Adaptive flood inundation range extraction method fusing multi-source remote sensing data

By integrating multi-source remote sensing data and deep learning models, the high confidence water area is extracted, and a single data source and shadow interference problem extracted from flood flood range in the existing technology is solved, the accuracy and efficiency of monitoring are improved, and the flood prevention and disaster relief is provided. More support.

CN118968237AActive Publication Date: 2024-11-15FUDAN UNIVERSITY

Patent Information

Application Number
CN202410905083.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-08
Publication Date
2024-11-15
Estimated Expiration
2044-07-08

AI Technical Summary

Technical Problem

The existing flood range extraction methods have limitations of a single data source, difficulty in determining thresholds and complex shadow interference problems, the deep learning model has poor generalization capabilities and is difficult to adapt to different environments, resulting in low monitoring accuracy and efficiency.

Method used

Adaptive methods that integrate multi-source remote sensing data include obtaining optical and SAR remote sensing image data, performing pre-processing and high-precision registration, extracting high-confidence water body areas, building high-quality water body data sets, and training U-Net models for water body Identify and automatically extract the flooding range.

Benefits of technology

It improves the accuracy and robustness of flooding range extraction, overcomes the limitations of a single data source and traditional threshold segmentation method, and provides stronger technical support for flood prevention and disaster relief.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968237A_ABST
    Figure CN118968237A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of flood disaster monitoring, and particularly relates to a self-adaptive flood inundation range extraction method fusing multi-source remote sensing data. The method comprises the following steps: acquiring optical and SAR remote sensing image data; preprocessing the multi-source remote sensing image; optical and SAR remote sensing images are fused; extracting a high-confidence discrete water body region; extracting a continuous water body region, fusing the continuous water body region with the digital elevation model, and constructing a high-quality water body data set; training a water body recognition deep learning model based on the water body data set; and extracting a flood inundation range by using a water body recognition model. The method solves the problems that a threshold value is difficult to determine, complex shadow interferences exist and training data of deep learning is less when a flood inundation range is extracted by using a remote sensing image. In addition, the method is high in robustness and high in recognition precision, the flood inundation range can be rapidly calculated, and technical support is provided for flood prevention and disaster relief.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of flood disaster monitoring, and in particular relates to a method for extracting a flood submerged range. Background Art

[0002] As a natural disaster, flood disasters have a great impact on human society and the natural environment. In order to effectively deal with flood disasters, it is crucial to accurately and quickly extract the flood inundation range. At present, remote sensing technology is widely used in flood monitoring and assessment because it can provide large-scale and highly timely surface information. Despite this, the existing flood inundation range extraction methods still have limitations due to the use of only a single type of image data, such as optical remote sensing images or SAR remote sensing images. Optical remote sensing images can provide high resolution and rich surface feature information under good weather conditions and without cloud cover, but they are almost ineffective under severe weather conditions such as clouds and rain, and floods are mostly caused by heavy rainfall. In contrast, SAR remote sensing images can penetrate clouds for all-weather observations, but they are limited by the wavelength and incident angle of radar waves, and their ability to describe fine surface features is not as good as optical images. Flood inundation range extraction methods that rely on a single data source are prone to misjudgment under complex surface conditions, such as urban areas and vegetation-covered areas. For example, the angular reflection phenomenon in SAR remote sensing images may cause the reflection characteristics of water bodies to be similar to those of wet surfaces, increasing the difficulty of classification. By fusing optical remote sensing images and SAR remote sensing images, we can combine the advantages of both and overcome the limitations of a single data source.

[0003] The threshold method can quickly extract water areas from SAR remote sensing images, but pixels at the boundaries of water bodies are often difficult to define and are also affected by complex factors such as mountain shadows. Remote sensing image water extraction based on deep learning can effectively solve the classification problem of water body boundary pixels. However, due to the lack of training data, deep learning-based methods are prone to learning the noise of the data in the training set rather than the real rules behind it, and the generalization ability of the model is poor.

[0004] Summarizing the above practical needs and problems, it can be seen that the current field still lacks a complete adaptive flood inundation range extraction method. The existing extraction methods have problems such as difficult threshold determination and complex shadow interference. In addition, when using deep learning methods to extract flood inundation ranges, insufficient training data leads to poor model generalization ability, and water bodies can only be identified in specific areas. Therefore, how to develop an adaptive flood inundation range extraction method based on deep learning that is suitable for different environments, improve the accuracy and efficiency of flood monitoring, speed up the acquisition of disaster information, and provide strong support for flood prevention and disaster relief is a technical problem that technicians in this field urgently need to solve. Summary of the invention

[0005] The purpose of the present invention is to provide an adaptive flood inundation range extraction method integrating multi-source remote sensing data to improve the accuracy and efficiency of flood monitoring, speed up the acquisition of disaster information, and provide technical support for flood prevention and disaster relief.

[0006] The adaptive flood inundation range extraction method provided by the present invention integrates multi-source remote sensing data, including six steps: acquiring multi-source remote sensing image data, preprocessing the multi-source remote sensing image data, extracting high-confidence discrete water body areas, constructing a high-quality water body data set, training a water body recognition deep learning model, and automatically extracting the flood inundation range. The multi-source remote sensing images include optical remote sensing images and SAR remote sensing images, aiming to overcome the limitation of a single data source; the preprocessing of multi-source remote sensing image data includes radiation correction, atmospheric correction and image registration, so as to achieve the optimization processing and high-precision registration of remote sensing image data; high-confidence discrete water body areas are used to generate prompt points of the Segment Anything Model [1] (SAM) to obtain continuous water body areas; the obtained water body areas are fused with digital elevation model (DEM) data to generate high-quality water body data sets for training the water body recognition model (referred to as U-net model) constructed by the present invention; the trained water body recognition model is used to complete the automatic extraction of water bodies in remote sensing images and obtain their flood inundation range.

[0007] The specific contents of each step are as follows:

[0008] Step 1: Obtain multi-source remote sensing image data

[0009] Multi-source remote sensing images include optical remote sensing images and SAR remote sensing images; specifically, optical remote sensing image data before the flood disaster event occurs and synthetic aperture radar remote sensing image data (SyntheticAperture Radar) before and during the flood disaster event are obtained as the image data source for automatic extraction of the flood inundation range. The SAR remote sensing image data is Sentinel-1 SLC format data with a polarization mode of VH, and the optical remote sensing image data is Landsat 8OLI / TIRS C1Level-1 optical image. Digital elevation model data is used as auxiliary data for screening water body areas of SAR remote sensing images. The present invention uses the global digital elevation model data set of ASTER GDEM 30m resolution.

[0010] Step 2: Preprocessing of multi-source remote sensing image data;

[0011] That is, the acquired optical remote sensing images and SAR remote sensing images are preprocessed. Among them, the SAR remote sensing images in SLC format are multi-viewed to obtain intensity maps. The output of multi-view processing is an optimized image with lower speckle noise, improved spatial and radiometric resolution, and improved image availability and interpretability. Based on the obtained intensity map, Lee filtering (7*7) is performed, and the output of the filter is controlled by the local statistical characteristics of the image to remove the inherent coherent speckle noise of the SAR image. The image noise after filtering is reduced, the image is clearer, and the water body is more obvious. The filtering results are calibrated by radiation, and the intensity value of the remote sensing image is converted into a backscattering coefficient. The acquired Landsat 8OLI / TIRS C1 Level-1 optical image has been geometrically corrected and terrain corrected, and is directly subjected to radiation calibration and atmospheric correction. Among them, radiation calibration is used to eliminate the error generated by the sensor itself, and atmospheric correction is used to eliminate the influence of factors such as atmosphere and light on the reflection of ground objects. The radiation calibration type is the radiation brightness value (Radiance), the coefficient is 0.1, and the atmospheric correction adopts the FLASSH model.

[0012] The optical remote sensing images and SAR remote sensing images that have been corrected for radiation and atmosphere are registered. Registration of optical remote sensing images and SAR remote sensing images based on NDF and improved SIFT: When constructing the scale space of optical and SAR remote sensing images, NDF is used instead of Gaussian filtering to construct the Harris scale space. This method better retains the edge information of the image while denoising. In addition, in terms of gradient calculation, multi-scale Sobel and multi-scale ROEWA operators are used to calculate the gradients of optical and SAR images, respectively, so as to obtain gradient information with good consistency. The descriptor is constructed using the GLOH template to increase the robustness of the descriptor. Finally, the feature matching is constrained by using the scale, position and direction of the matching points to effectively eliminate false matches. High-precision registration of optical images and SAR remote sensing images is achieved.

[0013] Step 3: Extract high confidence discrete water area;

[0014] Calculate the water index of optical remote sensing images and SAR remote sensing images.

[0015] Optical remote sensing images use the normalized water index, the formula is:

[0016]

[0017] Where Green and NIR are the brightness values ​​of the green band and near infrared band respectively.

[0018] The range of NDWI values ​​is [-1, 1]. In order to facilitate threshold selection and improve the accuracy of threshold selection, the range of NDWI index values ​​is stretched from [-1, 1] to [0, 2000]. Using the histogram bimodal method, the grayscale corresponding to the valley between the two peaks is selected as the image segmentation threshold. The grayscale above the threshold is the high-confidence water area of ​​the image.

[0019] For SAR remote sensing images, based on the relationship between VV and VH polarization images and water body information, the SDWI water body index is used to extract the range of water bodies in SAR images. The formula is:

[0020] SDWI=ln(10*VV*VH), (2)

[0021] In the formula, SDWI is the water extraction index, VV and VH represent the values ​​obtained by two polarization methods respectively. When the SDWI value is greater than 0, it is a high-confidence water body.

[0022] The overlapping areas of high-confidence water bodies obtained from optical remote sensing images and SAR remote sensing images are taken as high-confidence water body candidate areas.

[0023] Step 4: Build a high-quality water body dataset;

[0024] Based on the high-confidence water body candidate area, randomly select pixel points as the point prompt of the Segment Anything Model to extract the water body in the continuous area. Use the digital elevation model DEM to calculate the terrain slope data to eliminate the interference of mountain shadows in the continuous water body area, and obtain high-quality water body range data after processing. Based on the SAR remote sensing image data and the high-quality water body range, manually calibrate the labels to construct a high-quality water body dataset.

[0025] Step 5: Train the deep learning model for water body identification

[0026] The deep learning model for water body recognition adopts the U-Net model, whose structure includes three parts: encoder, decoder and jump connection; among them:

[0027] The encoder adopts ResNet101 (Residual Network-101) network (a deep convolutional neural network architecture used for computer vision tasks such as image classification and target detection), which extracts features through residual blocks, alleviates the gradient vanishing problem, and can effectively capture the deep features of the image; the encoder consists of multiple convolutional layers, batch normalization layers and ReLU activation functions, and each residual block contains several convolutional layers; each stage of the encoder reduces the spatial size of the feature map and increases the number of channels, thereby extracting features at different levels. The encoder first gradually reduces the spatial dimension of the input image through a series of convolution operations, which ensures the extraction of high-level features; the introduction of batch normalization layers helps to accelerate the training process and improve the stability of the model, while the ReLU activation function enhances the nonlinear expression ability of the model. In order to better restore detail information during the decoding process, the U-Net model introduces jump connections between the encoder and the decoder.

[0028] The decoder can gradually restore the spatial resolution of the feature map and expand the size of the feature map through upsampling operations (such as transposed convolution). Each decoder stage contains a convolution layer, a batch normalization layer, and a ReLU activation function, which is fused with the feature map transmitted by the jump connection. The decoder restores the spatial size of the feature map layer by layer, and finally outputs a water body range mask with the same size as the input image. Specifically, the decoder part restores the spatial resolution of the image through layer-by-layer upsampling operations, and further enhances the accuracy of the segmentation results by fusing with the feature map transmitted by the jump connection. The upsampling process of each layer is accompanied by the application of convolution operations and activation functions, thereby ensuring that the output image has high-quality spatial and semantic information.

[0029] The skip connection directly transfers the feature maps of each layer of the encoder to the corresponding layer of the decoder, ensuring that high-resolution features can be fully utilized during the decoding process. This design enables the model to combine low-level and high-level features during the decoding process, thereby improving segmentation accuracy. Specifically, the skip connection operates on each layer of the U-Net model. By combining the high-resolution features retained in the encoder stage with the upsampling results of the decoder stage, the model can better restore the boundaries and detail information of the target area, thereby improving the accuracy of the segmentation results.

[0030] The specific process of model training is as follows: First, a high-quality water body dataset is input into the U-Net model as training data. Secondly, the encoder uses ResNet101 to extract deep features of the image, and obtains feature maps of different scales through multi-layer convolution operations. The jump connection passes the feature maps of each layer of the encoder to the decoder part to ensure that the high-resolution features are fully utilized during the decoding process. The decoder restores the spatial resolution of the feature map layer by layer, expands the size of the feature map through upsampling operations, and fuses it with the feature map of the jump connection. Finally, the model outputs a water body range mask of the same size as the input image, extracts water body edge pixels through post-processing steps, and provides a high-precision water body range.

[0031] Through the above process, the U-Net model can accurately extract the scope of water bodies in complex remote sensing images and solve the problem that the edge pixels of water bodies are difficult to define. Combined with high-quality water body datasets and ResNet101 feature extractors, the model performs well in extracting the scope of water bodies.

[0032] Step 6: Automatically extract flood inundation area

[0033] The water body identification deep learning model is used to extract the water body range in the SAR remote sensing images before and during the flood period. The extracted water body range is converted into a vector surface format to calculate the difference between the water body range in the middle and early stages of the flood period, and the result is the flood inundation range of this flood.

[0034] The present invention overcomes the limitations of a single data source and traditional threshold segmentation methods, improves the accuracy and robustness of flood inundation range extraction, and provides more powerful technical support for flood prevention and disaster relief. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Flow chart of the method of the present invention.

[0036] Figure 2 This is the Gaofen-1 remote sensing image of Jiashan County.

[0037] Figure 3 This is the Sentinel-1 remote sensing image of Jiashan County.

[0038] Figure 4 This is the DEM data of Jiashan County.

[0039] Figure 5 This is a preprocessed Sentinel-1 image (partial).

[0040] Figure 6 This is the preprocessed GF-1 image (partial).

[0041] Figure 7The water body extraction result (partial). Among them, a is the water body extraction result of GF-1 optical image, b is the water body extraction result of Sentinel-1 image, and c is the superposition result of the two.

[0042] Figure 8 This is the result of high-quality water extraction (partial).

[0043] Fig. 9 It is the flood inundation area (local). DETAILED DESCRIPTION

[0044] The present invention is further described by taking Jiashan County, Zhejiang Province as an embodiment in conjunction with the accompanying drawings.

[0045] The adaptive flood inundation range extraction method provided by fusing multi-source remote sensing data is shown in the following process. Figure 1 shown.

[0046] Step 1: Obtain multi-source remote sensing image data

[0047] Jiashan County, Zhejiang Province is taken as the study area. In order to realize the automatic extraction of flood inundation range, multi-source remote sensing image data is used as the data source. The optical remote sensing image uses the image data acquired by Gaofen-1 on August 12, 2019; the synthetic aperture radar (Synthetic Aperture Radar) remote sensing image data uses the Sentinel-1 SLC format image acquired on August 4, 2019, with the polarization mode of VH, see the attached Figure 2 , 3 In addition, to assist in the screening of water areas in SAR remote sensing images, the ASTER GDEM global digital elevation model (DEM) dataset with a resolution of 30 meters is used. Figure 4 .

[0048] Step 2: Preprocessing of multi-source remote sensing image data

[0049] The acquired optical remote sensing image and SAR remote sensing image data need to be preprocessed. First, the SAR remote sensing image in SLC format is multi-viewed to generate an intensity map. Multi-view processing can optimize the image, reduce speckle noise, and improve spatial and radiometric resolution, thereby improving the usability and interpretability of the image. Based on the generated intensity map, a 7×7 window Lee filter is applied to control the filter output by utilizing the local statistical characteristics of the image to remove the coherent speckle noise inherent in the SAR image. The image noise after filtering is significantly reduced, the image is clearer, and the water body is more obvious. Next, the filtering result is radiometrically calibrated to convert the intensity value of the remote sensing image into a backscattering coefficient. Figure 5 , Figure 6 This is the image after preprocessing.

[0050] The acquired GF-1 optical images are calibrated by radiation and corrected by atmosphere. Radiation calibration is used to eliminate the error introduced by the sensor, and atmospheric correction is used to eliminate the influence of factors such as atmosphere and illumination on the reflection of objects. The radiation calibration type is radiance, the coefficient is 0.1, and the atmospheric correction uses the FLAASH model.

[0051] The optical remote sensing image and SAR remote sensing image after radiation correction and atmospheric correction are registered. The registration process is based on NDF and improved SIFT method. When constructing the scale space of optical and SAR images, NDF is used instead of Gaussian filtering to better retain the edge information of the image while denoising. In terms of gradient calculation, multi-scale Sobel operator and multi-scale ROEWA operator are used to calculate the gradients of optical and SAR images respectively, so as to obtain gradient information with good consistency. The descriptor is constructed using GLOH template to enhance the robustness of the descriptor. Finally, by constraining the feature matching by using the scale, position and direction of the matching points, the false matching is effectively eliminated, and the high-precision registration of optical and SAR images is achieved.

[0052] Step 3: Extract high confidence discrete water areas

[0053] Calculate water index in optical remote sensing images and SAR remote sensing images.

[0054] For optical remote sensing images, the normalized difference water index (NDWI) is used, and its calculation formula is as follows:

[0055]

[0056] Where Green and NIR are the brightness values ​​of the green band and the near infrared band, respectively. The calculated NDWI value range is [-1, 1]. In order to facilitate threshold selection and improve accuracy, the NDWI value range is stretched to [0, 2000]. Through the histogram bimodal method, the gray level corresponding to the valley between the two peaks is selected as the image segmentation threshold. The area above the threshold is considered to be a high-confidence water area.

[0057] For SAR remote sensing images, the relationship between VV and VH polarization images and water body information is used to calculate the water extraction index (SDWI). The calculation formula is as follows:

[0058] SDWI=ln(10*VV*VH)

[0059] In the formula, SDWI is the water extraction index, VV and VH represent the values ​​obtained by two polarization methods respectively. When the SDWI value is greater than 0, it is a high-confidence water body.

[0060] The overlapping areas of high-confidence water bodies obtained from optical remote sensing images and SAR remote sensing images are used as high-confidence water body candidate areas. Figure 7 shown.

[0061] Step 4: Build a high-quality water dataset

[0062] Based on the high-confidence water body candidate area, randomly select pixel points as the point prompt of SegmentAnything Model to extract the water body in the continuous area. Use the digital elevation model DEM to calculate the terrain slope data to eliminate the mountain shadow interference in the continuous water body area, and obtain high-quality water body range data after processing. Based on the SAR remote sensing image data and the high-quality water body range, manually calibrate the labels to construct a high-quality water body dataset. See Figure 8 .

[0063] Step 5: Train the deep learning model for water body identification

[0064] The constructed high-quality water body dataset is imported into the U-Net model for training. The ResNet101 network is used as a feature extractor, and the gradient vanishing problem is solved by introducing a residual learning mechanism, thereby effectively extracting deep features in the image. Integrating ResNet101 into the encoder part of the U-Net model can significantly enhance the model's feature extraction capability when processing complex remote sensing images. When extracting the range of water bodies, this method can effectively solve the problem of difficult definition of water body edge pixels and provide a high-precision water body range mask.

[0065] Step 6: Automatically extract flood inundation area

[0066] The trained water body recognition deep learning model is used to extract the water body range in the SAR remote sensing images before and during the flood period. The extracted water body range is converted into a vector surface format to calculate the change in the water body range in the mid-flood period relative to the early period. The result is the inundation range of this flood. Fig. 9 .

[0067] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0068] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein. References:

[0069] [1] Segment Anything Model (SAM) comes from Meta AI Lab of Facebook. The model can recognize and segment any object based on text instructions or image recognition. SAM reference paper: Kirillov A, Mintun E, Ravi N, et al. Segment anything [C] / / Proceedings of the IEEE / CVF International Conference on Computer Vision. 2023: 4015-4026.

[0070] [2] Digital Elevation Model (DEM) is a digital simulation of the ground topography (i.e. a digital expression of the terrain surface morphology) using limited terrain elevation data. It is a physical ground model that represents the ground elevation in the form of an ordered array of numerical values.

Claims

1. An adaptive flood inundation range extraction method integrating multi-source remote sensing data, characterized in that: The specific steps are as follows: Step 1: Obtain multi-source remote sensing image data; Multi-source remote sensing images include optical remote sensing images and SAR remote sensing images. Specifically, optical remote sensing image data before the flood disaster event occurs and synthetic aperture radar remote sensing image data before and during the flood disaster event are obtained as the image data source for automatic extraction of flood inundation range. The SAR remote sensing image data is Sentinel-1 SLC format data with VH polarization mode, and the optical remote sensing image data is Landsat 8OLI / TIRS C1 Level-1 optical image. Digital elevation model data is used as auxiliary data for water area screening of SAR remote sensing images. Step 2: Preprocessing of multi-source remote sensing image data; That is, the acquired optical remote sensing images and SAR remote sensing images are preprocessed; among them, the SAR remote sensing images in SLC format are multi-viewed to obtain intensity maps; the output of multi-view processing is an optimized image; based on the obtained intensity map, Lee filtering (7*7) is performed, and the output of the filter is controlled by using the local statistical characteristics of the image to remove the coherent speckle noise inherent in the SAR image; the filtering results are calibrated by radiation, and the intensity value of the remote sensing image is converted into a backscattering coefficient; the acquired Landsat8OLI / TIRS C1 Level-1 optical image has been geometrically corrected and terrain corrected, and is directly subjected to radiation calibration and atmospheric correction processing; among them, radiation calibration is used to eliminate the error generated by the sensor itself, and atmospheric correction is used to eliminate the influence of factors such as atmosphere and light on the reflection of ground objects; Register the optical remote sensing images and SAR remote sensing images that have undergone radiation correction and atmospheric correction; Step 3: Extract high confidence discrete water area; Calculate the water index of optical remote sensing images and SAR remote sensing images; where: Optical remote sensing images use the normalized water index, the formula is: In the formula, Green and NIR are the brightness values ​​of the green band and the near infrared band respectively; The range of NDWI value is [-1,1], and the range of NDWI index value is stretched from [-1,1] to [0,2000]. Using the histogram bimodal method, the gray level corresponding to the valley between the two peaks is selected as the image segmentation threshold, and the area above the threshold is the high confidence water area of ​​the image. For SAR remote sensing images, based on the relationship between VV and VH polarization images and water body information, the SDWI water index is used to extract the range of water bodies in SAR images. The formula is: SDWI=ln(10*VV*VH), (2) In the formula, SDWI is the water extraction index, VV and VH represent the values ​​obtained by two polarization methods, and when the SDWI value is greater than 0, it is a high-confidence water body; The overlapping areas of high-confidence water bodies obtained from optical remote sensing images and SAR remote sensing images are used as high-confidence water body candidate areas; Step 4: Build a high-quality water body dataset; Based on the high-confidence water body candidate area, randomly select pixel points as the points (prompt) for segmenting all models to extract the water body in the continuous area; use the digital elevation model (DEM) to calculate the terrain slope data to eliminate the interference of mountain shadows in the continuous water body area, and obtain high-quality water body range data after processing; based on the SAR remote sensing image data and the high-quality water body range, manually calibrate the labels to construct a high-quality water body data set; Step 5: Train the deep learning model for water body recognition; The deep learning model for water body recognition adopts the U-Net model, whose structure includes three parts: encoder, decoder and jump connection; among them: The encoder adopts the ResNet101 network, which extracts features through residual blocks, alleviates the gradient vanishing problem, and can effectively capture the deep features of the image; the encoder is composed of multiple convolutional layers, batch normalization layers and ReLU activation functions, and each residual block contains several convolutional layers; each stage of the encoder reduces the spatial size of the feature map and increases the number of channels, thereby extracting features at different levels; the encoder first gradually reduces the spatial dimension of the input image through a series of convolution operations, and this process ensures the extraction of high-level features; the batch normalization layer is used to accelerate the training process and improve the stability of the model, and the ReLU activation function is used to enhance the nonlinear expression ability of the model; The decoder can gradually restore the spatial resolution of the feature map and expand the size of the feature map through upsampling operations; each decoder stage includes a convolution layer, a batch normalization layer and a ReLU activation function, and is fused with the feature map transmitted by the jump connection; the decoder restores the spatial size of the feature map layer by layer, and finally outputs a water body range mask with the same size as the input image; The skip connection directly transfers the feature maps of each layer of the encoder to the corresponding layer of the decoder, ensuring that the high-resolution features can be fully utilized in the decoding process; this design enables the model to combine low-level and high-level features in the decoding process, thereby improving the segmentation accuracy; specifically, the skip connection operates on each layer of the U-Net model, and by combining the high-resolution features retained in the encoder stage with the upsampling results in the decoder stage, the model can better restore the boundary and detail information of the target area, thereby improving the accuracy of the segmentation result; The specific process of model training is as follows: a high-quality water body dataset is input into the U-Net model as training data; the encoder uses ResNet101 to extract deep features of the image, and obtains feature maps of different scales through multi-layer convolution operations; the jump connection transfers the feature maps of each layer of the encoder to the decoder part to ensure that high-resolution features are fully utilized in the decoding process; the decoder restores the spatial resolution of the feature map layer by layer, expands the size of the feature map through upsampling operations, and fuses it with the feature map of the jump connection; the model outputs a water body range mask of the same size as the input image, extracts water body edge pixels through post-processing steps, and provides a high-precision water body range; Step 6: Automatically extract flood inundation area The trained deep learning model for water body identification is used to extract the water body range in the SAR remote sensing images before and during the middle stage of the flood. The extracted water body range is converted into a vector surface format to calculate the difference in the water body range between the middle stage and the early stage of the flood. The result is the inundation range of this flood.

2. The adaptive flood inundation range extraction method according to claim 1 is characterized in that: In step (2), the radiation calibration type is the radiation brightness value, the coefficient is 0.1, and the atmospheric correction adopts the FLASSH model.

3. The adaptive flood inundation range extraction method according to claim 1, characterized in that: In step (2), the optical remote sensing image and the SAR remote sensing image that have undergone radiation correction and atmospheric correction are registered; specifically, the optical remote sensing image and the SAR remote sensing image are registered based on NDF and improved SIFT: when constructing the scale space of the optical and SAR remote sensing images, NDF is used instead of Gaussian filtering to construct the Harris scale space; in addition, multi-scale Sobel and multi-scale ROEWA operators are used to calculate the gradients of the optical and SAR images respectively to obtain gradient information with good consistency; the descriptor is constructed using the GLOH template to increase the robustness of the descriptor; finally, the feature matching is constrained by using the scale, position and direction of the matching point to effectively eliminate false matches; Realize high-precision registration of optical images and SAR remote sensing images.

Citation Information

Patent Citations

  • Rapid extraction method for outburst flood inundation range

    CN113191292A

  • Water body extraction method based on GEE cloud platform and optical and radar data fusion

    CN114021656A

  • Flood disaster remote sensing monitoring evaluation method based on machine learning

    CN114444791A

  • Flood inundated area extraction method and system

    CN116630426A

  • Irrigation pit-pond water storage monitoring method and device, electronic equipment and storage medium

    CN116935236A

Cited By

  • Method for determining water body area of reservoir by using remote sensing image based on deep learning network

    CN119339255A

  • Method for automatically extracting contour of ice lake

    CN119360047A

  • Remote sensing identification method and system for area suitable for layout of photovoltaic facilities in pond aquaculture

    CN119810397A

  • A method and system for remotely identifying a suitable layout area of a photovoltaic facility in a pond aquaculture

    CN119810397B

  • River flow measuring and calculating method based on multi-source remote sensing data and deep learning

    CN119810647A