SAR image flood region extraction method and system based on multi-source feature fusion

Abstract

The application provides a SAR image flood area extraction method and system based on multi-source feature fusion, and belongs to the technical field of remote sensing disaster monitoring. The extraction method is applied to a system comprising an image processing module, a model training module and a flood area extraction module. Specifically, the SAR image is preprocessed, and corresponding polarization physical features and time sequence abnormal features are extracted. A flood learning sample is constructed in combination with preset label data to train and learn a U-shaped network of a fusion residual network and obtain a pre-training model. Based on transfer learning, the pre-training model is adjusted in a geometric target scene to obtain a flood area recognition model. Based on the flood area recognition model, a flood area mask graph corresponding to a target SAR image is output. The application introduces a transfer model on the basis of the fusion of multi-dimensional features, combines an improved U-shaped network architecture, effectively improves the recognition and extraction accuracy of the flood area, avoids the risk of missed detection and false detection, and has high adaptability.

Description

Technical Field

[0001] This invention relates to the field of remote sensing disaster monitoring technology, and in particular to a method and system for extracting flood-prone areas from SAR images based on multi-source feature fusion. Background Technology

[0002] Most existing methods for identifying flood-prone areas rely on optical imagery as the data source. However, optical imagery, which depends on visible and near-infrared imaging, is easily obscured by common weather conditions in flood-prone areas, such as rain and clouds. This makes it difficult to penetrate cloud and rain interference to obtain clear information about ground features, leading to low accuracy in flood disaster monitoring. In contrast, synthetic aperture radar (SAR) imagery can effectively penetrate cloud and rain barriers and can reflect surface cover characteristics through differences in the backscattering coefficients of different ground features. It has gradually become the main data source for identifying flood-prone areas.

[0003] Currently, methods for extracting flooded areas based on SAR imagery mainly include traditional threshold segmentation, single-feature machine learning, and U-shaped network segmentation. Traditional threshold segmentation heavily relies on the bimodal distribution of the gray-level histogram of the backscattering coefficient in SAR images, resulting in a high false positive rate and low extraction accuracy in complex scenes. Single-feature machine learning relies solely on the backscattering coefficient, easily confusing the physical differences between temporary flooded areas and areas with higher humidity, leading to missed detections of flooded areas. While U-shaped network segmentation can improve feature extraction capabilities through deep learning, overcoming the problems of traditional threshold segmentation and single-feature machine learning, and improving extraction accuracy while avoiding missed detections, it still suffers from issues such as limited input features, insufficient ability to distinguish complex scattering mechanisms, gradient vanishing as network depth increases, unstable model training, low segmentation accuracy, and inability to adapt to feature differences in different scenes. Directly transferring a model trained in one area to other areas results in a significant decrease in extraction accuracy. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of low accuracy, weak anti-interference ability, and poor generalization ability when using U-shaped networks for SAR-based flood area extraction. It provides a method and system for SAR image flood area extraction based on multi-source feature fusion. By fusing polarimetric physical features and temporal anomaly features, it distinguishes flood areas from easily confused ground features, avoiding missed and false detections. Furthermore, by improving the U-shaped network architecture, it solves the gradient vanishing problem in deep networks, improving the accuracy of flood area boundary segmentation. Finally, through transfer learning, it enhances the model's adaptability to different distribution feature scenarios, ensuring cross-regional monitoring accuracy and achieving high-precision flood area extraction.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] Methods for extracting flood-affected areas from SAR images based on multi-source feature fusion include:

[0007] SAR images are preprocessed, and polarization physical features and temporal anomaly features are extracted based on the preprocessed SAR images. Flood learning samples are then constructed by combining them with pre-defined label data.

[0008] The U-shaped network with fused residual network is trained and learned using flood learning samples to obtain a pre-trained model;

[0009] Based on transfer learning, the pre-trained model is adjusted by combining the target scene corresponding to the target SAR image to obtain a flood area identification model;

[0010] Based on the flood area identification model, the output is a flood area mask map corresponding to the target SAR image.

[0011] Furthermore, the preprocessing of the SAR image includes:

[0012] Radiometric correction of SAR images is performed by combining absolute calibration factor and offset.

[0013] Geometric correction is performed on the radiometrically corrected SAR image using the orbital parameters and ground control points of the SAR image.

[0014] The geometrically corrected SAR image is filtered to obtain the preprocessed SAR image.

[0015] Furthermore, the extraction of polarization physical features and temporal anomaly features based on the preprocessed SAR image includes:

[0016] Acquire single-view complex data of preprocessed SAR images, construct a dual-polarization scattering matrix based on the single-view complex data, and convert the scattering matrix into a target vector;

[0017] Based on the target vector, the corresponding coherence matrix is ​​calculated, and the coherence matrix is ​​decomposed into eigenvalues ​​to obtain the polarization physical characteristics.

[0018] Obtain ground distance detection data from preprocessed SAR images of historical time phases, use the ground distance detection data as the historical baseline dataset, and calculate the mean and standard deviation of the backscattering coefficient of each pixel in the ground distance detection data;

[0019] Based on the preprocessed SAR image of the current time phase, the temporal anomaly score of each pixel in the preprocessed SAR image of the current time phase is calculated by combining the mean and standard deviation of the calculated scattering coefficients, and the temporal anomaly features are identified by combining the temporal anomaly score threshold.

[0020] Furthermore, the polarization physical characteristics include at least polarization entropy, scattering angle, and anisotropy.

[0021] Furthermore, the construction of flood learning samples by combining preset label data includes:

[0022] Using pre-labeled flood disaster distribution maps as preset label data, polarization physical features and temporal anomaly features are matched with the corresponding pixels of the preset label data according to pixel position, and corresponding feature label pairs are generated based on the matching results;

[0023] The feature label pairs are cropped into flood learning samples of a preset pixel size.

[0024] Furthermore, the step of training the U-shaped network with fused residual network using flood learning samples to obtain a pre-trained model includes:

[0025] The flood learning samples are preprocessed and split into training and validation sets according to a preset ratio;

[0026] Initialize the U-shaped network with fused residual networks and set the training parameters;

[0027] Based on the training parameters, during each training iteration, batch samples are randomly selected from the training set and input into the U-shaped network of the fused residual network for forward propagation. The average loss value of each batch of samples is calculated, and the network parameters are updated through the backpropagation algorithm.

[0028] The network performance metrics are calculated using the validation set, and the network parameters are updated based on these metrics until the iteration termination condition in the training parameters is met, thus obtaining the pre-trained model.

[0029] Furthermore, the step of adjusting the pre-trained model based on transfer learning and combining it with the target scene corresponding to the target SAR image to obtain the flood area identification model includes:

[0030] Acquire historical SAR images of the target scene corresponding to the target SAR image, preprocess the historical SAR images, and construct a fine-tuning training set;

[0031] Adjust the training parameters, freeze the low-level encoder weights in the pre-trained model, and train the pre-trained model by fine-tuning the training set to obtain a flood area identification model.

[0032] Furthermore, the step of outputting a flood area mask map corresponding to the target SAR image based on the flood area identification model includes:

[0033] Preprocess the target SAR image and extract the corresponding polarization physical features and temporal anomaly features;

[0034] The extracted polarization physical features and temporal anomaly features are input into the flood area identification model, and the corresponding flood attribution probability map is output.

[0035] The flood attribution probability map is binarized to obtain the corresponding flood area mask map.

[0036] A SAR image flood area extraction system based on multi-source feature fusion, used to perform any of the above-mentioned SAR image flood area extraction methods based on multi-source feature fusion, includes:

[0037] The image processing module is used to preprocess SAR images and extract polarization physical features and temporal anomaly features based on the preprocessed SAR images, and construct flood learning samples by combining them with preset label data.

[0038] The model training module is used to train and learn the U-shaped network of the fused residual network through flood learning samples, obtain a pre-trained model, and adjust the pre-trained model based on transfer learning and the target scene corresponding to the target SAR image to obtain a flood area identification model.

[0039] The flood area extraction module is used to output a flood area mask map corresponding to the target SAR image based on the flood area identification model.

[0040] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the SAR image flood area extraction method based on multi-source feature fusion as described above.

[0041] The beneficial effects of this invention are:

[0042] Polarimetric physical features are extracted from single-view complex data to accurately distinguish flooded areas based on differences in scattering mechanisms. Temporal anomaly features are calculated based on historical ground distance detection data to identify temporary flooding and easily confused features by analyzing temporal variations in backscattering coefficients. These two feature sets enhance the subsequent model's ability to distinguish easily confused features, reducing the risk of false positives and false negatives. Furthermore, a residual network is introduced into the U-shaped network architecture to address the vanishing gradient problem in deep networks, ensuring accurate segmentation of flooded area boundaries. Transfer learning is further combined, freezing the low-level weights of the pre-trained model encoder and fine-tuning only the high-level encoder and decoder to adapt to the target scene, thereby improving the model's cross-regional generalization ability and achieving efficient and accurate identification of SAR-affected flooded areas. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of a process of the present invention;

[0044] Figure 2 This is a U-shaped network architecture diagram of a fused residual network according to an embodiment of the present invention. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0046] Example: A method for extracting flood-affected areas from SAR images based on multi-source feature fusion, such as... Figure 1 As shown, it includes:

[0047] SAR images are preprocessed, and polarization physical features and temporal anomaly features are extracted based on the preprocessed SAR images. Flood learning samples are then constructed by combining them with pre-defined label data.

[0048] The U-shaped network with fused residual network is trained and learned using flood learning samples to obtain a pre-trained model;

[0049] Based on transfer learning, the pre-trained model is adjusted by combining the target scene corresponding to the target SAR image to obtain a flood area identification model;

[0050] Based on the flood area identification model, the output is a flood area mask map corresponding to the target SAR image.

[0051] Floods are often accompanied by rain and cloud cover. Optical imagery, relying on visible and near-infrared bands, is easily obscured by clouds and rain, making it difficult to obtain effective surface information. SAR imagery, operating on the microwave band, has all-weather, day-and-night capability. Regardless of day or night or complex weather conditions, it can penetrate cloud and rain barriers to reliably acquire monitoring data, ensuring data continuity and timeliness during disasters. Furthermore, the backscattering coefficient of SAR imagery directly reflects the scattering characteristics of ground features to radar waves. The core features in flooded areas are water bodies, which primarily exhibit specular reflection, resulting in a significantly lower backscattering coefficient than features in non-flooded areas such as buildings and vegetation. Utilizing this physical difference, flooded and non-flooded areas can be quickly and accurately distinguished. Therefore, this embodiment specifically uses SAR imagery as the data source for flood area identification, ensuring the reliability of the identification results.

[0052] In this embodiment, Sentinel-1 dual-polarization SAR imagery is specifically used, which can provide two types of data: single-view complex data and ground distance detection data. The single-view complex data retains amplitude and phase information and can be used to construct a scattering matrix and extract polarization physical features to reflect the differences in the scattering mechanism of ground objects. The ground distance detection data has spatial consistency after geometric correction and can be used to construct historical baseline datasets and calculate temporal anomaly features to identify dynamic changes in ground objects.

[0053] The original SAR images are affected by systematic errors during the acquisition process, resulting in certain deviations that can affect the accuracy of subsequent feature extraction and model training. Therefore, all acquired SAR images, including historical SAR images and target SAR images, are preprocessed.

[0054] The preprocessing of SAR images includes:

[0055] Radiometric correction of SAR images is performed by combining absolute calibration factor and offset.

[0056] Geometric correction is performed on the radiometrically corrected SAR image using the orbital parameters and ground control points of the SAR image.

[0057] The geometrically corrected SAR image is filtered to obtain the preprocessed SAR image.

[0058] The digital quantization value of SAR imagery is an integer obtained by electronically quantizing the radar echo signal from the sensor. It has no actual physical unit and is affected by system factors such as sensor gain fluctuations, electronic noise, and orbital altitude differences. Radiometric correction can solve the problem that the digital quantization value of the original SAR imagery has no physical meaning and is affected by system error interference. Specifically, it converts the SAR imagery into backscattering coefficients through absolute calibration factors and offsets, and then converts them into dB form. This gives the data a physical meaning that reflects the scattering ability of ground objects, and eliminates system errors such as sensor gain and electronic noise. This makes images from different time phases and different regions comparable and avoids feature calculation deviations caused by the lack of physical meaning in the data.

[0059] The conversion formula for the backscattering coefficient is:

[0060] ;

[0061] in, The backscattering coefficient is... For SAR image digital quantization values, As the absolute scaling factor, The offset is the signal gain coefficient calibrated through satellite calibration experiments, used to compensate for signal attenuation during radar echo propagation and reception. The offset is a system noise correction term used to counteract the interference of sensor's own electronic noise on the digital quantization value.

[0062] Regarding the backscattering coefficients obtained from the conversion, specifically through... Converting it to dB form, compressing the numerical range, and transforming the multiplication relationship of the signal into an addition relationship are more in line with the logic of radar signal analysis.

[0063] Since SAR images employ the side-looking imaging principle, the original images may suffer from geometric distortion due to terrain undulations. The spatial positions of images from different time phases may also be misaligned due to orbital offsets. For radiometrically corrected SAR images, based on the WGS84 coordinate system, the terrain distortion and geometric offset are corrected by using the orbital parameters of the SAR images and ground control points to ensure that all image pixels correspond to fixed ground positions.

[0064] The orbital parameters include satellite longitude, latitude, altitude, and flight direction, which are used to initially determine the spatial coordinate system of the image. The ground control points are landmarks with known longitude and latitude, GPS measurement points, etc., which are used to correct minor deviations in the orbital parameters.

[0065] Based on orbital parameters and ground control points, each pixel in the SAR image is mapped to the WGS84 coordinate system, so that each pixel corresponds to a unique latitude and longitude position on the ground, thus eliminating geometric distortion.

[0066] Due to the coherent imaging characteristics of radar waves, SAR images inherently produce speckle noise, manifested as alternating bright and dark granular textures. This noise does not reflect the true scattering characteristics of ground objects but is the result of coherent signal superposition, interfering with subsequent extraction of ground object scattering features. Therefore, the Lee filtering algorithm is further employed to filter the geometrically corrected SAR images to reduce noise interference and ensure the reliability of the SAR images.

[0067] Based on the preprocessed SAR effects, a multi-source feature system is constructed from two dimensions: extreme values ​​of ground object scattering and spatiotemporal dynamic changes, in order to achieve accurate differentiation between flooded areas and easily confused ground objects, such as vegetated wet areas, permanent water bodies, and radar shadows.

[0068] To ensure the accuracy and reliability of feature extraction, the extraction of polarization physical features and temporal anomaly features based on preprocessed SAR images includes:

[0069] Acquire single-view complex data of the preprocessed SAR image, construct a dual-polarization scattering matrix based on the single-view complex data, and convert the scattering matrix into a target vector;

[0070] Based on the target vector, the corresponding coherence matrix is ​​calculated, and the coherence matrix is ​​decomposed into eigenvalues ​​to obtain the polarization physical characteristics.

[0071] Obtain ground distance detection data from preprocessed SAR images of historical time phases, use the ground distance detection data as the historical baseline dataset, and calculate the mean and standard deviation of the backscattering coefficient of each pixel in the ground distance detection data;

[0072] Based on the preprocessed SAR image of the current time phase, the temporal anomaly score of each pixel in the preprocessed SAR image of the current time phase is calculated by combining the mean and standard deviation of the calculated scattering coefficients, and the temporal anomaly features are identified by combining the temporal anomaly score threshold.

[0073] Furthermore, the polarization physical characteristics include at least polarization entropy, scattering angle, and anisotropy.

[0074] The preprocessed single-view complex data retains the amplitude and phase information of the radar echo. Based on this, a dual-polarization scattering matrix is ​​constructed and converted into a target vector. Then, the polarization entropy, scattering angle, and anisotropy obtained through eigenvalue decomposition of the coherence matrix can directly reflect the scattering mechanism of the ground object. This feature difference based on the scattering mechanism can effectively distinguish ground objects with similar backscattering coefficients but different physical properties, avoiding the risk of misjudgment caused by relying solely on scattering intensity.

[0075] The dual-polarization scattering matrix constructed using single-view complex data is as follows:

[0076] ;

[0077] in, This is the dual-polarization scattering matrix. and For the same polarization component, and For cross-polarization components, and ignoring non-reciprocal effects, .

[0078] The scattering matrix is ​​then transformed into a target vector using Pauli basis expansion, converting the matrix-form scattering information into a low-dimensional vector. This process preserves the core polarization features while reducing subsequent computational complexity. The transformation formula is as follows:

[0079] ;

[0080] in, Let T be the target vector, and T be the matrix transpose.

[0081] The 2×2 coherence matrix is ​​calculated based on the target vector to eliminate local noise interference through spatial averaging and stably characterize the polarization scattering properties of ground objects. Its expression is:

[0082] ;

[0083] in, For coherence matrix, The average value is for a 5x5 window space. This is the conjugate transpose.

[0084] For coherence matrix Eigenvalue decomposition yields two eigenvalues. and eigenvalues Then, the corresponding polarization entropy, scattering angle, and anisotropy are calculated to obtain the polarization physical characteristics.

[0085] The calculation expression for polarization physical characteristics is as follows:

[0086] ;

[0087] ;

[0088] ;

[0089] in, The polarization entropy has a value range of [0, 1]. The scattering angle ranges from [0° to 90°]. It is anisotropic, and its value range is [0, 1]. The modulus of the first element in the eigenvector corresponding to the largest eigenvalue in the coherence matrix. and The first The eigenvalue and the eigenvalue Each feature value.

[0090] The temporal anomaly score is then calculated based on the preprocessed SAR images of the current time phase and historical statistical values, and the temporal anomaly characteristics are identified in order to capture the dynamic changes of the flood area from a spatiotemporal perspective.

[0091] Extract the backscattering coefficient of each pixel from the ground distance detection data of the current time phase, i.e., the flood season. Combine this with the mean and standard deviation of the backscattering coefficient of that pixel's historical baseline to calculate the temporal anomaly score, i.e., the Z score. The calculation expression is as follows:

[0092] ;

[0093] in, For time-series anomaly scores, The backscattering coefficient of the pixel in the current time phase. This represents the mean of the corresponding backscattering coefficients. This represents the standard deviation of the corresponding backscattering coefficient.

[0094] In this embodiment, the timing anomaly score threshold is set to 1.5. When the timing anomaly score is lower than 1.5, it is identified as a timing anomaly feature.

[0095] The U-shaped network that integrates residual networks is a supervised learning model. Its training essentially involves iterative optimization through feature input and label feedback to learn the mapping rules between multi-source features and attributes of flooded and non-flooded areas. In order to adapt to the training requirements of the U-shaped network that integrates residual networks, flood learning samples are constructed based on pre-labeled data.

[0096] Specifically, the construction of flood learning samples by combining preset label data includes:

[0097] Using pre-labeled flood disaster distribution maps as preset label data, polarization physical features and temporal anomaly features are matched with the corresponding pixels of the preset label data according to pixel position, and corresponding feature label pairs are generated based on the matching results;

[0098] The feature label pairs are cropped into flood learning samples of a preset pixel size.

[0099] The preset label data is a real flood-affected area benchmark that has been labeled, such as the UNOSAT flood disaster distribution map, which can clearly define the attributes of each pixel.

[0100] In the early SAR image preprocessing, both the images from which multi-source features are extracted and the flood disaster distribution maps with preset labels have undergone geometric correction based on the WGS84 coordinate system. The pixels of both correspond one-to-one with the ground locations. Therefore, samples are constructed by matching the pixel locations to ensure that the feature data of each pixel can be bound to the flood label of the corresponding location.

[0101] Based on the bound feature data and flood labels, corresponding feature label pairs are generated. The feature label pairs are then cropped into flood learning samples of 256×256 pixels. At the same time, the dual polarization mode of the Sentinel-1 dual polarization SAR image corresponding to the feature label pairs can also be added to the flood learning samples.

[0102] Furthermore, considering that flooded areas typically account for a very small proportion of images during flood disasters, directly constructing samples according to the original proportions would result in a class imbalance problem where non-flooded area samples dominate. Therefore, when generating flood learning samples, a constraint condition is set that the pixel ratio of flooded areas to non-flooded areas in each sample is 1:1 to avoid model bias caused by sample imbalance and improve the sensitivity of flooded area recognition.

[0103] This embodiment specifically uses a U-shaped network with fused residual networks as the basic model for flood area identification. The U-shaped network with fused residual networks consists of an encoder, a bottleneck layer, a decoder, and an output layer, and its overall network architecture diagram is shown below. Figure 2 As shown.

[0104] The encoder comprises four modules, each consisting of two residual blocks. Each residual block is formed by a convolutional layer, batch normalization, ReLU activation, another convolutional layer, batch normalization, skip connections, and another ReLU activation. The convolutional layers have 3×3 kernels with 1 unit of padding. The number of output channels for each module is 32, 64, 128, and 256, respectively, and each module is followed by a 2×2 max-pooling downsampling layer.

[0105] The bottleneck layer includes two convolutional layers with a kernel size of 3×3, padding of 1, and 256 output channels, followed by batch normalization and ReLU activation.

[0106] The decoder consists of three modules, each consisting of a 2×2 transposed convolutional layer and two residual blocks, with output channels of 128, 64, and 32 respectively. The transposed convolutional layer is then concatenated with the feature map of the corresponding module of the encoder.

[0107] The output layer includes a 1×1 convolutional layer, which can reduce the number of channels to 1. The convolutional layer is followed by a sigmoid activation function, which can output a flood attribution probability map in the interval [0, 1].

[0108] Based on the U-shaped network architecture of the fusion residual network, it is first trained on flood training samples to learn the general features of SAR images, such as edges, textures, and basic scattering patterns. These general features are common in different scenarios and do not require retraining. Subsequently, the model parameters are fine-tuned through transfer learning to obtain a flood area identification model adapted to the corresponding scenario, which can effectively address the problem of scarce labeled samples in actual disaster scenarios.

[0109] Specifically, the step of training the U-shaped network with fused residual network using flood learning samples to obtain a pre-trained model includes:

[0110] The flood learning samples are preprocessed and split into training and validation sets according to a preset ratio;

[0111] Initialize the U-shaped network with fused residual networks and set the training parameters;

[0112] Based on the training parameters, during each training iteration, batch samples are randomly selected from the training set and input into the U-shaped network of the fused residual network for forward propagation. The average loss value of each batch of samples is calculated, and the network parameters are updated through the backpropagation algorithm.

[0113] The network performance metrics are calculated using the validation set, and the network parameters are updated based on these metrics until the iteration termination condition in the training parameters is met, thus obtaining the pre-trained model.

[0114] Because the numerical ranges of multi-source features differ significantly, if directly input into the model, features with larger numerical ranges will dominate gradient updates, causing the model to favor learning those features while ignoring other key features. Therefore, we first use Min-Max normalization or Z-score standardization to unify the numerical scale of all features.

[0115] Based on the preprocessed flood learning samples, the training set and validation set are divided in a 7:3 ratio, and the feature distribution of the two types of samples must be consistent. That is, the proportion of flooded areas and non-flooded areas and the statistical regularity of multi-source features in the training set and validation set are consistent with those in the original flood learning sample set.

[0116] Network initialization requires parameter initialization based on a U-shaped network of a pre-defined fusion residual network. The convolution kernel parameters of the four modules of the encoder, the bottleneck layer, and the three modules of the decoder are initialized using He to ensure that the mean and variance of the initial parameters are reasonable and to avoid gradient explosion or vanishing due to excessively large or small parameters in the early stage of training. The mean and variance of the batch normalization layer are initialized to 0 and 1 respectively to ensure the stability of the feature distribution of each input layer.

[0117] The training parameters were set based on the binary classification task in flood-prone areas and the U-shaped network characteristics of the fused residual network. The optimizer used was Adam (adaptive momentum estimation), with an initial learning rate of 0.001, which decayed to 0.9 every 10 iterations. The maximum number of iterations was set to 50, and training stopped when the F1 score on the validation set showed no improvement for 5 consecutive iterations. A binary cross-entropy loss function was used as the loss function during training, and its expression is:

[0118] ;

[0119] in, The loss value. To train a large number of flood learning samples, For the training set The label values ​​of each flood learning sample. For the training set The predicted probability of a flood training sample.

[0120] Furthermore, a regularized Dropout layer with a dropout rate of 0.3 is added after the encoder residual block to prevent overfitting of the model.

[0121] Iterative training is conducted based on the set training parameters. The model's ability to identify flood-prone area features is gradually improved by calculating the loss through forward propagation and updating the parameters through backpropagation. In each training iteration, a fixed batch of samples is randomly selected from the training set and input into the model to avoid the influence of sample order on training and to prevent the model from memorizing sample arrangement patterns.

[0122] During forward propagation, features are extracted layer by layer through residual blocks, and max pooling layers progressively compress the spatial dimension, transforming multi-source features into a highly abstract representation of flood-affected areas. Bottleneck layers further optimize the feature map output by the encoder, strengthening key feature signals. The decoder gradually restores the spatial dimension through transposed convolutional layers, and simultaneously concatenates the decoder's feature map with the encoder's feature map of the same dimension to supplement edge details captured by shallower networks. Finally, the output layer generates a flood attribution probability map. Subsequently, based on the binary cross-entropy formula, the loss is calculated for the predicted probability and the true label of each pixel. The average pixel loss of all samples within a batch is then taken to obtain the average loss for that batch.

[0123] During the backpropagation parameter update process, the loss gradient propagates backward along the direction of the decoder, bottleneck layer, and encoder. Skip connections directly pass the gradient to the shallow network, avoiding parameter update failure in the deep network due to gradient decay. The Adam optimizer combines gradient and momentum to adjust the convolution kernel weights and batch normalization parameters according to the current learning rate, so that the loss of the next training is closer to the minimum value, thus realizing network parameter update.

[0124] Furthermore, after each iteration, network performance metrics are calculated using the validation set, and the iteration termination condition is determined. Specifically, this embodiment employs... , The performance of the model is evaluated using four network performance metrics: score, recall, and precision. and The expression for calculating fractions is:

[0125] ;

[0126] ;

[0127] in, For intersection, union, and comparison, The probability of a true positive result. This represents the probability of a false positive. The probability of a true negative. for Score For accuracy, This refers to the recall rate.

[0128] If the validation set corresponding to the current iteration period If the score improves compared to the previous iteration, immediately save the current network parameters as temporary optimal parameters. If the validation set score improves for 5 consecutive iterations... If the score does not improve, the early stopping mechanism is triggered; or if the preset 50 iteration cycles have been completed, training is terminated. The network model corresponding to the final temporarily optimal parameters is the pre-trained model.

[0129] The acquired pre-trained model has learned the mapping rules between multi-source features and flooded areas in the source scene. The encoder's lower layer has learned general features such as SAR image edges and basic scattering rules, which can accurately extract flooded areas from the source scene. The source scene is a ground feature distribution scene with complete Sentinel-1 dual-polarization SAR data, sufficient historical baseline data, and relevant annotation labels.

[0130] Considering the differences between the source and target scenes, if the pre-trained model of the source scene is directly applied to the target scene, the accuracy of the model will drop significantly due to the mismatch between the memorized source scene features and the actual features of the target scene. Therefore, the pre-trained model is further adjusted through transfer learning to adapt to the feature differences of the target scene, so as to ensure the accuracy of the final flood area identification results.

[0131] The step of adjusting the pre-trained model based on transfer learning and combining it with the target scene corresponding to the target SAR image to obtain the flood area identification model includes:

[0132] Acquire historical SAR images of the target scene corresponding to the target SAR image, preprocess the historical SAR images, and construct a fine-tuning training set;

[0133] Adjust the training parameters, freeze the low-level encoder weights in the pre-trained model, and train the pre-trained model by fine-tuning the training set to obtain a flood area recognition model.

[0134] After acquiring historical SAR images of the target scene corresponding to the target SAR image, preprocessing is performed on them. The preprocessing operation is consistent with the preprocessing operation used in the previous model training process. The corresponding polarization physical features and temporal anomaly features are obtained, and a fine-tuning training set is constructed by combining the corresponding label data and the corresponding dual polarization mode.

[0135] The training parameters were adjusted, and the learning rate was reduced from 0.001 in the source scene to 0.0001, allowing the model to slowly learn the features of the target scene. This avoids the learning rate being too high and covering the low-level parameters that have already been optimized in the source scene. The training cycle was shortened from 50 to 30, reducing the risk of overfitting from small sample training. At the same time, the Dropout layer after the encoder residual block was retained to continuously suppress overfitting and ensure the model's generalization ability.

[0136] Since the first two modules of the encoder, namely the lower layer of the encoder, learn the general features of SAR images, these features are common to both the source and target scenes and do not need to be retrained, the weights of the lower layer of the encoder are frozen during the fine-tuning process.

[0137] The latter two modules of the encoder, namely the encoder high-level module, learn specific features and need to further adapt to the scattering extreme value differences of the target scene, while the decoder needs to optimize the spatial resolution recovery and boundary segmentation logic to adapt to the differentiated ground feature boundaries of the target scene. Therefore, during the fine-tuning process, the encoder high-level weights and decoder are unfrozen so that they can relearn the flood area distribution features of the target scene.

[0138] When training the pre-trained model by fine-tuning the training set, the flood attribution probability map of the target scene is output through pre-propagation. Then, the difference between the predicted value and the true label is calculated by combining the binary cross-entropy loss function. The parameters of the unfrozen part are updated only through backpropagation. The recognition logic of the target scene is gradually optimized to achieve the adaptation of the pre-trained model to the target scene. Finally, a flood area recognition model that can accurately identify the flood area of ​​the target scene is obtained.

[0139] Based on this, the flood area identification model outputs a flood area mask map corresponding to the target SAR image, including:

[0140] Preprocess the target SAR image and extract the corresponding polarization physical features and temporal anomaly features;

[0141] The extracted polarization physical features and temporal anomaly features are input into the flood area identification model, and the corresponding flood attribution probability map is output.

[0142] The flood attribution probability map is binarized to obtain the corresponding flood area mask map.

[0143] The preprocessing and feature extraction operations for target SAR images are consistent with those used in the previous model training process.

[0144] The extracted polarization physical features and temporal anomaly features are then input into the flood area identification model along with the dual polarization mode of the target SAR image to output the corresponding flood attribution probability map.

[0145] Then, using 0.5 as the binarization threshold, the flood attribution probability map is segmented. Areas with a probability value greater than or equal to 0.5 are identified as flooded areas and marked as 1, while areas with a probability value less than 0.5 are identified as non-flooded areas and marked as 0. The binarized two-dimensional matrix is ​​the flood area mask map, and the spatial resolution of the flood area mask map is consistent with that of the target SAR image.

[0146] Another aspect of this embodiment provides a SAR image flood area extraction system based on multi-source feature fusion, including:

[0147] The image processing module is used to preprocess SAR images and extract polarization physical features and temporal anomaly features based on the preprocessed SAR images, and construct flood learning samples by combining them with preset label data.

[0148] The model training module is used to train and learn the U-shaped network of the fused residual network through flood learning samples, obtain a pre-trained model, and adjust the pre-trained model based on transfer learning and the target scene corresponding to the target SAR image to obtain a flood area identification model.

[0149] The flood area extraction module is used to output a flood area mask map corresponding to the target SAR image based on the flood area identification model.

[0150] The image processing module, model training module, and flood area extraction module are all equipped with computers and other devices that have corresponding data processing capabilities, and each has a corresponding external communication interface to obtain the required data.

[0151] Another aspect of this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the SAR image flood area extraction method based on multi-source feature fusion described in this embodiment.

[0152] Taking the extraction of flood-prone areas from a single-peak water body distribution scenario as an example, a pre-trained model is first constructed using a double-peak water body distribution scenario as the source scenario. The data source is the Sentinel-1 IW image of this scenario on a certain day, with a resolution of 5m×20m. The historical baseline data is the ground distance detection data of the same period in the past 4 years. The model parameters are: encoder with 4 residual blocks, decoder transposed convolution stride of 2, optimizer Adam with an initial learning rate of 0.001, and maximum training epochs of 50. Preprocessing and feature extraction are performed. In the preprocessing, the absolute scaling factor for radiometric correction is set to 0.0015, and the offset is set to 0.02.

[0153] After preprocessing and feature extraction, corresponding flood training samples are constructed, and model training is performed. After training, the validation set is then... The score was 0.9394, the precision was 0.9380, and the recall was 0.9408. The value was 0.882, resulting in a pre-trained model with a radar shadow misclassification rate of 3.2% and a permanent water body misclassification rate of 2.8%.

[0154] Based on the pre-trained model, Sentinel-1 IW mode imagery of the target scene from a past day was extracted, with a resolution of 5m × 20m. Preprocessing and feature extraction were performed, yielding a polarization entropy of 0.32, a scattering angle of 35°, an anisotropy of 0.28, and a Z-score of -1.8. A fine-tuning dataset was constructed based on the extracted feature data and corresponding annotations. Simultaneously, training parameters were adjusted: the learning rate was set to 0.0001, the maximum number of training epochs was set to 30, and the first two modules of the encoder were frozen. The network parameters of the last two modules of the encoder and the decoder were then adjusted according to the fine-tuning dataset. After training, a validation set was established. The score was 0.899, the precision was 0.912, and the recall was 0.887. With a value of 0.805, and based on the low radar shadow misjudgment rate and permanent water body misjudgment rate, a flood area identification model with a mountain shadow misjudgment rate of 4.5% and a solution depression water accumulation missed detection rate of 6.3% was obtained. This significantly reduced the probability of misjudgment and missed detection in the flood area identification process, and can achieve accurate flood area identification and extraction of target scenes.

[0155] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications are possible without departing from the technical solutions described in the claims.

Claims

1. A method for extracting flood-affected areas from SAR images based on multi-source feature fusion, characterized in that, include: SAR images are preprocessed, and polarization physical features and temporal anomaly features are extracted based on the preprocessed SAR images. Flood learning samples are constructed by combining them with pre-defined label data. The U-shaped network with fused residual network is trained and learned using flood learning samples to obtain a pre-trained model; Based on transfer learning, the pre-trained model is adjusted by combining the target scene corresponding to the target SAR image to obtain a flood area identification model; Based on the flood area identification model, output the flood area mask map corresponding to the target SAR image; The extraction of polarization physical features and temporal anomaly features based on preprocessed SAR images includes: Acquire single-view complex data of the preprocessed SAR image, construct a dual-polarization scattering matrix based on the single-view complex data, and convert the scattering matrix into a target vector; Based on the target vector, the corresponding coherence matrix is ​​calculated, and the coherence matrix is ​​decomposed into eigenvalues ​​to obtain the polarization physical characteristics. Obtain ground distance detection data from preprocessed SAR images of historical time phases, use the ground distance detection data as the historical baseline dataset, and calculate the mean and standard deviation of the backscattering coefficient of each pixel in the ground distance detection data; Based on the preprocessed SAR image of the current time phase, the temporal anomaly score of each pixel in the preprocessed SAR image of the current time phase is calculated by combining the calculated mean and standard deviation of the scattering coefficients. Temporal anomaly features are then identified by combining the temporal anomaly score threshold. When the time series anomaly score is lower than the time series anomaly score threshold, it is identified as a time series anomaly feature.

2. The method for extracting flooded areas from SAR images based on multi-source feature fusion according to claim 1, characterized in that, The preprocessing of SAR images includes: Radiometric correction of SAR images is performed by combining absolute calibration factor and offset. Geometric correction is performed on the radiometrically corrected SAR image using the orbital parameters and ground control points of the SAR image. The geometrically corrected SAR image is filtered to obtain the preprocessed SAR image.

3. The method for extracting flooded areas from SAR images based on multi-source feature fusion according to claim 1, characterized in that, The polarization physical characteristics include at least polarization entropy, scattering angle, and anisotropy.

4. The method for extracting flooded areas from SAR images based on multi-source feature fusion according to claim 1, characterized in that, The construction of flood learning samples by combining preset label data includes: Using pre-labeled flood disaster distribution maps as preset label data, polarization physical features and temporal anomaly features are matched with the corresponding pixels of the preset label data according to pixel position, and corresponding feature label pairs are generated based on the matching results; The feature label pairs are cropped into flood learning samples of a preset pixel size.

5. The method for extracting flooded areas from SAR images based on multi-source feature fusion according to claim 1, characterized in that, The process of training the U-shaped network with fused residual network using flood learning samples to obtain a pre-trained model includes: The flood learning samples are preprocessed and split into training and validation sets according to a preset ratio; Initialize the U-shaped network with fused residual networks and set the training parameters; Based on the training parameters, during each training iteration, batch samples are randomly selected from the training set and input into the U-shaped network of the fused residual network for forward propagation. The average loss value of each batch of samples is calculated, and the network parameters are updated through the backpropagation algorithm. The network performance metrics are calculated using the validation set, and the network parameters are updated based on these metrics until the iteration termination condition in the training parameters is met, thus obtaining the pre-trained model.

6. The method for extracting flooded areas from SAR images based on multi-source feature fusion according to claim 5, characterized in that, The method of adjusting the pre-trained model based on transfer learning and combining it with the target scene corresponding to the target SAR image to obtain a flood area identification model includes: Acquire historical SAR images of the target scene corresponding to the target SAR image, preprocess the historical SAR images, and construct a fine-tuning training set; Adjust the training parameters, freeze the low-level encoder weights in the pre-trained model, and train the pre-trained model by fine-tuning the training set to obtain a flood area identification model.

7. The method for extracting flooded areas from SAR images based on multi-source feature fusion according to claim 1, characterized in that, The flood-area identification model outputs a flood-area mask map corresponding to the target SAR image, including: Preprocess the target SAR image and extract the corresponding polarization physical features and temporal anomaly features; The extracted polarization physical features and temporal anomaly features are input into the flood area identification model, and the corresponding flood attribution probability map is output. The flood attribution probability map is binarized to obtain the corresponding flood area mask map.

8. A SAR image flood area extraction system based on multi-source feature fusion, used to execute the SAR image flood area extraction method based on multi-source feature fusion as described in any one of claims 1 to 7, characterized in that, include: The image processing module is used to preprocess SAR images and extract polarization physical features and temporal anomaly features based on the preprocessed SAR images, and construct flood learning samples by combining them with preset label data. The model training module is used to train and learn the U-shaped network of the fused residual network through flood learning samples, obtain a pre-trained model, and adjust the pre-trained model based on transfer learning and the target scene corresponding to the target SAR image to obtain a flood area identification model. The flood area extraction module is used to output a flood area mask map corresponding to the target SAR image based on the flood area identification model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the SAR image flood area extraction method based on multi-source feature fusion as described in any one of claims 1 to 7.

Images