF, FC-Siam-conc, and FC-Siam-diff, along with their variants, have been widely used in change detection, achieving good results. However, most current methods focus heavily on multi-scale feature extraction, repeatedly using skip connections to combine low-level features from the encoder with high-level features from the decoder in multi-scale information fusion, neglecting the interrelationships between different scales. This leads to insufficient utilization of multi-scale features and difficulty in identifying flood areas of different sizes. Furthermore, these research frameworks often process semantic and differential information independently, ignoring the correlation between them.

[0005] Floods typically occur over large areas, and CNN-based methods are limited by the size of the convolutional kernel, making them ineffective at capturing large-scale floods. Transformer-based methods utilize self-attention to acquire global information and are widely used in change detection tasks; for example, BIT, ChangeFormer, and DAM-Net have achieved good results. However, applying Transformers requires large-scale training samples and significant computational resources and memory. Furthermore, compared to CNNs, Transformers struggle with pixel-level feature extraction, focusing on local information, leading to difficulties in detecting small-scale changes and poor performance in detecting change boundaries.

[0006] To address the aforementioned challenges, this invention aims to provide a dual-branch, multi-scale fusion-based intelligent SAR flood detection method, offering crucial insights for rescue planning and post-disaster assessment, thereby supporting disaster relief and recovery efforts. Summary of the Invention

[0007] The purpose of this invention is to overcome the problems in the prior art and provide a dual-branch multi-scale fusion SAR flood intelligent detection method.

[0008] The technical concept of this invention is as follows: The SAR image flood detection model consists of a twin encoder and a dual-branch decoder. The encoder utilizes the first three stages of ResNet18 as the network backbone, adjusting the stride of the first 7×7 convolutional layer from the default value of 2 to 1 to extract multi-level features while retaining more spatial detail and reducing computational complexity, thus better adapting to the requirements of change detection tasks. Then, the Global-Local Adaptive Module (GLAM) dynamically fuses global and local information using adaptive parameters α and β. This adaptive fusion mechanism allows the model to prioritize change-related information while effectively suppressing irrelevant details. In the decoder, the semantic branch uses a 3D-Attention module to sequentially interact with dual-temporal features, enhancing temporal, channel, and spatial features, while the difference branch extracts difference information using element-wise subtraction. Next, ResidualBlock is applied to process the difference and semantic information. Subsequently, the Scale Information Aggregation Module (SIAM) is used to fuse multi-level features and identify flood regions of different shapes under the semantic and difference branches. Finally, the dual-branch fusion yields the final flood detection result.

[0009] To achieve the above-mentioned objectives, the present invention employs the following technical solution: a dual-branch, multi-scale fusion SAR flood intelligent detection method, comprising the following steps:

[0010] S1, Process SAR images and construct a SAR flood detection dataset;

[0011] S2, construct a dual-branch multi-scale fusion network including a twin encoder and a dual-branch decoder;

[0012] S3, Use the SAR flood detection dataset to train and validate the network to obtain a trained flood detection model;

[0013] S4. Input the dual-temporal SAR image pair to be detected into the flood detection model, and output a flood detection result image with the same size as the input image.

[0014] Further, in step S1, constructing the SAR flood detection dataset includes:

[0015] S1.1 Select SAR image pairs with uniform resolution and complete spatial registration;

[0016] S1.2, perform pixel-level annotation on the registered SAR image pairs, including flood change areas and non-change areas;

[0017] S1.3, Perform synchronous cropping on SAR image pairs to obtain sub-image pairs of the same size and discard images that are too small;

[0018] S1.4, select samples with a change region ratio of ≥3.5% to construct a standard dataset containing training and validation sets.

[0019] Furthermore, the dual-branch multi-scale fusion network includes:

[0020] S2.1, Construct a dual-branch multi-scale fusion network, wherein:

[0021] The GLAM module in the twin encoder dynamically fuses global and local information using adaptive parameters α and β.

[0022] The dual-branch decoder includes:

[0023] The 3D-Attention module in the dual-branch decoder is used to enhance the spatiotemporal interaction of dual-temporal features;

[0024] The SIAM module in the dual-branch decoder is used to fuse multi-level features and identify flood regions in the semantic branch and the difference branch;

[0025] S2.2: Based on step S2.1, set the parameter values ​​required for network training according to the SAR flood detection training dataset, including the initial learning rate, optimizer and loss function.

[0026] Further, in step S3, the training and validation include:

[0027] S3.1, Set the maximum number of training iterations for the dual-branch multi-scale fusion network;

[0028] S3.1 Input the dataset into the dual-branch multi-scale fusion network constructed in step 2 for model training and validation, and optimize the network parameters during the training process;

[0029] S3.1 After the network training is completed, save the network parameters that best perform on the validation set.

[0030] Furthermore, in this step S4, the processing of the SAR image to be detected includes:

[0031] S4.1, Obtain SAR image pairs to be detected with the same resolution and number of polarization channels as the training dataset;

[0032] S4.2, Perform synchronous cropping and edge padding according to the training data size to construct the test dataset;

[0033] S4.3, pair the dual-temporal images in the SAR flood detection test dataset and input them into the trained dual-branch multi-scale fusion network to obtain pixel-level detection results;

[0034] S4.4, the detection results of the SAR flood detection test dataset are stitched together using the cropping method described in step S4.2, and the fill-in parts of the images with insufficient size in step S4.2 are deleted, finally obtaining the complete SAR flood detection results.

[0035] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention proposes a dual-branch multi-scale fusion SAR flood intelligent detection method—a dual-branch multi-scale fusion network. The GLAM in the network combines an adaptive factor to dynamically coordinate a self-attention mechanism and a dilated convolution design. The self-attention mechanism enables the model to focus on large-scale flood-affected areas, while the dilated convolution enables the model to focus on smaller flood-affected areas, thus achieving more effective and accurate integration of global and local contextual information. The 3D-Attention module in this invention utilizes an attention mechanism, which interactively adjusts multi-dimensional features sequentially in the temporal, channel, and spatial dimensions, enabling the model to gradually acquire more refined multi-dimensional features. The SIAM in this invention uses grouped convolution and multi-scale convolution design. This module not only reduces feature redundancy but also utilizes multi-level change information to effectively detect irregular flood areas. The dual-branch multi-scale fusion network not only adaptively fuses global and local information but also effectively combines multi-scale change features from differential and semantic information, thereby achieving the goal of accurately detecting large-scale inundated areas and small, fragmented areas. Attached Figure Description

[0036] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0037] Figure 1 This is a flowchart of the method provided in an embodiment of the present invention;

[0038] Figure 2 This is a diagram of the dual-branch multi-scale fusion network structure provided in an embodiment of the present invention;

[0039] Figure 3 This is a GLAM structure diagram provided in an embodiment of the present invention;

[0040] Figure 4 This is a structural diagram of the 3D-Attention module provided in an embodiment of the present invention;

[0041] Figure 5 This is a SAIM structure diagram provided in an embodiment of the present invention;

[0042] Figure 6 This is a visualization of flood detection results for a portion of the test set provided in this embodiment of the invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0044] Example: Figure 1 As shown, a dual-branch, multi-scale fusion SAR flood intelligent detection method includes the following steps:

[0045] Step 1: Create a SAR flood detection dataset;

[0046] Specifically, step 1 includes the following sub-steps:

[0047] Step 1.1: Select SAR image pairs with uniform resolution and acceptable quality, and complete the spatial registration operation;

[0048] Step 1.2: Perform pixel-level annotation on the registered SAR image pairs, clearly dividing the pixels in the SAR image pairs into changed and unchanged regions;

[0049] Step 1.3: Perform synchronous cropping on the SAR image pairs to obtain several pairs of sub-SAR images with the same length and width. Discard images that are too short or too wide.

[0050] Step 1.4: Organize the SAR image pair files and corresponding change label files, select effective samples with a change area ratio of ≥3.5%, and construct a standard SAR flood detection dataset containing training and validation sets;

[0051] Step 2: Design a dual-branch multi-scale fusion network based on a deep learning framework;

[0052] Specifically, step 2 includes the following sub-steps:

[0053] Step 2.1: The dual-branch multi-scale fusion network designed in this invention adopts an encoder-decoder structure, such as... Figure 2 As shown. The encoder part uses the first three stages of ResNet18 as the backbone, and adjusts the stride of the first 7×7 convolutional layer from 2 to 1 to extract multi-scale features. GLAM uses adaptive parameters α and β to dynamically fuse global and local information. The decoder part adopts a dual-branch decoder. The semantic branch interacts with the dual-temporal features through the 3D-Attention module to enhance temporal, channel, and spatial features; the difference branch extracts difference information through element-wise subtraction. Subsequently, ResidualBlock processes the difference and semantic information, and SIAM fuses multi-level features and identifies flood areas of different shapes. Finally, the dual-branch fusion obtains the final flood detection result. The network input is a pair of dual-temporal SAR images, and the output is a flood detection result of the same size as the input SAR image. Before training, it is necessary to set the input of the dual-branch multi-scale fusion network. The input and output sizes of the dual-branch multi-scale fusion network are set according to the length, width, and number of polarization channels of the images in the SAR image pair dataset.

[0054] To effectively perceive both large-scale and small-scale flood areas, this invention utilizes a self-attention mechanism and dilated convolutions with different dilation rates to design GLAM, such as... Figure 3 As shown. This module, through parameter sharing via a twin network, can better adaptively couple dual-temporal features, thereby achieving effective fusion of global and local information to reduce the interference of redundant information on change detection. The GLAM operation formula is as follows:

[0055] Q = F in W Q K = F in W K V = F in W V (1)

[0056] F G =Softmax(Q·K) T )·V

[0057]

[0058]

[0059] In the formula, F in For input features, W Q W K and W V Let be a learnable weight matrix, and Softmax(·) be the activation function. This indicates a convolution operation with a kernel size of 3×3 and an inflation rate of r. `Concat(·)` is the concatenation operation. out For the output features, α and β are learnable weight parameters used to balance the global features F. G and local features F L . contributions.

[0060] To enhance the model's ability to represent water features, this invention designs a 3D-Attention module in the form of an attention mechanism, such as... Figure 4 As shown. This module enables the model to interact with information across different dimensions by concatenating temporal, channel, and spatial attention mechanisms, thereby extracting richer multidimensional feature representations. The operating formula of the 3D-Attention module is as follows:

[0061]

[0062]

[0063]

[0064] In the formula, and The input consists of two temporal features, where R(·) represents the reshaping operation, MLP(·) is a multilayer perceptron, AvgPool(·) and MaxPool(·) represent the average pooling and max pooling operations, respectively, and F... T and F C Features enhanced with time information and channel information, respectively, F 3D These are the output features resulting from information interaction across different dimensions.

[0065] To effectively capture the irregular variations in fragmented flood zones, this invention utilizes grouped convolution and multi-scale convolution to design SIAM, such as... Figure 5 As shown. This module not only reduces feature redundancy but also effectively perceives irregular flood areas by utilizing multi-scale variation information. The SIAM operation formula is as follows:

[0066]

[0067]

[0068] In the formula, Chunk(·) represents grouping, Up(·) represents bilinear interpolation upsampling, and f k,i f represents the k-th feature at the i-th scale. k Let i represent the feature of the k-th group after grouping, where i∈[1,2,3], k∈[1,2,3,4]. This represents a double convolution operation, where each convolution operation includes a 3×3 kernel with an inflation rate of r, normalization, and the ReLU activation function, D. agg This is the result of multi-scale change information fusion.

[0069] Step 2.2: Based on Step 2.1, set the parameter values ​​required for network training according to the SAR flood detection training dataset, mainly including the initial learning rate, optimizer, and loss function;

[0070] Step 3: Train and optimize the dual-branch multi-scale fusion network using the SAR flood detection dataset;

[0071] Specifically, step 3 includes the following sub-steps:

[0072] Step 3.1: Set the maximum number of training iterations for the dual-branch multi-scale fusion network;

[0073] Step 3.2: Input the SAR flood detection dataset into the dual-branch multi-scale fusion network constructed in Step 2 for model training and validation. During the training process, the dual-branch multi-scale fusion network will continuously optimize the network parameters to achieve the best learning effect.

[0074] Step 3.3: After the network training is completed, save the network parameters that perform best in the validation set for model testing;

[0075] Step 4: Use the trained dual-branch multi-scale fusion network for flood detection.

[0076] Specifically, step 4 includes the following sub-steps:

[0077] Step 4.1: Obtain the SAR image pair to be detected. The resolution and number of polarization channels of the SAR image pair must be the same as those of the SAR flood detection training dataset.

[0078] Step 4.2: Simultaneously crop the SAR image pairs according to the length and width of the images in the SAR image training dataset, and sort them to obtain several pairs of sub-SAR images that meet the input requirements of the dual-branch multi-scale fusion network. Fill in the SAR images whose length and width are insufficient to meet the input requirements, and finally form the SAR flood detection test dataset.

[0079] Step 4.3: Input the paired dual-temporal images from the SAR flood detection test dataset into the trained dual-branch multi-scale fusion network;

[0080] Step 4.4: Obtain the SAR flood detection results, i.e., the pixel-level detection results of the SAR flood detection test dataset;

[0081] Step 4.5: Stitch the images from the SAR flood detection test dataset according to the cropping method in Step 4.2, and delete the filler parts of the undersized images in Step 4.2 to obtain the complete SAR flood detection results.

[0082] The effectiveness of the embodiments of the present invention can be further verified through the following experiments:

[0083] 1) Experimental Environment

[0084] The proposed model is implemented in PyTorch and trained using a single NVIDIA RTX 4090 GPU.

[0085] 2) Experiment Content

[0086] The method of this invention, namely a dual-branch multi-scale fusion intelligent SAR flood detection method, is used to perform flood detection on a SAR flood detection dataset. To verify the advantages of this method in flood detection, we compared it with different change detection methods.

[0087] 3) Accuracy Evaluation

[0088] To quantitatively evaluate the effectiveness of this invention, the overall accuracy (OA), F1 score (F1), kappa coefficient (KC), and intersection-over-union ratio (IoU) are selected as evaluation indicators.

[0089] Table 1. Accuracy Evaluation of Flood Detection Results between the Method of the Present Invention and Other Methods

[0090]

[0091] 4) Analysis of experimental results

[0092] Figure 6 The experimental results in Table 1 show that, compared with other methods, the dual-branch multi-scale fusion network proposed in this invention provides flood detection results that are closer to the true value map of the flooded area, can better distinguish between inundated and non-inundated areas, significantly reduces false alarms and missed detections, and can effectively detect small-scale inundated areas. Furthermore, its quantitative evaluation indicators also outperform those of other methods by a significant margin. Therefore, the method proposed in this invention significantly improves flood detection performance in multiple aspects.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dual-branch, multi-scale fusion SAR flood intelligent detection method, characterized in that, Includes the following steps: S1, Process SAR images and construct a SAR flood detection dataset; S2, construct a dual-branch multi-scale fusion network including a twin encoder and a dual-branch decoder; S3, Use the SAR flood detection dataset to train and validate the network to obtain a trained flood detection model; S4, input the dual-temporal SAR image pair to be detected into the flood detection model, and output a flood detection result image with the same size as the input image; In step S2, the dual-branch multi-scale fusion network includes: S2.1, Construct a dual-branch multi-scale fusion network, wherein: The GLAM module in the twin encoder dynamically fuses global and local information using adaptive parameters α and β. The dual-branch decoder includes: The 3D-Attention module in the dual-branch decoder is used to enhance the spatiotemporal interaction of dual-temporal features; The SIAM module in the dual-branch decoder is used to fuse multi-level features and identify flood regions in the semantic branch and the difference branch; S2.2: Based on step S2.1, set the parameter values ​​required for network training according to the SAR flood detection training dataset, including the initial learning rate, optimizer and loss function; The GLAM module, through parameter sharing via a twin network, can better adaptively couple dual-temporal features, thereby achieving effective fusion of global and local information and reducing the interference of redundant information on change detection. The GLAM operation formula is as follows: In the formula, As input features, , and The weight matrix is ​​a learnable matrix. For activation function, This represents a convolution operation with a kernel size of 3×3 and an inflation rate of r. For splicing operations, For the output features, α and β are learnable weight parameters used to balance the global features. and local features Contributions; The 3D-Attention module, by concatenating temporal, channel, and spatial attention mechanisms, enables the model to interact with information across different dimensions, thereby extracting richer multidimensional feature representations. The operation formula of the 3D-Attention module is as follows: In the formula, and For the input dual-temporal features, This indicates a reshaping operation. It is a multilayer perceptron. and These represent average pooling and max pooling operations, respectively. and These are the features enhanced with time information and those enhanced with channel information, respectively. These are the output features resulting from information interaction across different dimensions; The SIAM module can reduce feature redundancy and effectively perceive irregular flood areas by utilizing multi-scale change information. The SIAM operation formula is as follows: In the formula, Indicates grouping, This indicates bilinear interpolation upsampling. Let k represent the feature group at the i-th scale. Let i represent the feature of the k-th group after grouping, where i∈[1,2,3], k∈[1,2,3,4]. This represents a double convolution operation, where each convolution operation includes a 3×3 kernel with an inflation rate of r, normalization, and the ReLU activation function. This is the result of multi-scale change information fusion.

2. The dual-branch multi-scale fusion SAR flood intelligent detection method according to claim 1, characterized in that, In step S1, constructing the SAR flood detection dataset includes: S1.1 Select SAR image pairs with uniform resolution and complete spatial registration; S1.2, perform pixel-level annotation on the registered SAR image pairs, including flood change areas and non-change areas; S1.3, Perform synchronous cropping on SAR image pairs to obtain sub-image pairs of the same size and discard images that are too small; S1.4, select samples with a change region ratio of ≥3.5% to construct a standard dataset containing training and validation sets.

3. The dual-branch multi-scale fusion SAR flood intelligent detection method according to claim 1, characterized in that, In step S3, the training and validation include: S3.1, Set the maximum number of training iterations for the dual-branch multi-scale fusion network; S3.1 Input the dataset into the dual-branch multi-scale fusion network constructed in step 2 for model training and validation, and optimize the network parameters during the training process; S3.1 After the network training is completed, save the network parameters that best perform on the validation set.

4. The dual-branch multi-scale fusion SAR flood intelligent detection method according to claim 1, characterized in that, In step S4, the processing of the SAR image to be detected includes: S4.1, Obtain SAR image pairs to be detected with the same resolution and number of polarization channels as the training dataset; S4.2, Perform synchronous cropping and edge padding according to the training data size to construct the test dataset; S4.3, pair the dual-temporal images in the SAR flood detection test dataset and input them into the trained dual-branch multi-scale fusion network to obtain pixel-level detection results; S4.4, the detection results of the SAR flood detection test dataset are stitched together using the cropping method described in step S4.2, and the fill-in parts of the images with insufficient size in step S4.2 are deleted, finally obtaining the complete SAR flood detection results.

Images