A priori knowledge-based multi-temporal SAR image flood mapping method and device
By employing a multi-temporal SAR image flood mapping method based on prior knowledge, and utilizing a pre-trained water feature extraction and feature fusion module, water change detection can be directly performed. This solves the problems of error accumulation and difficulty in obtaining training datasets in existing technologies, and achieves high-precision detection of flood inundation range.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- WUHAN UNIV
- Filing Date
- 2023-02-27
- Publication Date
- 2026-05-15
AI Technical Summary
Existing methods for detecting flood inundation ranges in SAR images suffer from problems such as error accumulation, insufficient robustness, and difficulty in obtaining training datasets, resulting in insufficient detection accuracy and timeliness.
A flood mapping method based on prior knowledge using multi-temporal SAR images is adopted. Through a pre-trained SAR image water feature extraction module, a two-branch twin feature fusion module, and a change feature convolution module, prior knowledge of the water body range before and after the disaster is directly used to detect water body changes. This enables cross-task transfer, reduces the number of training samples required, and improves detection accuracy.
It significantly improves the accuracy and efficiency of flood inundation range detection, is suitable for complex scenarios, reduces the requirement for the number of training samples, and has good transferability and applicability.
Smart Images
Figure CN116152305B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of SAR image flood inundation range detection technology, and in particular to a method and apparatus for multi-temporal SAR image flood mapping based on prior knowledge. Background Technology
[0002] Due to the suddenness, randomness, and complexity of floods, post-disaster emergency rescue often faces the dilemma of not knowing "where the disaster occurred, where to provide relief, and how effective the relief efforts are." Remote sensing Earth observation technology has advantages such as a wide field of view, rapid information acquisition, and short cycle time. Especially in rainy, foggy, or traffic-disrupted weather conditions, SAR imagery acquired by synthetic aperture radar satellites has become a crucial data source for flood disaster emergency monitoring and loss assessment. Therefore, using multi-temporal SAR imagery to quickly and accurately estimate the inundation area of flood disasters is of great significance for disaster emergency rescue and post-disaster loss assessment.
[0003] Currently, the methods for extracting flood inundation ranges using SAR image data mainly fall into the following two categories:
[0004] The first category is based on the water body extent results extracted from SAR images of different time phases, and the flood extent is detected by overlay comparison. Among them, the water body extraction methods include thresholding water body extraction and classification water body extraction.
[0005] The second category is change detection methods based on dual-temporal SAR images. These methods utilize the similarity features between dual-temporal SAR images to directly detect areas where water bodies have changed, using either supervised or unsupervised methods.
[0006] The first type of water change detection method based on image classification involves complex steps. Since the accuracy of flood inundation range detection depends on the accuracy of single-phase water body extraction, errors accumulate in the detection results. Furthermore, due to the complex background of flood-prone areas and interference from SAR image speckle noise, the adaptive thresholding method based on the assumption of a bimodal distribution of SAR image histograms is unsuitable for the actual situation of SAR images, resulting in insufficient robustness and applicability. The second type of change detection method based on unsupervised dual-phase SAR image change detection mainly relies on manually defined dual-phase SAR image similarity metrics. It uses thresholding to classify similarity features and extract changed areas. However, the similarity features selected by this type of method, such as difference, ratio, or logarithmic ratio, are only applicable to image areas with a single scene. When applied to complex flood-prone areas, the accuracy of flood inundation range detection is insufficient. Additionally, although supervised change detection methods based on dual-phase SAR imagery extract richer deep semantic features and have stronger flood detection robustness, these methods are limited by the difficulty in obtaining high-quality training datasets, failing to meet the needs of practical flood detection applications.
[0007] Flood-affected areas often present complex and variable landscapes. On the one hand, SAR images are susceptible to inherent image noise interference; on the other hand, the imaging mechanism of SAR images makes it difficult to distinguish between ground targets with low roughness, such as roads, and water bodies. In addition, deep learning models for detecting changes in multi-temporal SAR images used to detect flood inundation areas are limited by the difficulty in obtaining training samples, and cannot guarantee the accuracy and timeliness of flood inundation area detection. Summary of the Invention
[0008] To address the problems mentioned in the background, this invention proposes a multi-temporal SAR image flood mapping method based on prior knowledge. The proposed SAR image flood inundation range detection method innovatively links SAR image change detection with cross-task transfer learning. SAR image flood inundation range detection is essentially the detection of changes in water body extent before and after a flood. Therefore, the SAR image flood inundation detection problem generally consists of two steps: water body extraction and water body change detection. The water body change detection task relies on prior knowledge of the water body extent acquired in the previous step, which leads to error accumulation in the water body change detection results. This method aims to solve the error accumulation problem in existing technologies that use a classification-then-change detection approach, where the flood inundation detection accuracy is limited by the accuracy of water body extraction. It designs a flood water body change recognition method from the semantic feature level of prior knowledge transfer. Therefore, this invention differs from strategies that indirectly utilize pre- and post-disaster water body extent for water body change detection. Instead, it is based on a pre-trained network model with water body feature extraction capabilities. By transferring the prior knowledge of the water body extraction model to the water body change detection task through cross-task transfer, it enables the direct use of deep semantic features of water bodies in two-temporal SAR images to model flood water body change information and extract flood inundation range.
[0009] The technical solution of this invention is a flood mapping method based on prior knowledge using multi-temporal SAR images, comprising the following steps:
[0010] Step 1: Obtain two-phase registered SAR images before and after the flood disaster;
[0011] Step 2: Input the two-phase SAR images into the trained multi-phase SAR image flood mapping network based on prior knowledge, and output water body change characteristic images through the multi-phase SAR image flood mapping network based on prior knowledge.
[0012] The multi-temporal SAR image flood mapping network based on prior knowledge includes a pre-trained SAR image water feature extraction module, a two-branch twin feature fusion module, and a change feature convolution module. Pre-disaster and post-disaster SAR images are input into the SAR water feature extraction module, which outputs two-temporal water semantic feature images. The two-temporal water semantic feature images are input into the two-branch twin feature fusion module, which outputs an initial water feature fused image. The initial water feature fused image is input into the change feature convolution module, which outputs a water change feature image.
[0013] Step 3: Extract binary images from water body change feature images based on thresholds to obtain the flood inundation range detection results.
[0014] Moreover, the process of obtaining dual-temporal SAR images registered before and after the flood in step 1 is as follows: obtain SAR images of the same area in the disaster-stricken region before and after the flood, and perform preprocessing such as radiometric correction, geometric correction, and filtering. Then, use the corresponding points of the two temporal images to perform geometric registration so that the pixel coordinates of the two temporal images match each other.
[0015] Moreover, the pre-trained SAR water feature extraction module includes a multi-level encoding module with several convolutional layers and pooling layers, and a multi-level decoding module with several convolutional layers, up-sampling layers, fusion layers and channel attention layers.
[0016] In the multi-level coding module, the original SAR image is first downsampled through a 3×3 double convolution operation, followed by several max pooling operations and 3×3 double convolution operations to obtain depth-coded semantic features at different scales, which can extract important water body semantic features. In the multi-level decoding module, the deep semantic features are upsampled several times in accordance with the coding module to restore spatial information, and are then fused with the same-level coding features of the coding module through cascaded feature fusion, 3×3 double convolution, and channel attention mechanism processing to obtain a water body semantic feature image with multi-scale depth features after channel weight adjustment and restored spatial information.
[0017] The 3×3 double convolution operation includes a convolutional layer + a normalization layer + a convolutional layer + a normalization layer + a modified linear unit layer.
[0018] Furthermore, the training process of the pre-trained SAR water feature extraction module is as follows:
[0019] A training sample set for SAR image water body extraction is obtained, wherein the training sample set includes several training samples, and the training samples include SAR image data and SAR image water body range annotation images corresponding to the SAR image data; the training data is input into a preset initial SAR water body feature extraction network, and the preset initial SAR water body feature extraction network outputs a predicted water body extraction result image; wherein the preset initial SAR water body feature extraction network consists of the SAR water body feature extraction module and a 1×1 convolutional layer; a loss function is obtained based on the predicted water body extraction result image and the SAR image water body range annotation image; based on the loss function, the initial SAR water body feature extraction network is trained to obtain a pre-trained SAR water body feature extraction module.
[0020] Moreover, the dual-branch twin feature fusion module is a vector splicing layer. The vector splicing layer is used to fuse the two-phase deep semantic features of water obtained by the pre-trained SAR water feature extraction module, and to provide the initial water feature fusion image for the variable feature convolution module. The specific operation is shown in Equation (1). The operation of the vector splicing layer is a feature splicing in the form of Concatenate.
[0021] F_union=Concat(F_pro,F_post)(1)
[0022] Wherein, F_union represents the output initial water feature fusion image. Concat represents the vector concatenation operation, and F_pro represents the pre-temporal deep semantic features of the water body output by the pre-trained SAR water feature extraction module. F_post represents the post-temporal deep semantic features of the water body output by the pre-trained SAR water feature extraction module. .
[0023] Moreover, the change feature convolution module consists of several convolutional layers, which extract change information from the initial water body feature fusion image to obtain a water body change feature image with enhanced change information. The specific operation is shown in Equation (2), where the first two convolution operations are convolutional layers with a window size of 3×3, and the third convolution operation is a convolutional layer with a window size of 1×1.
[0024] Fout=Conv(RL(BN(Conv (RL(BN(Conv (F_union)))))))(2)
[0025] Wherein, Fout represents the output water body change feature image, Conv represents the convolution operation, BN represents the batch normalization operation, RL represents the modified linear unit operation, and F_union represents the initial water body feature fusion image output by the dual-branch twin feature fusion module.
[0026] Furthermore, the training process of the multi-temporal SAR image flood mapping network based on prior knowledge is as follows:
[0027] Obtain a SAR flood inundation training dataset, wherein the SAR flood inundation training dataset includes a small number of training samples, the training samples include two time-phase training image data before and after the disaster, and a ground truth image of the inundation range corresponding to the training image data;
[0028] The flood inundation training dataset is input into a pre-defined multi-temporal SAR image flood mapping network model based on prior knowledge, and the initial network model outputs the predicted flood inundation range detection result image.
[0029] The loss function is obtained based on the predicted flood inundation range detection result image and the inundation range reference true image;
[0030] Based on the loss function, the initial network model is trained to obtain a multi-temporal SAR image flood mapping network based on prior knowledge.
[0031] Moreover, the loss function is implemented by calculating a binary cross-entropy loss function based on the predicted flood inundation range detection result image and the inundation range reference true image.
[0032] Moreover, the implementation process of step 3 is as follows: the pixel value range of the flood change feature map is [0, 1], the threshold for judging the flood range is 0.5, the pixels greater than the threshold are classified as flood flooded, and the pixels less than the threshold are classified as unchanged areas, thereby obtaining the flood flood range detection result: flooded or unchanged.
[0033] Secondly, the present invention also provides a multi-temporal SAR image flood mapping device based on prior knowledge, comprising the following units:
[0034] The SAR image acquisition unit is used to acquire SAR images from two time phases: before and after the flood disaster.
[0035] The flood inundation intensity image acquisition unit is used to input two temporal SAR images, one before and one after the flood, into a pre-trained multi-temporal SAR image flood mapping network based on prior knowledge, and output images of water body change characteristics through the multi-temporal SAR image flood mapping network based on prior knowledge.
[0036] The multi-temporal SAR image flood mapping network based on prior knowledge includes a pre-trained SAR water feature extraction module, a two-branch twin feature fusion module, and a change feature convolution module. It inputs pre-flood and post-flood SAR images into the pre-trained SAR water feature extraction module, outputting a two-temporal water semantic feature image; inputs the two-temporal water semantic feature image into the two-branch twin feature fusion module, outputting an initial water feature fused image; and inputs the initial water feature fused image into the change feature convolution module, outputting a water change feature image.
[0037] The flood mapping unit is used to classify water body change feature images into binary images using the threshold method, thereby obtaining the flood inundation range identification results.
[0038] Thirdly, the present invention also provides a smart terminal, including at least one memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors, the one or more programs including a method for performing a priori knowledge-based multi-temporal SAR image flood mapping as described in any of the preceding claims.
[0039] Fourthly, the present invention also provides a non-transitory computer-readable storage medium, wherein when the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is able to perform the multi-temporal SAR image flood mapping method based on prior knowledge as described in any one of the above.
[0040] The beneficial effects of this invention are as follows: When the flood inundation annotation information of SAR images is limited, by transferring the pre-trained module for water body feature extraction with prior knowledge to the flood mapping task through task transfer, the number of training samples required for the flood inundation range detection model can be greatly reduced. With a small number of training samples, the training efficiency of the flood inundation range detection model can be obtained faster, and the accuracy of flood inundation range detection in SAR images can be significantly improved. Furthermore, this technique has good transferability and applicability. Attached Figure Description
[0041] To more clearly illustrate the technical methods in the specific embodiments of this application or the prior art, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0042] Figure 1 This is a schematic diagram of the method flow proposed in this invention;
[0043] Figure 2A schematic diagram of the overall framework of the multi-temporal SAR image flood mapping method based on prior knowledge provided in an embodiment of the present invention;
[0044] Figure 3 A functional structure diagram of the multi-temporal SAR image flood mapping device based on prior knowledge provided in an embodiment of the present invention;
[0045] Figure 4 This is a schematic diagram of the detection results of flood inundation range in an embodiment of the present invention. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0047] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to identify selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0048] First, a brief introduction to the design concept of the embodiments of this application will be given.
[0049] To address the technical shortcomings of existing SAR image flood inundation range detection methods, this application provides a multi-temporal SAR image flood mapping method based on prior knowledge. This method extracts water body change areas before and after floods using an inundation detection model: a pre-trained SAR image water body feature extraction module, consisting of two branches with weight-sharing twin network structures, extracts deep semantic features of water bodies from pre- and post-flood SAR images; a dual-branch twin feature fusion module fuses the deep water body features from the two phases into a feature set containing information on water body changes in both phases; a change feature convolution module generates a flood inundation feature map from the initial fused water body feature image; and the flood inundation feature map is then binary-classified to ultimately achieve flood inundation range detection.
[0050] The method proposed in this application, under the strategy of task transfer, not only effectively reduces the number of labeled samples required for model training, but also avoids the error accumulation problem caused by traditional methods that first identify water bodies and then identify water body changes. This significantly improves the accuracy of flood inundation range detection in dual-temporal SAR images and can be applied to flood detection tasks of different SAR satellite images.
[0051] The technical solutions provided in the embodiments of this application will be described in detail below.
[0052] The flowchart of the technical solution of this invention is as follows: Figure 1 As shown in the embodiment, a flood mapping method based on prior knowledge using multi-temporal SAR images specifically includes the following steps:
[0053] Step S100: Acquire SAR images registered before and after the flood disaster;
[0054] Specifically, the SAR images acquired before and after the floods came from the same satellite and the same area at different times. By performing preprocessing such as radiometric correction, geographic correction, and filtering on the original SAR images, matching points were selected manually or using matching algorithms. The matching algorithms were then used to perform geometric registration between the two temporal images, ensuring that the pixel coordinates of the two temporal images corresponded one-to-one, thus preparing data for subsequent detection of the flood inundation range.
[0055] Once the registered pre-disaster and post-disaster SAR images are obtained, the following can be executed: Figure 1 The steps are as follows:
[0056] S200: Input the two-phase SAR images into the trained multi-phase SAR image flood mapping network based on prior knowledge, and output water body change characteristic images through the multi-phase SAR image flood mapping network based on prior knowledge.
[0057] See Figure 2 In this embodiment, the multi-temporal SAR image flood mapping network based on prior knowledge includes a pre-trained SAR image water feature extraction module, a two-branch twin feature fusion module, and a change feature convolution module.
[0058] The pre-trained SAR water feature extraction module possesses prior knowledge about the semantic features of water bodies in SAR imagery. This knowledge can assist the model in learning the deep semantic features of water bodies in flood inundation range detection tasks, which can then be used for subsequent two-temporal water feature fusion and water change feature extraction. In one implementation, the pre-trained SAR water feature extraction module includes a multi-level encoding module with several convolutional and pooling layers, and a multi-level decoding module with several convolutional layers, upsampling layers, fusion layers, and channel attention layers. The detailed operation of each module is shown in Table 1. In the multi-level coding module, the original SAR image is first downsampled through a 3×3 double convolution operation, followed by several max pooling operations and 3×3 double convolution operations to obtain depth-coded semantic features at different scales, which can extract important water body semantic features. In the multi-level decoding module, the deep semantic features are upsampled several times to restore spatial information, and then fused with the same-level coding features of the multi-level coding module through cascaded feature fusion, 3×3 double convolution, and channel attention mechanism processing to obtain a water body semantic feature image with multi-scale depth features after channel weight adjustment and restored spatial information.
[0059] Table 1 Water Feature Extraction Module
[0060]
[0061] In the table, H and W represent the height and width of the input data, MaxPool represents the maximum pooling layer, Up represents the upsampling layer, Concat represents the feature fusion of upsampled features and the same-level encoded features obtained by the encoding module, DoubleConv represents the convolutional layer + normalization layer + convolutional layer + normalization layer + corrected linear unit layer, and CA represents the channel attention mechanism layer.
[0062] In one embodiment, the training process of the pre-trained SAR water feature extraction module includes the following steps: obtaining a SAR image water feature extraction training sample set, wherein the SAR image water feature extraction training sample set includes several training samples, the training samples include SAR image data and SAR image water feature range annotation images corresponding to the SAR image data; inputting the training data into a preset initial SAR water feature extraction network, and outputting a predicted water feature extraction result image through the initial SAR water feature extraction network; wherein the preset initial SAR water feature extraction network consists of the SAR water feature extraction module and a 1×1 convolutional layer; obtaining a loss function based on the predicted water feature extraction result image and the SAR image water feature range annotation image; and training the initial network model based on the loss function to obtain the pre-trained SAR water feature extraction module.
[0063] In this embodiment, the acquired SAR water body extraction dataset is divided into a training dataset and a validation dataset. The training samples include SAR image data and corresponding SAR image water body extent annotations. These annotations are obtained through manual visual interpretation, with pixel 0 representing non-water bodies and pixel 255 representing water bodies. This results in 100 pairs of 512×512 pixel training datasets and 4 pairs of 512×512 pixel validation datasets. The SAR water body extraction dataset is input into a pre-defined initial SAR water body feature extraction network, which outputs a predicted water body extraction result image. Based on the predicted water body extraction result image and the SAR image water body extent annotations, a loss function is obtained, specifically a binary cross-entropy loss. The initial network model is then trained based on this loss function to obtain a pre-trained SAR water body feature extraction module.
[0064] The dual-branch twin feature fusion module is a vector concatenation layer, such as... Figure 2 As shown. The vector stitching layer is used to fuse the two-phase deep semantic features of water obtained by the pre-trained SAR water feature extraction module, providing the initial water feature fusion image for the variable feature convolution module. In one implementation, the specific operation of the dual-branch twin feature fusion module is shown in Equation (1), and the operation of the vector stitching layer is a concatenate-style feature stitching.
[0065] F_union=Concat(F_pro,F_post)(1)
[0066] Wherein, F_union represents the output initial water feature fusion image. Concat represents the vector concatenation operation, and F_pro represents the pre-temporal deep semantic features of the water body output by the pre-trained SAR water feature extraction module. F_post represents the post-temporal deep semantic features of the water body output by the pre-trained SAR water feature extraction module. .
[0067] The variable feature convolution module consists of several convolutional layers. For example... Figure 2 As shown. The change feature convolution module extracts change information from the initial water feature fusion image to obtain a water change feature image with enhanced change information. In one implementation, the specific operation of the change feature convolution module is shown in equation (2), where the first two convolution operations are convolutional layers with a window size of 3×3, and the third convolution operation is a convolutional layer with a window size of 1×1.
[0068] Fout=Conv(RL(BN(Conv (RL(BN(Conv (F_union)))))))(2)
[0069] Wherein, Fout represents the output water body change feature image, Conv represents the convolution operation, BN represents the batch normalization operation, RL represents the modified linear unit operation, and F_union represents the initial water body feature fusion image output by the dual-branch twin feature fusion module.
[0070] To obtain images of water body change characteristics, the processing of a multi-temporal SAR image flood mapping network based on prior knowledge can be considered to include the following steps:
[0071] S201. Input the SAR images registered in two time phases into the pre-trained SAR water body feature extraction module, and output the two-time phase SAR water body deep semantic feature images through the pre-trained SAR water body feature extraction module.
[0072] S202. Input the two-temporal SAR water body deep semantic feature images into the dual-branch twin feature fusion module, and output the initial water body feature fusion image through the dual-branch twin feature fusion module;
[0073] S203. Input the initial water feature fusion image into the change feature convolution module, and output the water change feature image through the change feature convolution module.
[0074] In one embodiment, the training process of the SAR image flood inundation detection network based on task transfer includes the following steps: acquiring a flood inundation training dataset, wherein the flood inundation training dataset includes a small number of training samples, the training samples include pre-disaster and post-disaster two-phase training image data, and the inundation range reference ground image corresponding to the training image data; inputting the flood inundation training dataset into a preset initial flood inundation range detection network model, and outputting a predicted flood inundation range detection result image through the initial network model; obtaining a loss function based on the predicted flood inundation range detection result image and the inundation range reference ground image; and training the initial network model based on the loss function to obtain a multi-temporal SAR image flood mapping network based on prior knowledge.
[0075] In this embodiment, the acquired flood inundation training dataset is divided into a training dataset, a validation dataset, and a test dataset in a 4:4:1 ratio. The training samples include pre-disaster and post-disaster training image data, and corresponding ground truth inundation range images. The ground truth inundation range images are obtained through SAR water body extraction combined with manual visual interpretation and annotation. Pixel 0 represents unchanged water bodies, and pixel 255 represents changed water bodies. This results in four pairs of 512×512 pixel training datasets, four pairs of 512×512 pixel validation datasets, and one pair of 512×512 pixel test datasets. The flood inundation training dataset is input into a pre-defined initial task-transfer based SAR image flood inundation detection network. The task-transfer based SAR image flood inundation detection network outputs a predicted water body change feature image. Based on the predicted water body change feature image and the reference ground truth image of the inundation range, a loss function is obtained, wherein the loss function is a binary cross-entropy loss. Based on the loss function, the initial network model is trained to obtain a multi-temporal SAR image flood mapping network based on prior knowledge.
[0076] After obtaining images of water body change characteristics, perform the following steps: Figure 1 The following steps are shown:
[0077] S300: Based on threshold extraction of water body change feature images, binary images are obtained to obtain the flood inundation range detection results.
[0078] Specifically, the pixel value range of the flood change feature map is [0, 1]. The threshold for judging the flood range is 0.5. Pixels with values greater than the threshold are classified as flood-inundated areas, and pixels with values less than the threshold are classified as unchanged areas, thus obtaining the flood-inundated range detection result: flooded or unchanged.
[0079] pass Figure 3 By examining the flood inundation detection results of SAR images before and after floods in different regions, it can be found that the flood inundation range detection results obtained by the embodiments of the present invention are closest to the true reference value.
[0080] Based on the above embodiments, this application provides a multi-temporal SAR image flood mapping device 400 based on prior knowledge, such as... Figure 3 As shown, the device includes at least:
[0081] SAR image acquisition unit 401 is used to acquire SAR remote sensing images of two phases before and after the flood disaster.
[0082] The flood inundation intensity image acquisition unit 402 is used to input two-phase SAR images before and after the flood into a trained multi-phase SAR image flood mapping network based on prior knowledge, and output water body change characteristic images through the multi-phase SAR image flood mapping network based on prior knowledge.
[0083] The multi-temporal SAR image flood mapping network based on prior knowledge includes a pre-trained SAR water feature extraction module, a two-branch twin feature fusion module, and a change feature convolution module. It inputs pre-flood and post-flood SAR images into the pre-trained SAR water feature extraction module, outputting a two-temporal water semantic feature image; inputs the two-temporal water semantic feature image into the two-branch twin feature fusion module, outputting an initial water feature fused image; and inputs the initial water feature fused image into the change feature convolution module, outputting a water change feature image.
[0084] The flood mapping unit 403 is used to classify water body change feature images into binary images using the threshold method, and obtain the flood inundation range identification results.
[0085] It should be noted that the principle of the multi-temporal SAR image flood mapping device based on prior knowledge provided in this application embodiment to solve the technical problem is similar to the multi-temporal SAR image flood mapping method based on prior knowledge provided in this application embodiment. Therefore, the implementation of the multi-temporal SAR image flood mapping device based on prior knowledge provided in this application embodiment can refer to the implementation of the multi-temporal SAR image flood mapping method based on prior knowledge provided in this application embodiment, and the repeated parts will not be described again.
[0086] This application also provides a smart terminal, including at least one memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors. The one or more programs include a method for performing a priori knowledge-based multi-temporal SAR image flood mapping as described in any of the above embodiments.
[0087] This application also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the multi-temporal SAR image flood mapping method based on prior knowledge provided in this application.
[0088] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0089] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited thereto. Any simple variations or equivalent substitutions of the technical methods that can be obviously obtained by those skilled in the art within the scope of the technology disclosed in the present invention shall fall within the scope of protection of the present invention.
Claims
1. A method for flood mapping based on prior knowledge using multi-temporal SAR images, characterized in that, Includes the following steps: Step 1: Obtain two-phase registered SAR images before and after the flood disaster; Step 2: Input the two-phase SAR images into the trained multi-phase SAR image flood mapping network based on prior knowledge, and output water body change characteristic images through the multi-phase SAR image flood mapping network based on prior knowledge. The prior knowledge-based multi-temporal SAR image flood mapping network includes a pre-trained SAR water body feature extraction module, a two-branch twin feature fusion module, and a change feature convolution module. Pre-disaster and post-disaster SAR images are input into the SAR water body feature extraction module, which outputs two-temporal water body semantic feature images. The two-temporal water body semantic feature images are input into the two-branch twin feature fusion module, which outputs an initial water body feature fused image. The initial water body feature fused image is input into the change feature convolution module, which outputs a water body change feature image. The dual-branch twin feature fusion module is a vector splicing layer. The vector splicing layer is used to fuse the two-phase deep semantic features of water obtained by the pre-trained SAR water feature extraction module, and to provide the initial water feature fusion image for the variable feature convolution module. The specific operation is shown in Equation (1). The operation of the vector splicing layer is a feature splicing in the form of concatenation. F_union=Concat(F_pro,F_post)(1) Wherein, F_union represents the output initial water feature fusion image. Concat represents the vector concatenation operation, and F_pro represents the pre-temporal deep semantic features of the water body output by the pre-trained SAR water feature extraction module. F_post represents the post-temporal deep semantic features of the water body output by the pre-trained SAR water feature extraction module. ; The change feature convolution module consists of several convolutional layers. The change feature convolution module extracts change information from the initial water body feature fusion image to obtain a water body change feature image with enhanced change information. The specific operation is shown in Equation (2), where the first two convolution operations are convolutional layers with a window size of 3×3, and the third convolution operation is a convolutional layer with a window size of 1×1. Fout=Conv(RL(BN(Conv (RL(BN(Conv (F_union)))))))(2) Wherein, Fout represents the output water body change feature image, Conv represents the convolution operation, BN represents the batch normalization operation, RL represents the modified linear unit operation, and F_union represents the initial water body feature fusion image output by the dual-branch twin feature fusion module. Step 3: Extract binary images from water body change feature images based on thresholds to obtain the flood inundation range detection results.
2. The flood mapping method based on prior knowledge using multi-temporal SAR images according to claim 1, characterized in that: The process of obtaining pre- and post-flood registered dual-temporal SAR images in step 1 is as follows: acquire preprocessed SAR images of the same area in the disaster-stricken region before and after the flood, and perform radiometric correction, geometric correction, and filtering preprocessing. Then, use the corresponding points of the two temporal images to perform geometric registration so that the pixel coordinates of the two temporal images match each other.
3. The flood mapping method based on prior knowledge using multi-temporal SAR images according to claim 1, characterized in that: The pre-trained SAR water feature extraction module includes a multi-level encoding module with several convolutional layers and pooling layers, and a multi-level decoding module with several convolutional layers, upsampling layers, fusion layers and channel attention layers. In the multi-level coding module, the original SAR image is first downsampled through a 3×3 double convolution operation, followed by several max pooling operations and 3×3 double convolution operations to obtain depth-coded semantic features at different scales, which can extract important water body semantic features. In the multi-level decoding module, the deep semantic features are upsampled several times according to the coding module to restore spatial information, and then cascaded feature fusion, 3×3 double convolution, and channel attention mechanism processing are performed with the same-level coding features of the coding module to obtain a water body semantic feature image with multi-scale depth features after channel weight adjustment and restored spatial information.
4. The flood mapping method based on prior knowledge using multi-temporal SAR images according to claim 3, characterized in that: The 3×3 double convolution operation consists of a convolutional layer + a normalization layer + a convolutional layer + a normalization layer + a modified linear unit layer.
5. The flood mapping method based on prior knowledge using multi-temporal SAR images according to claim 3, characterized in that: The training process for the pre-trained SAR water feature extraction module is as follows: A training sample set for SAR image water body extraction is obtained, wherein the training sample set includes several training samples, and the training samples include SAR image data and SAR image water body range annotation images corresponding to the SAR image data; the training samples are input into a preset initial SAR water body feature extraction network, and the preset initial SAR water body feature extraction network outputs a predicted water body extraction result image; wherein the preset initial SAR water body feature extraction network consists of the SAR water body feature extraction module and a 1×1 convolutional layer; a loss function is obtained based on the predicted water body extraction result image and the SAR image water body range annotation image; based on the loss function, the initial SAR water body feature extraction network is trained to obtain a pre-trained SAR water body feature extraction module.
6. The flood mapping method based on prior knowledge using multi-temporal SAR images according to claim 1, characterized in that: The training process of the multi-temporal SAR image flood mapping network based on prior knowledge is as follows. Obtain a SAR flood inundation training dataset, wherein the SAR flood inundation training dataset includes a small number of training samples, the training samples include two time-phase training image data before and after the disaster, and a ground truth image of the inundation range corresponding to the training image data; The flood inundation training dataset is input into a preset multi-temporal SAR image flood mapping network model based on prior knowledge, and the network model outputs the predicted flood inundation range detection result image. The loss function is obtained based on the predicted flood inundation range detection result image and the inundation range reference true image; Based on the loss function, the network model is trained to obtain a multi-temporal SAR image flood mapping network based on prior knowledge.
7. The flood mapping method based on prior knowledge using multi-temporal SAR images according to claim 1, characterized in that: Step 3 is implemented as follows: the pixel value range of the flood change feature map is [0, 1]. The threshold for judging the flood range is 0.
5. Pixels with values greater than the threshold are classified as flood flooded, and pixels with values less than the threshold are classified as unchanged areas, thus obtaining the flood flood range detection result: flooded or unchanged.
8. A flood mapping device based on prior knowledge using multi-temporal SAR images, implementing the method described in any one of claims 1-7, characterized in that, Includes the following units: The SAR image acquisition unit is used to acquire SAR images from two time phases: before and after the flood disaster. The flood inundation intensity image acquisition unit is used to input two temporal SAR images, one before and one after the flood, into a pre-trained multi-temporal SAR image flood mapping network based on prior knowledge, and output images of water body change characteristics through the multi-temporal SAR image flood mapping network based on prior knowledge. The multi-temporal SAR image flood mapping network based on prior knowledge includes a pre-trained SAR water feature extraction module, a two-branch twin feature fusion module, and a change feature convolution module. It inputs pre-flood and post-flood SAR images into the pre-trained SAR water feature extraction module, outputting a two-temporal water semantic feature image; inputs the two-temporal water semantic feature image into the two-branch twin feature fusion module, outputting an initial water feature fused image; and inputs the initial water feature fused image into the change feature convolution module, outputting a water change feature image. The flood mapping unit is used to classify water body change feature images into binary images using the threshold method, thereby obtaining the flood inundation range identification results.