A method, system, equipment and storage medium for flood disaster prediction

Abstract

This disclosure relates to the field of disaster prediction technology, and in particular to a method, system, device, and storage medium for flood disaster prediction. The method includes: extracting features from optical remote sensing images and radar images respectively; inputting these features into the optical encoding channel and radar encoding channel of a preset dual-channel convolutional neural network; extracting optical and radar features at each encoding level; calculating the channel weights and spatial weights of the optical and radar features, and performing weighted fusion to obtain fused features; performing cross-level feature fusion between features from any layer of the preset dual-channel convolutional neural network decoding layer and features from adjacent decoding layers to obtain a flood water body probability map; segmenting the flood water body probability map into a binary mask based on a slope threshold from a digital elevation model; calculating an interferometric coherence coefficient threshold based on the radar image, and suppressing interference information in the binary mask based on the interferometric coherence coefficient threshold to obtain a flood disaster extent map. This method accurately extracts flood areas and improves the ability to predict flood disasters.

Description

Technical Field

[0001] This disclosure relates to the field of disaster prediction technology, and in particular to a method, system, device and storage medium for predicting flood disasters. Background Technology

[0002] Floods are among the most common natural disasters globally, posing a significant threat to human lives and property. Remote sensing technology has become an important tool for effectively monitoring, warning, and assessing flood disasters. Remote sensing imagery can rapidly acquire information about large areas of the Earth's surface, providing data support for timely responses to floods. However, accurately extracting flood disaster information from remote sensing imagery is a complex and challenging task.

[0003] In remote sensing imagery, water bodies can easily be confused with other land features (such as still, muddy water, deep water, shadows, asphalt pavements, or dark roofs) due to their similar spectral characteristics. Similarly, in SAR imagery, calm water surfaces and smooth roads, squares, airport runways, and other land features may exhibit similar backscattering characteristics under radar illumination, leading to insufficient extraction accuracy in complex scenes. Specifically, during floods, water body boundaries are often blurred, especially in areas mixed with land, resulting in inaccurate flood boundary extraction from remote sensing imagery. Furthermore, small water bodies in cities, such as narrow rivers and street flooding, are easily overlooked or incompletely extracted by existing models due to their small size. Topographical and land feature interference also plays a role; in mountainous areas, terrain shadows can severely interfere with water body identification in optical imagery. In urban areas, the tall, dense buildings create shadows and occlusion effects, making the extraction of flooded areas from remote sensing imagery extremely difficult.

[0004] Therefore, existing technologies suffer from inaccurate extraction of flood areas and are prone to misjudgment. Summary of the Invention

[0005] To address the aforementioned technical problems, this disclosure provides a flood disaster prediction method, system, device, and storage medium.

[0006] This disclosure provides a method for predicting flood disasters, including: Acquire optical remote sensing images, radar images, and digital elevation model data, and unify them into the same coordinate system for spatial registration; After spatial registration, features of optical remote sensing images and radar images are extracted respectively, and then input into the optical coding channel and radar coding channel of a preset dual-channel convolutional neural network to extract optical features and radar features at each coding level. Based on the optical and radar features of each coding level, the channel weights of the optical and radar features are calculated respectively. The spatial weights of the optical and radar features are calculated using digital elevation model data as prior information for spatial weighting, and then weighted fusion is performed to obtain the fused features. The fused features are fused with features from any layer of the preset dual-channel convolutional neural network decoding layer and features from adjacent decoding layers to obtain a flood probability map. The slope threshold of the digital elevation model is calculated using digital elevation model data, and the probability map of the flood water body is segmented into a binary mask based on the slope threshold of the digital elevation model. The interferometric coherence coefficient threshold is calculated based on the radar image, and the interference information in the binary mask is suppressed based on the interferometric coherence coefficient threshold to obtain a flood disaster range map.

[0007] Furthermore, the dual-channel convolutional neural network is provided with optical coding channels and radar coding channels having the same structure, and the parameters of the optical coding channels and radar coding channels are set respectively; The optical encoding channel extracts water body index features based on optical characteristics; The radar coding channel extracts the backscattering and polarization features of the radar characteristics.

[0008] Further, the process of calculating the channel weights of optical and radar features based on the optical and radar features at each coding level, calculating the spatial weights of optical and radar features using digital elevation model data as spatial weighting prior information, and performing weighted fusion to obtain fused features includes: Based on the optical and radar features of each coding level, the average optical features of the optical coding channel and the average radar features of the radar coding channel are calculated respectively. The optical coding channel and the radar coding channel are shared, and the optical feature channel weight and the radar feature channel weight are obtained based on the average value of the average value of the optical features. Calculate the slope map based on the digital elevation model data; The optical coding channel and the radar coding channel are spliced ​​together. Based on the terrain information of the slope map, a spatial attention map is calculated through a 7 x 7 convolutional layer to obtain the spatial weights of optical features and radar features. The weighted optical features and radar features are fused together to obtain the fused features.

[0009] Furthermore, after the steps of calculating the digital elevation model slope threshold using digital elevation model data and segmenting the flood water body probability map into a binary mask based on the digital elevation model slope threshold, the method further includes: Construct a circular structural element with a preset radius; The binary mask is expanded by the circular structural element and then eroded to obtain the closing operation result. After performing an erosion operation on the closing operation result, an expansion operation is performed to form an opening operation result; The result of the opening operation is used as the optimized binary mask.

[0010] Further, the step of calculating the interferometric coherence coefficient threshold based on radar imagery and suppressing interference information in the binary mask based on the interferometric coherence coefficient threshold to obtain a flood disaster range map includes: Calculate the interference coherence coefficient between pre-disaster and post-disaster radar images based on radar imagery. A coherence coefficient threshold is set based on the interference coherence coefficient; The binary mask is evaluated pixel by pixel, and pixels with a coherence coefficient less than the coherence coefficient threshold are retained to obtain a flood disaster range map.

[0011] Furthermore, the cross-level feature fusion is used to calculate and output a probability map of flood water bodies through a 1x1 convolutional layer.

[0012] Furthermore, prior to spatial registration, it also includes: Atmospheric correction is performed on the optical image; The radar image is subjected to noise suppression and radiometric calibration.

[0013] This disclosure also provides a flood disaster prediction system, including: The acquisition module is used to acquire optical remote sensing images, radar images, and digital elevation model data, and to unify them into the same coordinate system for spatial registration. The extraction module is used to extract features from optical remote sensing images and radar images after spatial registration, and input them into the optical coding channel and radar coding channel of a preset dual-channel convolutional neural network to extract optical features and radar features at each coding level. The calculation module is used to calculate the channel weights of optical features and radar features based on the optical features and radar features of each coding level, respectively. It uses digital elevation model data as prior information for spatial weighting to calculate the spatial weights of optical features and radar features, and performs weighted fusion to obtain fused features. The fusion module is used to perform cross-level feature fusion of the fused features with features of any layer in the preset dual-channel convolutional neural network decoding layer and features of adjacent decoding layers to obtain a flood water body probability map; The segmentation module is used to calculate the slope threshold of the digital elevation model through the digital elevation model data, and to segment the flood water body probability map into a binary mask based on the slope threshold of the digital elevation model. The suppression module is used to calculate the interferometric coherence coefficient threshold based on the radar image, and suppress the interference information in the binary mask based on the interferometric coherence coefficient threshold to obtain a flood disaster range map.

[0014] This disclosure also provides a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the flood disaster prediction method.

[0015] This disclosure also provides a computer-readable storage medium storing computer-readable instructions, which, when executed by a processor, implement the steps of the flood disaster prediction method.

[0016] The technical solution provided in this disclosure has the following advantages compared with the prior art: By acquiring optical remote sensing imagery, radar imagery, and digital elevation model (DEM) data, and unifying them into the same coordinate system for spatial registration, features from both optical and radar images are extracted and input into the optical and radar coding channels of a pre-defined dual-channel convolutional neural network. Optical and radar features are extracted at each coding level. Channel weights for the optical and radar features at each coding level are calculated, and the spatial weights of the optical and radar features are calculated using DEM data as prior information for spatial weighting. The system performs weighted fusion to obtain fused features; it then performs cross-level feature fusion between any layer of the preset dual-channel convolutional neural network decoding layer and features from adjacent decoding layers to obtain a flood probability map; it calculates the digital elevation model (DEM) slope threshold using DEM data and segments the flood probability map into a binary mask based on the DEM slope threshold; it calculates the interferometric coherence coefficient (ICC) threshold based on radar imagery and suppresses interference information in the binary mask based on the IIC C, thereby obtaining a flood disaster range map, accurately extracting flood areas, and improving the ability to predict flood disasters. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0018] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1A schematic diagram of the flood disaster prediction method provided in the embodiments of this disclosure; Figure 2 A schematic diagram illustrating the acquisition of fusion features provided in an embodiment of this disclosure; Figure 3 This is a schematic diagram of a flood disaster prediction system provided in an embodiment of the present disclosure. Detailed Implementation

[0020] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0021] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.

[0022] Figure 1 A schematic diagram of the flood disaster prediction method provided in this embodiment of the disclosure; as shown Figure 1 As shown, a flood disaster prediction method includes: Step S1: Acquire optical remote sensing images, radar images, and digital elevation model data, and unify them into the same coordinate system for spatial registration; In this embodiment, optical remote sensing images, radar images, and digital elevation model (DEM) data are acquired and spatially registered under the same coordinate system to solve the problem of spatial inconsistency in multi-source remote sensing data. Since optical images typically originate from optical satellites such as Gaofen-2, their imaging mechanism is based on solar reflection, while radar images originate from synthetic aperture radar systems, their imaging geometry and side-view characteristics differ from optical images. Furthermore, DEM data often have different resolutions and projection methods. Therefore, in the initial state, the three types of data suffer from different projection references, different pixel sizes, and spatial misalignment. By unifying them to the same geographic coordinate system and performing resampling and geometric correction, pixel-level one-to-one correspondence can be achieved, providing a fundamental guarantee for subsequent multimodal feature fusion. Without strict registration, subsequent feature fusion will produce spatial offset errors, leading to misjudgments or blurred boundaries. This eliminates geometric differences between different sensors, improves the spatial consistency of multi-source data, and fundamentally ensures spatial comparability and correspondence when the subsequent dual-channel network extracts features, thereby improving the overall accuracy and stability of flood identification.

[0023] Step S2: After spatial registration, extract the features of optical remote sensing images and radar images respectively, and input them into the optical coding channel and radar coding channel of the preset dual-channel convolutional neural network respectively to extract the optical features and radar features of each coding level. In this embodiment, a dual-channel structure is employed to achieve modal separation modeling. This means that optical and radar images have independent parameters and identical structures during the feature extraction stage, avoiding interference between different physical imaging mechanisms. The optical encoding channel primarily learns spectral features, such as the difference in water reflection between green and near-infrared bands, as well as texture and edge information. The radar encoding channel focuses on learning features such as backscattering intensity, polarization differences, and microwave texture structure. Through multi-layer convolution and downsampling operations, semantic feature representations can be gradually extracted from shallow local texture information to deep semantic features, enabling the network to possess both detailed characterization and overall semantic understanding capabilities. This fully leverages the complementary advantages of optical and radar data, allowing the network to still rely on radar features for recognition even under cloud cover or changing lighting conditions. Simultaneously, it avoids feature confusion caused by simple splicing, improving the purity and discriminative power of multimodal feature representation.

[0024] Step S3: Calculate the channel weights of the optical and radar features based on the optical and radar features of each coding level, calculate the spatial weights of the optical and radar features using digital elevation model data as prior information for spatial weighting, and perform weighted fusion to obtain the fused features. In this embodiment, channel weights are used to measure the importance of each feature channel for flood identification. A channel-level descriptive vector is generated by performing global statistical calculations on each channel within the spatial range. This vector is then processed by a weight generation mechanism to obtain adaptive channel weights, thereby strengthening key features and suppressing irrelevant features. Spatial weights utilize DEM slope or terrain information as priors, emphasizing the responsiveness of low-lying areas and reducing the probability of misclassification in high-slope areas. Finally, through the dual weighting of channel and spatial weights, adaptive fusion of optical and radar features is achieved. Combining data-driven deep feature learning with physical terrain constraints preserves the expressive power of deep networks while introducing terrain priors to enhance discrimination stability, thus significantly reducing the misclassification rate in urban and mountainous areas and improving the model's physical rationality and generalization ability.

[0025] Step S4: Perform cross-level feature fusion between the features of any layer in the preset dual-channel convolutional neural network decoding layer and the features of the adjacent decoding layer to obtain a flood water body probability map. In this embodiment, the decoding stage restores spatial resolution through layer-by-layer upsampling, while fusing high-level semantic features with low-level detail features from adjacent layers, thus balancing semantic information and boundary details. Cross-layer feature fusion can compensate for the deficiencies of single-layer features in terms of resolution or semantic expression, enabling the network to maintain overall consistency when identifying large bodies of water and clear boundaries when identifying small channels or localized water accumulation. The output is a probability map, where each pixel corresponds to a probability value belonging to a flooded body, realizing a mapping from the feature space to the probability space. Multi-scale semantic reconstruction improves segmentation accuracy, avoiding oversmoothing or boundary breakage problems caused by single-scale prediction, and providing high-quality input for subsequent threshold segmentation and rule suppression.

[0026] Step S5: Calculate the slope threshold of the digital elevation model using the digital elevation model data, and segment the flood water body probability map into a binary mask based on the slope threshold of the digital elevation model. In this embodiment, the slope value of each pixel is calculated using DEM data. When the slope exceeds a preset threshold, the area is considered to not meet the terrain conditions for flood retention and is thus excluded during the probabilistic map segmentation process. Subsequently, threshold segmentation is performed on the probabilistic map, assigning a value of 1 to pixels with a probability value greater than the threshold and a value of 0 to pixels with a probability value less than the threshold, forming an initial binary mask. This incorporates the physical laws of terrain into the segmentation decision process, preventing high-slope areas from being misclassified as water bodies due to shadows or texture similarity. This enhances the terrain consistency of the results, reduces false alarms in mountainous areas, and improves the overall reliability and interpretability of the results.

[0027] Step S6: Calculate the interferometric coherence coefficient threshold based on the radar image, and suppress the interference information in the binary mask based on the interferometric coherence coefficient threshold to obtain the flood disaster range map.

[0028] In this embodiment, the interference coherence coefficient reflects the stability of ground features over time. Stable buildings typically exhibit high coherence, while flooded areas show significantly reduced coherence due to increased randomness in water surface scattering. By setting a threshold, pixels with coherence coefficients exceeding the threshold are zeroed out, retaining only areas with reduced coherence, thus effectively suppressing false alarms from urban buildings. Incorporating temporal information into the discrimination process allows the model to rely not only on spatial features but also on change information for filtering. This significantly reduces the false detection rate in urban areas, improves the accuracy and practicality of flood extent extraction, and ultimately outputs a physically reasonable, spatially coherent flood disaster extent map with minimal error.

[0029] In some possible implementations, the dual-channel convolutional neural network is configured with optical encoding channels and radar encoding channels of identical structure, and the parameters of the optical encoding channels and radar encoding channels are set separately; the optical encoding channel extracts the water index features of optical features; the radar encoding channel extracts the backscattering features and polarization features of radar features.

[0030] In this embodiment, the dual-channel convolutional neural network is designed with an optical encoding channel and a radar encoding channel that have identical structures but independent parameters. Identical structure means that the two encoding channels maintain consistency in terms of the number of convolutional layers, kernel size, downsampling method, and the number of feature levels, thus ensuring the hierarchical semantic alignment of features from different modalities. Parameter independence means that each channel learns data distribution patterns applicable to its respective modality, avoiding feature representation confusion caused by shared parameters. The optical encoding channel focuses on extracting features related to the spectral response of water bodies, such as the water index features reflected by the difference in reflection between green light and near-infrared bands, while automatically learning boundary, texture, and spectral combination information through convolutional layers. The radar encoding channel mainly extracts backscattering and polarization features, which reflect the roughness, structural morphology, and polarization response differences of ground features, and are crucial for identifying cloud-covered areas or water accumulation around urban buildings. Through separate modeling, the two types of features are fully expressed under their respective physical mechanisms, avoiding information conflicts or imbalances in dominance caused by simple splicing. By fully leveraging the spectral sensitivity of optical data and the all-weather penetration capability of radar data, complementary advantages can be achieved, maintaining stable flood identification performance even in complex weather, terrain, and urban environments, and improving the network's generalization ability and robustness.

[0031] Figure 2 A schematic diagram illustrating the acquisition of fusion features provided in embodiments of this disclosure; as shown Figure 2 As shown, in some possible implementations, step S3, calculating the channel weights of optical and radar features based on the optical and radar features of each coding level, and using digital elevation model data as prior information for spatial weighting to calculate the spatial weights of optical and radar features, and performing weighted fusion to obtain fused features, includes: step S31, calculating the average optical feature value of the optical coding channel and the average radar feature value of the radar coding channel based on the optical and radar features of each coding level; step S32, sharing the optical and radar coding channels, and obtaining the optical and radar feature channel weights based on the average radar feature value of the optical feature value; step S33, calculating the slope map based on the digital elevation model data; step S34, concatenating the optical and radar coding channels, and calculating the spatial attention map through a 7x7 convolutional layer based on the terrain information of the slope map to obtain the spatial weights of optical and radar features; and step S35, fusing the weighted optical and radar features to obtain fused features.

[0032] In this embodiment, the average values ​​of optical and radar features are calculated at each encoding level. This is essentially a global statistical operation used to obtain the overall response intensity of each channel within the spatial range, thus forming a channel-level descriptive vector. Subsequently, the optical and radar encoded channels are shared in a modeling manner, that is, the average values ​​of optical and radar features are mapped using the same weight generation mechanism to obtain corresponding channel weights. This approach ensures modal independence while maintaining consistent discrimination logic in weight allocation across different modes. A slope map is calculated based on digital elevation model data and introduced as spatial prior information into the spatial attention calculation process. A spatial attention map is generated by performing a 7×7 convolution operation on the stitched optical and radar features, thereby obtaining spatial weights. The 7×7 convolution kernel can capture terrain change trends within a large receptive field, allowing spatial weights to be adjusted not only based on feature responses but also in conjunction with terrain conditions. The optical and radar features, weighted by both channel and spatial weights, are then fused to obtain the fused features. The system achieves dual optimization by enhancing selective channel dimension and physical constraints in spatial dimension, enabling the network to adaptively emphasize key modalities and key regions, significantly reducing the misclassification rate in mountainous and urban areas, and improving the physical rationality and refined expression of the model output.

[0033] In some possible implementations, after calculating the digital elevation model slope threshold using digital elevation model data and segmenting the flood water body probability map into a binary mask based on the digital elevation model slope threshold, the method further includes: constructing a circular structuring element with a preset radius; performing an expansion operation on the binary mask using the circular structuring element and then performing an erosion operation to obtain a closing operation result; performing an erosion operation on the closing operation result and then performing an expansion operation to form an opening operation result; and using the opening operation result as the optimized binary mask.

[0034] In this embodiment, a circular structural element with a preset radius is constructed. This structural element serves as the basic kernel for morphological operations, and its shape more closely resembles the natural diffusion pattern of water bodies. Compared to square structural elements, it is more conducive to maintaining the natural smoothness of water body boundaries. Subsequently, this structural element is used to perform dilation and erosion operations on the binary mask, resulting in a closing operation. The closing operation fills small holes within the water body area and enhances regional connectivity. Next, the closing operation result is subjected to erosion and then dilation operations, resulting in an opening operation. The opening operation removes isolated noise points, eliminates small false detection areas, and smooths the boundaries. Finally, the opening operation result is output as the optimized binary mask. By correcting small errors in network prediction through geometric structural constraints, regional coherence and boundary smoothness are improved, noise interference is reduced, and the output flood range map more closely matches the actual water body distribution pattern, while also improving the reliability of subsequent statistical analysis and disaster assessment.

[0035] In some possible implementations, an interferometric coherence coefficient threshold is calculated based on radar imagery, and interference information in the binary mask is suppressed based on the interferometric coherence coefficient threshold to obtain a flood disaster extent map. This includes: calculating the interferometric coherence coefficients of pre-disaster and post-disaster radar images based on radar imagery; setting a coherence coefficient threshold based on the interferometric coherence coefficients; performing pixel-by-pixel judgment on the binary mask, retaining pixels whose coherence coefficients are less than the coherence coefficient threshold, and obtaining the flood disaster extent map.

[0036] In this embodiment, the interferometric coherence coefficient is calculated based on pre-disaster and post-disaster radar images. This coefficient reflects the stability of the same ground feature across different time phases. Stable buildings typically maintain high coherence, while flooded areas experience a significant decrease in coherence due to changes in water surface scattering characteristics. Subsequently, a coherence coefficient threshold is set based on statistical results, and a pixel-by-pixel assessment is performed on the binary mask, retaining only pixels with coherence coefficients less than the threshold and discarding pixels with higher coherence. This pixel-by-pixel filtering method essentially introduces temporal dimension change information into the spatial segmentation results, achieving a combination of change detection and semantic segmentation. This significantly reduces false alarms caused by urban buildings, roads, or other stable ground features, improving the authenticity and accuracy of flood extent identification. It is particularly effective in reducing misjudgments in high-density urban areas, thereby enhancing the overall reliability and application value of the results.

[0037] In some possible implementations, cross-level feature fusion outputs a probability map of flood bodies by computing a 1x1 convolutional layer.

[0038] In this embodiment, cross-level feature fusion calculates and outputs a flood probability map through a 1×1 convolutional layer. The 1×1 convolutional layer performs linear combination and channel compression on the fused multi-channel features, mapping high-dimensional features to a single-channel probability space while maintaining spatial resolution. Cross-level feature fusion integrates shallow detail information with deep semantic information, while the 1×1 convolution performs discriminative mapping on these fused features, weighting the responses of different feature channels to generate a probability value for each pixel belonging to a flood body. This avoids the parameter redundancy problem caused by complex fully connected layers, while maintaining computational efficiency and model stability. It improves the granularity and accuracy of probability prediction, giving the model stronger discriminative ability and higher computational efficiency in the output stage, providing a stable and reliable probabilistic foundation for subsequent threshold segmentation.

[0039] In some possible implementations, prior to spatial registration, the following steps are also included: atmospheric correction of optical images; noise suppression and radiometric calibration of radar images.

[0040] In this embodiment, atmospheric correction of optical images is used to eliminate the influence of atmospheric scattering and absorption on reflectivity, making images acquired at different times comparable and thus ensuring the accuracy of water index calculation and spectral feature extraction. Noise suppression of radar images can effectively reduce the interference of speckle noise on backscattering features, while radiometric calibration converts the original echo values ​​into physically meaningful backscattering coefficients, ensuring consistent scale for data in different scenarios. This can significantly improve data quality, providing stable input for subsequent deep network learning. It reduces the risk of error accumulation caused by data noise, improves the reliability and consistency of feature extraction, and enhances the accuracy and robustness of the overall flood identification process from the source.

[0041] In some possible ways, data acquisition includes: acquiring domestically produced Gaofen-2 optical images of the disaster area (panchromatic resolution 0.8m, multispectral resolution 3.2m), ESA Sentinel-1 SAR (Synthetic Aperture Radar) images (IW (Interferometric Wide Swath) mode, VV+VH (representing dual-polarization data combination under vertical transmission conditions) polarization), and 30-meter resolution ASTER GDEM (Advanced Spaceborne Thermal and Reflective Radiometer Global Digital Elevation Model) data.

[0042] Preprocessing: Optical images were atmospherically corrected using the FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) module in ENVI (Environment for Visualizing Images) software to eliminate the effects of atmospheric scattering and absorption. Image fusion technology was then used to fuse the multispectral and panchromatic images to obtain high-resolution multispectral images.

[0043] SAR imagery was precisely processed using SNAP software, including radiometric calibration, application of Lee Sigma filtering (7x7 window size) to suppress speckle noise, and topographic correction and geocoding, and was accurately registered to the WGS84 UTM coordinate system.

[0044] The DEM data is corrected and a slope map is calculated.

[0045] Feature input preparation involves extracting the Normalized Difference Water Index (NDWI) as the key spectral feature from the preprocessed optical image. The formula for calculating NDWI is: ; in, It is in the green light band. Located in the near-infrared band, this index can effectively enhance water body information.

[0046] VV polarization backscattering coefficients (expressed in dB) and polarization characteristics are extracted from SAR images.

[0047] The NDWI feature map, SAR backscattering coefficient map, Pauli decomposition (a polarization decomposition method for analyzing fully polarimetric SAR data) feature map, and DEM slope map are used as inputs to the network.

[0048] The dual-branch coding and attention fusion method uses two encoder branches with the same structure but independent parameters to process optical features (NDWI, Normalized Difference Water Index) and SAR features (backscattering, polarization) respectively.

[0049] At each encoding level, perform the following operations: Channel attention, for optical feature maps and SAR feature map Global average pooling is performed separately to obtain one-dimensional channel descriptors. Then, channel attention weights are generated through a shared two-layer neural network (MLP, Multi-Layer Perceptron). . ; in, It is the Sigmoid activation function, GAPGAP represents global average pooling, and MLP is a multilayer perceptron.

[0050] Spatial attention (integrating terrain priors) stitches optical and SAR features along the channel dimension and inputs the DEM slope map S as guiding information, generating a spatial attention map through a 7x7 convolutional layer. To highlight important spatial locations and suppress shadow / overlay areas.

[0051] ; Feature weighting and fusion, ultimately, the fused features at this level. It is calculated by the following formula: ; It can dynamically and adaptively assign appropriate weights to features from different sources.

[0052] The decoder gradually restores spatial resolution through upsampling and convolution operations. During feature fusion, the decoder not only uses standard skip connections (fusing features from the same level encoder) but also introduces cross-level connections, fusing features from deeper layers (higher semantic levels) with the currently decoded features to capture multi-scale information and optimize boundaries. The network ultimately outputs a probability map of each pixel representing a flood body through a 1x1 convolutional layer and a sigmoid activation function. The value range is [0,1].

[0053] Binarization and morphological optimization, setting a threshold Convert the probability map into a binary mask. .

[0054] ; Subsequently, morphological closing operations (3x3 circular structuring elements) were used to fill the small holes, and opening operations were used to remove isolated noise points.

[0055] Multi-rule interference suppression: Terrain shadow suppression: Using the DEM slope map S, a slope threshold is set. The region in the initial mask with a slope greater than this threshold is set to 0 (non-water body), and the mask is generated. .

[0056] Urban Overlay / Shadow Suppression: Calculating the Interferometric Coherence Coefficient of SAR Imagery The formula for calculating the coherence coefficient is: ; in, and These are SAR composite images from before and after the disaster, respectively. * indicates complex conjugate. This calculates the number of pixels within the window (typically 5x5 or 7x7). Coherence values ​​range from [0,1], with lower values ​​indicating poorer ground feature scattering stability, typically corresponding to unreliable areas such as overlays and shadows. A coherence coefficient threshold is set. ,Will Regions with a coherence coefficient below this threshold are set to 0, resulting in the final, optimized flood inundation extent map. .

[0057] After processing the aforementioned urban flood-affected areas, the final result achieved an Intersection over Union (IoU) ratio of 92.5% and an F1 score of 96.1%. Detailed comparison results with the baseline model (standard U-Net) are shown in Table 1.

[0058] Table 1 The results show that the present invention significantly improves performance, especially in densely built-up urban areas and shaded areas in front of mountains, fully demonstrating the effectiveness and advancement of the present invention.

[0059] Figure 3 This is a schematic diagram of a flood disaster prediction system provided in an embodiment of this disclosure; as shown below. Figure 3 As shown, this disclosure also provides a flood disaster prediction system, including: an acquisition module 401, used to acquire optical remote sensing images, radar images, and digital elevation model data, and unify them into the same coordinate system for spatial registration; an extraction module 402, used to extract features from the optical remote sensing images and radar images respectively after spatial registration, and input them respectively into the optical encoding channel and radar encoding channel of a preset dual-channel convolutional neural network to extract optical features and radar features at each encoding level; and a calculation module 403, used to calculate the channel weights of the optical features and radar features based on the optical features and radar features at each encoding level, using digital elevation model data as spatial weighting. The system calculates the spatial weights of optical and radar features based on prior information and performs weighted fusion to obtain fused features. The fusion module 404 is used to perform cross-level feature fusion of the fused features with features from any layer of the preset dual-channel convolutional neural network decoding layer and features from adjacent decoding layers to obtain a flood water body probability map. The segmentation module 405 is used to calculate the digital elevation model slope threshold using digital elevation model data and segment the flood water body probability map into a binary mask based on the digital elevation model slope threshold. The suppression module 406 is used to calculate the interferometric coherence coefficient threshold based on radar imagery and suppress interference information in the binary mask based on the interferometric coherence coefficient threshold to obtain a flood disaster range map.

[0060] This disclosure also provides a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of a flood disaster prediction method.

[0061] This disclosure also provides a computer-readable storage medium storing computer-readable instructions, which, when executed by a processor, implement the steps of a flood disaster prediction method.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0063] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting flood disasters, characterized in that, include: Acquire optical remote sensing images, radar images, and digital elevation model data, and unify them into the same coordinate system for spatial registration; After spatial registration, features of optical remote sensing images and radar images are extracted respectively, and then input into the optical coding channel and radar coding channel of a preset dual-channel convolutional neural network to extract optical features and radar features at each coding level. Based on the optical and radar features of each coding level, the channel weights of the optical and radar features are calculated respectively. The spatial weights of the optical and radar features are calculated using digital elevation model data as prior information for spatial weighting, and then weighted fusion is performed to obtain the fused features. The fused features are fused with features from any layer of the preset dual-channel convolutional neural network decoding layer and features from adjacent decoding layers to obtain a flood probability map. The slope threshold of the digital elevation model is calculated using digital elevation model data, and the probability map of the flood water body is segmented into a binary mask based on the slope threshold of the digital elevation model. The interferometric coherence coefficient threshold is calculated based on the radar image, and the interference information in the binary mask is suppressed based on the interferometric coherence coefficient threshold to obtain a flood disaster range map.

2. The flood disaster prediction method according to claim 1, characterized in that, The dual-channel convolutional neural network is configured with optical encoding channels and radar encoding channels having the same structure, and the parameters of the optical encoding channels and radar encoding channels are set respectively; The optical coding channel extracts water body index features based on optical characteristics; The radar coding channel extracts the backscattering and polarization features of the radar characteristics.

3. The flood disaster prediction method according to claim 1, characterized in that, The process involves calculating the channel weights of optical and radar features based on the optical and radar features at each coding level, using digital elevation model data as prior information for spatial weighting to calculate the spatial weights of the optical and radar features, and then performing weighted fusion to obtain the fused features, including: Based on the optical and radar features of each coding level, the average optical features of the optical coding channel and the average radar features of the radar coding channel are calculated respectively. The optical coding channel and the radar coding channel are shared, and the optical feature channel weight and the radar feature channel weight are obtained based on the average value of the average value of the optical features. Calculate the slope map based on the digital elevation model data; The optical coding channel and the radar coding channel are spliced ​​together. Based on the terrain information of the slope map, a spatial attention map is calculated through a 7 x 7 convolutional layer to obtain the spatial weights of optical features and radar features. The weighted optical features and radar features are fused together to obtain the fused features.

4. The flood disaster prediction method according to claim 1, characterized in that, After the steps of calculating the digital elevation model slope threshold using digital elevation model data and segmenting the flood water body probability map into a binary mask based on the digital elevation model slope threshold, the method further includes: Construct a circular structural element with a preset radius; The binary mask is expanded by the circular structural element and then eroded to obtain the closing operation result. After performing an erosion operation on the closing operation result, an expansion operation is performed to form an opening operation result; The result of the opening operation is used as the optimized binary mask.

5. The flood disaster prediction method according to claim 1, characterized in that, The step of calculating the interferometric coherence coefficient threshold based on radar imagery and suppressing interference information in the binary mask based on the interferometric coherence coefficient threshold to obtain a flood disaster range map includes: Calculate the interference coherence coefficient between pre-disaster and post-disaster radar images based on radar imagery. A coherence coefficient threshold is set based on the interference coherence coefficient; The binary mask is evaluated pixel by pixel, and pixels with a coherence coefficient less than the coherence coefficient threshold are retained to obtain a flood disaster range map.

6. The flood disaster prediction method according to any one of claims 1 to 5, characterized in that, The cross-level feature fusion is used to calculate and output a probability map of flood water bodies through a 1x1 convolutional layer.

7. The flood disaster prediction method according to any one of claims 1 to 5, characterized in that, Before spatial registration, the following is also included: Atmospheric correction is performed on the optical image; The radar image is subjected to noise suppression and radiometric calibration.

8. A flood disaster prediction system, characterized in that, include: The acquisition module is used to acquire optical remote sensing images, radar images, and digital elevation model data, and to unify them into the same coordinate system for spatial registration. The extraction module is used to extract features from optical remote sensing images and radar images after spatial registration, and input them into the optical coding channel and radar coding channel of a preset dual-channel convolutional neural network to extract optical features and radar features at each coding level. The calculation module is used to calculate the channel weights of optical features and radar features based on the optical features and radar features of each coding level, respectively. It uses digital elevation model data as prior information for spatial weighting to calculate the spatial weights of optical features and radar features, and performs weighted fusion to obtain fused features. The fusion module is used to perform cross-level feature fusion of the fused features with features of any layer in the preset dual-channel convolutional neural network decoding layer and features of adjacent decoding layers to obtain a flood water body probability map; The segmentation module is used to calculate the slope threshold of the digital elevation model through the digital elevation model data, and to segment the flood water body probability map into a binary mask based on the slope threshold of the digital elevation model. The suppression module is used to calculate the interferometric coherence coefficient threshold based on the radar image, and suppress the interference information in the binary mask based on the interferometric coherence coefficient threshold to obtain a flood disaster range map.

9. A computer device, characterized in that, The method includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the flood disaster prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the flood disaster prediction method as described in any one of claims 1 to 7.

Images