A method and device for identifying a flood inundation area based on S-band SAR data and a visual basic model

Abstract

The application discloses a flood inundated area identification method and device based on S-band SAR data and a visual basic model, and belongs to the technical field of remote sensing image processing. The method comprises the following steps: acquiring S-band SAR data and auxiliary data; performing preprocessing on the S-band SAR data and the auxiliary data; constructing a water body sample set based on the preprocessed S-band SAR data and the auxiliary data; constructing an SSUNet model and training the model based on the water body sample set; inputting the preprocessed S-band SAR data to be measured into the trained SSUNet model to obtain an identification result of a flood area. The application fully utilizes the strong capability of the visual model in image feature extraction and the fine advantage of the windmill convolution module in detail capture, effectively enhances the identification capability of small water bodies in the S-band SAR image, and further improves the flood monitoring precision based on the S-band SAR satellite.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing image processing, disaster monitoring and deep learning, and particularly relates to a flood inundated area identification method and device based on S-band SAR data and a visual basic model. BACKGROUND

[0002] Floods, as a common and severe natural disaster, pose a significant threat to human life and property safety, social and economic development, and the ecological environment. Rapid and accurate acquisition of flood inundated area information is a core step in responding to disasters, conducting rescue operations, assessing disaster conditions, and post-disaster reconstruction. Traditional flood monitoring methods mainly rely on ground observations and manual patrols, which can provide detailed water condition data in local areas. However, due to their high cost, low efficiency, and inability to quickly cover large areas, they are difficult to meet the needs of emergency response and large-scale monitoring. Therefore, remote sensing technology, especially synthetic aperture radar (SAR), has become an important technical means for flood monitoring and disaster warning as an efficient and real-time remote sensing tool. Synthetic aperture radar (SAR) has all-weather and all-day imaging capabilities and is not affected by adverse weather conditions such as clouds and rain, so it has a significant advantage in flood monitoring. SAR can provide reliable image data under different weather conditions, making it an ideal tool for emergency response and post-disaster monitoring.

[0003] Although SAR has important value in flood monitoring, current research mostly focuses on C-band data, and research on S-band SAR data is still insufficient.

[0004] In recent years, deep learning technology has made significant progress in the field of remote sensing image processing, providing new ideas for flood identification. However, existing deep learning methods are mostly based on general models and lack targeted optimization for SAR image characteristics and flood disaster features, resulting in insufficient precision and robustness in S-band SAR flood identification tasks. Visual basic models (such as Segment Anything Model, SAM) have shown strong generalization ability and zero-shot learning ability in natural image segmentation. However, direct application of SAM to SAR image flood identification does not yield good results, mainly due to significant differences in imaging mechanisms and data characteristics between natural images and SAR images.

[0005] In summary, although deep learning and visual basic models perform well in image segmentation, they still face many challenges in the task of SAR image flood identification. Therefore, it is necessary to optimize existing methods to better adapt to the characteristics of S-band SAR images and improve the precision and robustness of flood identification. SUMMARY

[0006] To solve the above technical problems, the present invention provides a flood inundation area identification method and device based on S-band SAR data and a visual basic model, aiming to improve the accuracy and robustness of flood identification in S-band SAR images, especially the ability to identify scattered small water bodies.

[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] A flood inundation area identification method based on S-band SAR data and a visual base model, the method comprising:

[0009] Step S1, acquiring S-band SAR data and auxiliary data;

[0010] Step S2: preprocessing the S-band SAR data and auxiliary data;

[0011] Step S3: constructing a water sample set based on the preprocessed S-band SAR data and auxiliary data;

[0012] Step S4: constructing an SSUNet model and training the SSUNet model based on the water sample set. The SSUNet model includes a visual base model and a windmill convolution module. The visual base model is used to extract image features, and the windmill convolution module is used to capture small water bodies and process blurred image boundary areas.

[0013] Step S5: Input the pre-processed S-band SAR data to be identified into the trained SSUNet model to obtain the identification result of the flood area.

[0014] In another aspect, the present invention provides a flood inundation area identification device based on S-band SAR data and a visual base model, comprising:

[0015] A data acquisition unit, used to acquire S-band SAR data and auxiliary data;

[0016] A data preprocessing unit, used for preprocessing S-band SAR data and auxiliary data;

[0017] A sample set construction unit is used to construct a water sample set based on the preprocessed S-band SAR data and auxiliary data;

[0018] A model training unit, configured to construct an SSUNet model and train the SSUNet model based on a water sample set. The SSUNet model includes a visual base model and a windmill convolution module. The visual base model is configured to extract image features, and the windmill convolution module is configured to capture small water bodies and process blurred image boundary areas.

[0019] The flood distribution map acquisition unit is configured to input the preprocessed S-band SAR data to be identified into the trained SSUNet model to obtain an identification result of a flood area.

[0020] In a third aspect, the present application provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned flood inundated area identification method based on S-band SAR data and a visual basic model.

[0021] In a fourth aspect, the present application provides a computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, enable the processor to implement the aforementioned flood inundated area identification method based on S-band SAR data and a visual basic model.

[0022] The present application has the following beneficial effects:

[0023] 1. The SSUNet model of the present application combines a visual basic model and a pinwheel convolution module, and has high water body classification accuracy. Compared with the commonly used U-Net model, the SSUNet model improves the IoU (intersection over union) index by 2.35%, can fully exert the powerful feature extraction capability of the visual basic model, and enhance the capture ability of small water bodies through the pinwheel convolution module, and is especially suitable for water body identification tasks in S-band SAR data.

[0024] 2. The SSUNet model effectively utilizes the powerful generalization ability of the visual basic model and the detail capture ability of the pinwheel convolution module, so that the model can effectively distinguish small water bodies, significantly improve the adaptability to S-band SAR data, and thus show better accuracy and robustness in the flood identification task of HJ-2E SAR data. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 A flowchart of the flood inundated area identification method based on S-band SAR data and a visual basic model of the present application is shown in Figure 1.

[0026] Figure 2 A structure diagram of the SSUNet model proposed by the present application is shown in Figure 2.

[0027] Figure 3 A structure diagram of the pinwheel convolution (PSM) module is shown in Figure 3.

[0028] Figure 4 A comparison result diagram of the SSUNet model and the FCN and UNet models is shown in Figure 4.

[0029] Figure 5 A flood distribution result identified by using the trained SSUNet model is shown in Figure 5. DETAILED DESCRIPTION

[0030] The application will be further described below in conjunction with the accompanying drawings and examples.

[0031] As Figure 1 shown, it is a flow chart of a flood inundation area identification method based on S-band SAR data and visual basic model of the application, specifically comprising:

[0032] Step S1. Obtain S-band SAR data and auxiliary data;

[0033] The data used by the application includes HJ-2E SAR data and auxiliary data. Specifically:

[0034] HJ-2E adopts Level-1 SLC product of NSCAN scanning mode, contains HH (horizontal-horizontal polarization) and HV (horizontal-vertical polarization) features, and the spatial resolution is 20m. That is, the data collected this time is HJ-2E SAR data before and after the disaster, containing two polarization modes: HH polarization (horizontal transmission-horizontal reception) and HV polarization (horizontal transmission-vertical reception), the imaging date before the disaster is 2024.08.18, and the imaging date after the disaster is 2024.09.22.

[0035] The auxiliary data is GF-2 remote sensing image, land cover / land use data and digital elevation model (DEM) data, etc. GF-2 satellite provides visible and near-infrared band data, and the spatial resolution is 3.2 meters. The GF-2 data collected this time is the image after the disaster, the imaging date is 2024.09.21, and it is mainly used to assist in flood interpretation.

[0036] Step S2. Preprocess the S-band SAR data and auxiliary data;

[0037] Step S2-1: For HJ-2E SAR data, use Python script for radiation correction, spot filtering, geographic coding and resampling, extract HH and HV polarization features, ensure data quality and meet the needs of subsequent analysis; preferably, in step S2-1, the number of channels of the finally preprocessed HJ-2E SAR data is 2.

[0038] Step S2-2: For GF-2 data, use ENVI platform to perform radiation calibration, atmospheric correction and orthorectification, and extract multispectral features, including blue, green, red and near-infrared band information. After all the data are processed, they are uniformly converted to EPSG:4326 (WGS84) geographic coordinate system to ensure the spatial consistency between different data sets and facilitate subsequent comprehensive analysis. Preferably, in step S2-2, the number of channels of the finally preprocessed GF-2 optical data is 4.

[0039] Step S3. Constructing water body sample set based on pre-processed S-band SAR data and auxiliary data;

[0040] Step S3-1: Combining pre-processed HJ-2E SAR data before and after disaster and auxiliary optical remote sensing data, water body and non-water body labels are made by visual interpretation method;

[0041] Step S3-2: Combining water body / non-water body labels and dual-polarization S-band SAR features, water body segmentation sample set is constructed. The size of sample slice is 352x352, and there are a total of 710 slices;

[0042] Step S3-3: The constructed sample set is divided into training set and test set in the ratio of 5:5.

[0043] Step S4. Constructing SSUNet model and training the model based on the training set in the water body segmentation sample set;

[0044] Step S4-1: Constructing SSUNet model, which is a U-shaped encoder and decoder structure, composed of encoding block, decoding block and pinwheel convolution module (PSM). The pinwheel convolution module and the encoding block one by one correspond to constitute a multi-layer network, which is used for fine feature acquisition of image features extracted by the encoding. Wherein, the fine features acquired by the first layer network are sent into the second layer encoding block, the second layer encoding block decodes the fine features extracted by the pinwheel convolution module of the first layer network and sends them into the third layer network, and the process is repeated until the acquired fine feature image is the same size as the input image.

[0045] Preferably, four encoding blocks, three decoding blocks and four pinwheel convolution modules (PSMs) are selected in the embodiment. The encoding blocks use a Hiera block pre-trained based on SAM2 (Segment Anything Model 2) as the backbone, and the parameters of the Hiera are frozen. The Hiera backbone is composed of four Hiera blocks (corresponding to the four encoding blocks), and a SAR adapter is inserted before each Hiera block. Each adapter includes an up-sampling linear layer, a GeLU activation function, a down-sampling linear layer and a GeLU activation function connected in sequence, so that the knowledge learned by SAM2 in the natural image field can be quickly applied to the SAR flood segmentation task. The input of the adapter is the input S-band SAR feature or the output of the Hiera block in the previous encoding block, and the output is given to the Hiera block in the next encoding block. The decoding block is similar to the U-Net model, which uses double convolution modules. Each convolution module includes a 3x3 convolution layer, a batch normalization layer and a ReLU activation function connected in sequence. In the network architecture, four pinwheel convolution modules (PSMs) are inserted, one PSM for each encoding block, a total of four layers, to enhance the model's ability to capture small water bodies. As shown in Figure 2 , the first layer of the encoding block sends the input picture feature to the second layer of the decoding block after extracting the input picture feature and using the PSM to extract fine features. That is, the second layer of the decoding block not only receives the fine features of the PSM of the current layer, but also fuses the fine features extracted by the PSM of the previous layer, and decodes layer by layer to restore the final fine features to the same size as the input image.

[0046] Compared with the traditional convolution module, the pinwheel convolution module creates horizontal and vertical convolution kernels for different regions of the image by asymmetric padding, thereby achieving more fine feature extraction. Its structure is shown in Figure 3 . Specifically, for the input feature, the pinwheel convolution module uses an asymmetric padding strategy in the left and right and top and bottom directions. For example, when the padding parameters are (1, 0, 0, 3), it means filling 1 pixel on the left side of the input feature and 3 pixels on the bottom, while not filling on the right and top. Then, the module uses a 1x3 convolution kernel and a 3x1 convolution kernel to extract features in the horizontal and vertical directions, respectively. The feature maps obtained by different direction convolution are spliced together to integrate multi-directional context information, thereby obtaining richer semantic features. Finally, a 2x2 convolution layer is used to fuse the spliced features and output the final result. After the above operation, the model outputs the water and non-water binary classification results consistent with the size of the input feature map through the up-sampling module.

[0047] The loss function includes the sum of the weighted IoU loss and the binary cross-entropy (BCE) loss, i.e. To enhance the training effect of the model, a deep supervision strategy is introduced, which combines water labels to supervise the output features of each decoding block of the network (as shown by the small cubes in Figure 2 ). This further improves the performance and robustness of the model, resulting in losses , and . The total loss is . As shown in Figure 2 , the loss function corresponds to each decoding block.

[0048] Step S4-2: Train the SSUNet model using the AdamW optimizer. Use the divided training set as input, set the initial learning rate of the training process to 1e-3, the weight decay coefficient to 5e-4, and the total number of training epochs to 100.

[0049] Step S4-3: To verify the effectiveness of the SSUNet model, compare it with the FCN and UNet models on the test set and use multiple indicators (such as IoU, F1 score, Recall, and Precision) for accuracy evaluation. The specific evaluation indicator formulas are as follows:

[0050] ,

[0051] ,

[0052] ,

[0053] ,

[0054] where TP (True Positive) represents the number of pixels correctly classified as water, FP (False Positive) represents the number of non-water pixels misclassified as water, TN (True Negative) represents the number of pixels correctly classified as non-water, FN (False Negative) represents the number of water pixels misclassified as non-water, IoU represents the ratio of the intersection area of the model's predicted water area and the true water area to the union area of the two, Recall (Recall Rate) represents the proportion of actual water pixels correctly identified, Precision (Precision) represents the proportion of pixels identified as water that are actually water, and F1 score is the harmonic mean of Precision and Recall, used to comprehensively evaluate the model's performance.

[0055] The accuracy evaluation results of different models are shown in Table 1. The results show that the SSUNet model performs better than the FCN and U-Net models in the water body identification task, especially in the IoU, F1 score, Precision and Recall indicators, which are improved by 2.35% compared with the UNet. The inference results of different models are shown in Figure 4 The results show that the SSUNet model can not only effectively identify large-area water bodies, but also more accurately capture small water bodies, especially in post-disaster images, and the detection performance for small-scale water bodies is particularly outstanding.

[0056] Table 1

[0057]

[0058] Step S5. The preprocessed S-band SAR data to be identified is input into the trained SSUNet model to obtain the identification result of the flood area.

[0059] Step S5-1: According to the SSUNet model trained in step S4, the preprocessed HJ-2E SAR data before and after the disaster is inferred, and the SSUNet model automatically identifies the water body distribution before and after the disaster to generate a water body distribution map before and after the disaster.

[0060] Step S5-2: Comparing the water body distribution maps before and after the disaster, the change of water body distribution is analyzed to identify the flood area during the disaster. The flood area of Yongcheng City is obtained, and the results are shown in Figure 5 .

[0061] In step S5-2, the water body classification result after the disaster and the water body identification result before the disaster are subtracted, and the area greater than 0 becomes the flood area.

[0062] On the other hand, the present application provides a flood inundation area identification device based on S-band SAR data and visual basic model, which includes various units that can realize each step of the aforementioned method. Specifically, it includes a data acquisition unit for acquiring S-band SAR data and auxiliary data.

[0063] A data preprocessing unit for preprocessing S-band SAR data and auxiliary data;

[0064] A sample set construction unit for constructing a water body sample set based on the preprocessed S-band SAR data and auxiliary data;

[0065] The model training unit is configured to construct an SSUNet model and train the SSUNet model based on a water body sample set, the SSUNet model comprising a visual base model and a pinwheel convolution module, the visual base model being configured to extract image features, and the pinwheel convolution module being configured to capture small water bodies and process image boundary blur areas.

[0066] The flood distribution map acquisition unit is configured to input the preprocessed S-band SAR data to be identified into the trained SSUNet model to obtain a recognition result of a flood area.

[0067] In a third aspect, the present application provides an electronic device, comprising: one or more processors; a memory configured to store one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned flood inundation area identification method based on S-band SAR data and a visual base model.

[0068] In a fourth aspect, the present application provides a computer-readable storage medium having stored executable instructions, which, when executed by a processor, enable the processor to implement the above-mentioned flood inundation area identification method based on S-band SAR data and a visual base model.

[0069] The above-mentioned specific embodiments further illustrate the purpose, technical solutions and advantages of the present application. It should be understood that the above-mentioned specific embodiments are only examples of the present application and are not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A flood inundation area identification method based on S-band SAR data and visual basic model, characterized by: The method comprises: Step S1, acquiring S-band SAR data and auxiliary data; the S-band SAR data is HJ-2E SAR data acquired before and after the flood disaster; the auxiliary data includes GF-2 optical remote sensing data, land cover / land use data and DEM data; Step S2: preprocessing the S-band SAR data and auxiliary data; Step S3: constructing a water sample set based on the preprocessed S-band SAR data and auxiliary data; Step S4: construct an SSUNet model and train the SSUNet model based on the water sample set. The SSUNet model includes a visual base model and a windmill convolution module. The visual base model is used to extract image features, and the windmill convolution module is used to capture small water bodies and process image boundary fuzzy areas. The SSUNet model includes: Step S4-1, constructing an SSUNet model, the SSUNet model includes an encoding block, a decoding block and a windmill convolution module; wherein, The encoding block is constructed based on the pre-trained Hiera block, the Hiera parameters are frozen, and a SAR adapter is inserted before the Hiera block. Each adapter includes an upsampling linear layer, a GeLU activation function, a downsampling linear layer, and a GeLU activation function connected in sequence; The decoding block adopts a dual convolution module, each of which includes a 3×3 convolution layer, a batch normalization layer and a ReLU activation function connected in sequence; The windmill convolution module and the encoding block correspond one-to-one to form a multi-layer network for obtaining fine features of the image features extracted by the encoding, wherein the fine features obtained by the first layer of the network are sent to the second layer of the encoding block, and the second layer of the encoding block decodes the fine features extracted by the windmill convolution module of this layer and the fine features extracted by the first layer of the network and then sends them to the third layer of the network, and so on, until the obtained fine feature image has the same size as the input image; wherein, the output features of each decoding block of the network are optimized through multi-level loss supervision in combination with the water body label; In step S4-2, the SSUNet model is trained using the AdamW optimizer based on the divided training set. The initial learning rate is set to 1e-3, the weight decay coefficient is 5e-4, and the loss function is a combination of weighted IoU loss and binary cross entropy loss. The pinwheel convolution module adopts an asymmetric padding strategy for the image features extracted by the encoder in the left-right and top-bottom directions, using a 1×3 convolution kernel and a 3×1 convolution kernel to extract features in the horizontal and vertical directions respectively, and performs feature splicing. The spliced ​​features are fused through a 2×2 convolution layer and the output is a water body and non-water body binary classification result with the same size as the image features extracted by the encoder. Step S5: Input the pre-processed S-band SAR data to be identified into the trained SSUNet model to obtain the identification result of the flood area.

2. The flood inundation area identification method based on S-band SAR data and visual basic model according to claim 1 is characterized in that: The step S2 comprises: Step S2-1, performing radiometric correction, speckle filtering, geocoding, and resampling operations on the HJ-2E SAR data to obtain HH and HV dual-polarization features; Step S2-2: performing radiometric calibration, atmospheric correction, and orthorectification on the GF-2 optical remote sensing data to extract multispectral features.

3. The flood inundation area identification method based on S-band SAR data and visual basic model according to claim 1 is characterized in that: The step S3 comprises: Step S3-1: Based on the pre-processed HJ-2E SAR data and auxiliary data, a visual interpretation method is used to manually mark water and non-water areas and construct water and non-water labels; Step S3-2, combining water body / non-water body labels and S-band SAR dual-polarization features to construct a water body semantic segmentation sample set; In step S3-3, the water body semantic segmentation sample set is divided into a training set and a test set in proportion.

4. The method for identifying flooded areas based on S-band SAR data and a visual base model according to claim 1, wherein: The step S5 comprises: Step S5-1: Based on the trained SSUNet model, the acquired pre-processed HJ-2E NSCAN image to be tested is inferred to identify the distribution of water bodies before and after the disaster. Step S5-2: Determine the flood area by combining the water area distribution map before and after the disaster.

5. A flood inundation area identification device based on S-band SAR data and a visual base model, applied to the method according to any one of claims 1 to 4, characterized in that: include: A data acquisition unit, used to acquire S-band SAR data and auxiliary data; A data preprocessing unit, used for preprocessing S-band SAR data and auxiliary data; A sample set construction unit is used to construct a water sample set based on the preprocessed S-band SAR data and auxiliary data; A model training unit, configured to construct an SSUNet model and train the SSUNet model based on a water sample set. The SSUNet model includes a visual base model and a windmill convolution module. The visual base model is configured to extract image features, and the windmill convolution module is configured to capture small water bodies and process blurred image boundary areas. The flood distribution map acquisition unit is used to input the pre-processed S-band SAR data to be identified into the trained SSUNet model to obtain the identification results of the flood area.

6. An electronic device, characterized in that: include: one or more processors; a memory for storing one or more programs; Wherein, when one or more programs are executed by the one or more processors, the one or more processors implement the flood inundation area identification method based on S-band SAR data and visual basic model as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that Executable instructions are stored thereon, which, when executed by a processor, enable the processor to implement the flood inundation area identification method based on S-band SAR data and a visual basic model as described in any one of claims 1 to 4.

Images