On-satellite identification method for water body in electric power disaster area

Abstract

The invention belongs to the technical field of satellite image recognition, and particularly discloses an on-satellite recognition method for a water body in an electric power affected area, and the method comprises the steps: carrying out the preprocessing of a multispectral image of a third satellite, and obtaining a processed multispectral image; pixel-level labeling is carried out on the processed multispectral image, and a special training sample set for satellite-borne water body recognition is constructed; performing iterative training on the special training sample set to construct a satellite-borne water body recognition model; performing quantitative compression and operator optimization on the satellite-borne water body recognition model to obtain an embedded reasoning model; when a target disaster area is monitored, acquiring a real-time multispectral image, and based on an embedded reasoning model, identifying a water body area and generating a suspected water body identification graph; and carrying out binarization compression and vector contour extraction on the suspected water body identification graph to generate a lightweight vector data packet, and downloading the lightweight vector data packet to a ground receiving center. According to the invention, the power disaster emergency efficiency and identification precision can be improved, and the data transmission pressure can be reduced.

Description

Technical Field

[0001] This invention belongs to the field of satellite image recognition technology, and in particular relates to a satellite-based method for identifying water bodies in power-affected areas. Background Technology

[0002] The Beijing-3 satellite (Beijing-3 satellite) is an optical remote sensing satellite with high-resolution, multispectral imaging capabilities. It can quickly acquire large-area multispectral images of the Earth's surface and has the ability to maneuver and revisit in emergencies. It is suitable for scenarios such as disaster monitoring, resource surveys, and power facility inspection. The multispectral sensor onboard the Beijing-3 satellite can simultaneously acquire data in the blue, green, red, and near-infrared bands, clearly reflecting the spectral differences of features such as water bodies, vegetation, buildings, and exposed surfaces, providing high-quality raw data support for water body identification in disaster-stricken areas.

[0003] Currently, the identification of water bodies in power-affected areas using multispectral imagery from the Beidou-3 satellite generally adopts a traditional model of on-board acquisition and storage followed by centralized ground processing: after the satellite captures the images, all raw images are first downloaded, and then ground stations perform radiometric correction, water index calculation, threshold segmentation, or manual interpretation to ultimately extract the extent of water bodies in the affected area. This satellite-ground separation processing model lacks real-time intelligent interpretation capabilities in orbit, making it difficult to obtain water body identification results immediately after the satellite passes overhead. Due to the massive amount of raw satellite imagery data, downloading all of it puts significant pressure on the satellite-ground data transmission channels. When communication facilities in power-affected areas are damaged, downloading large volumes of data is difficult and prone to errors. This prevents the power emergency command center from obtaining timely information on the distribution of water bodies in the affected area during the critical rescue window, severely impacting the timeliness of repair route planning and disaster assessment.

[0004] Current research has applied deep learning to remote sensing image recognition to improve automation and recognition accuracy. However, existing deep learning-based remote sensing image recognition methods typically train models using general remote sensing image datasets, lacking dedicated sample construction and model optimization processes tailored to the imaging characteristics of specific satellite models. When processing multispectral images acquired by specific satellites, these general models struggle to adapt to the unique spectral response and radiometric characteristics of satellite sensors, leading to decreased accuracy in water body identification. Furthermore, existing deep learning models are usually designed to run on cloud servers or high-performance workstations, lacking quantization compression and operator optimization for onboard embedded processing platform hardware resources, making efficient inference difficult in resource-constrained onboard environments. Summary of the Invention

[0005] This invention provides a satellite-based method for identifying water bodies in power-affected areas, which enables real-time satellite-based water body identification and minimized result downlink transmission, thereby improving the efficiency and accuracy of emergency response to power-affected areas and reducing data transmission pressure.

[0006] To achieve the above objectives, the present invention provides a satellite-based identification method for water bodies in power-affected areas, comprising: S1. Acquire multispectral images of the selected area from the No. 3 satellite and perform image preprocessing to obtain the processed multispectral images of the No. 3 satellite. S2. Pixel-level annotation of the real water body areas in the processed multispectral image is performed to obtain the label data of the processed multispectral image, and the label data is used to construct a special training sample set for water body identification on satellite No. 3. S3. Input the dedicated training sample set for spaceborne water body identification into the preset deep learning semantic segmentation network for iterative training to build the spaceborne water body identification model of the No. 3 satellite. S4. Quantize and compress the spaceborne water body identification model, and optimize the quantization and compression results to obtain the embedded inference model of the No. 3 satellite. S5. When the third satellite detects the target disaster area, it acquires real-time multispectral images of the target disaster area and performs forward reasoning on each frame of the real-time multispectral images based on the embedded reasoning model to identify water areas in each frame of the images and generate a suspected water body identification map of the target disaster area. S6. Perform binarization compression and vector contour extraction on the suspected water body identification map to generate a lightweight vector data package of the target disaster area, and then download the lightweight vector data package to the ground receiving center of the No. 3 satellite.

[0007] Preferably, in step S1, obtaining the processed multispectral image of satellite 3 includes: Radiometric calibration was performed on the multispectral images of the No. 3 satellite to obtain the radiance images of the No. 3 satellite; Atmospheric correction was performed on the radiance image to obtain the surface reflectance image from the No. 3 satellite; Geometric correction was performed on the surface reflectance image to obtain the processed multispectral image from the No. 3 satellite.

[0008] Preferably, in step S2, the processed multispectral image's label data is obtained, and the label data is used to construct a dedicated training sample set for satellite-borne water body identification on satellite-3, including: Multispectral images containing water areas were selected from historical multispectral images of the No. 3 satellite as images to be labeled. Radiometric normalization was performed on the image to be labeled to obtain the radiometric normalized image of the No. 3 satellite; Water bodies were labeled on the radiation-normalized image to obtain the initial label map of the radiation-normalized image; Morphological operations are performed on the initial label image to obtain the optimized label image; The optimized label map is paired with the radiometrically normalized image, and the pairing results are divided into training subset and validation subset according to a preset ratio; The training subset and the validation subset were used to construct a dedicated training sample set for onboard water body identification on the No. 3 satellite.

[0009] Preferably, in step S3, constructing the onboard water body identification model for satellite No. 3 includes: Input a training subset from the dedicated training sample set for spaceborne water body identification into a pre-defined deep learning semantic segmentation network; The deep learning semantic segmentation network is forward propagated so that the training subset passes through the convolutional layer, pooling layer and upsampling layer of the deep learning semantic segmentation network in sequence to generate a predicted segmentation map. The predicted segmentation map is compared with the true label map in the training subset to obtain the loss value of the training subset; The loss value is backpropagated in the deep learning semantic segmentation network to obtain the gradient information of the weight parameters of the loss value. Based on the gradient information of the weight parameters, the weight parameters in the deep learning semantic segmentation network are iteratively updated to obtain the iteratively updated deep learning semantic segmentation network. Based on the validation subset, the segmentation accuracy of the iteratively updated deep learning semantic segmentation network is validated. If the segmentation accuracy no longer improves, the training of the deep learning semantic segmentation network is stopped, and the preliminary training model of the No. 3 satellite is obtained. The weight structure pruning of the initial trained model is performed to obtain the pruned sparse model. Based on the training samples of the training subset, the pruned sparse model is fine-tuned to obtain the onboard water body identification model of the No. 3 satellite.

[0010] Preferably, in step S4, the embedded inference model of satellite three is obtained, including: Analyze the network structure of the spaceborne water body identification model to obtain the model structure information of the spaceborne water body identification model; Based on the model structure information and the integer operation instruction set of the onboard embedded processing platform in the No. 3 satellite, the onboard water body identification model is quantified. The quantized spaceborne water body identification model is calibrated using representative samples from the target disaster-stricken area to obtain the calibrated quantized model. The standard convolution operator in the calibrated quantization model is replaced with the depth-separable convolution operator of the spaceborne embedded processing platform to obtain the operator-optimized model. The optimized model was converted into a binary instruction sequence, and an in-loop test was performed on the binary instruction sequence to obtain the embedded inference model of the No. 3 satellite.

[0011] Preferably, in S5, the disaster-stricken area monitored by the third satellite is specifically: Satellite 3 receives emergency observation instructions sent by the ground emergency command center. The emergency observation instructions include the geographical coordinate range of the target disaster area. Based on the geographical coordinate range and the orbital parameters of the No. 3 satellite, an imaging plan is generated, which includes the side-swing angle of the No. 3 satellite and the imaging time. At the imaging moment, the control satellite No. 3 takes pictures of the target disaster area according to the side tilt angle to obtain real-time multispectral images of the target disaster area.

[0012] Preferably, in step S5, generating a suspected water body identification map of the target disaster area includes: When the third satellite detects the target disaster area, it continuously takes pictures of the target disaster area to obtain multiple frames of real-time multispectral images of the target disaster area, and integrates the multiple frames of real-time multispectral images into an image sequence. Rapid radiometric and geometric corrections are performed on real-time multispectral images in an image sequence to obtain corrected real-time image frames. The corrected real-time image frames are sequentially input into the embedded inference model, which outputs the water body probability value of the corrected real-time image frame pixel by pixel and generates a probability map frame of the corrected real-time image frame. By stitching and fusing probabilistic image frames according to the shooting time sequence and geographical coordinates, a suspected water body identification map of the target disaster area is generated.

[0013] Preferably, in step S6, generating a lightweight vector data packet for the target disaster area includes: Threshold segmentation is performed on the suspected water body identification image. Pixels with pixel values ​​greater than the preset adaptive segmentation threshold are marked as water bodies, and pixels with values ​​less than or equal to the preset adaptive segmentation threshold are marked as background, generating a binarized image. Connectivity analysis is performed on the binarized image to obtain the connected components of the binarized image; Boundary tracking is performed on connected components to obtain the water body contour of the binarized image, and the water body contour is converted into a vector polygon; The vector polygon and its attribute information are encapsulated into a lightweight vector data package for the target disaster area.

[0014] Preferably, the lightweight vector data packet is downloaded to the ground receiving center of the third satellite, including: Convert the vertex coordinates of the vector polygons in the lightweight vector data package to latitude and longitude coordinates to obtain the geographic coordinate vector data of the lightweight vector data package; Integrate geographic coordinate vector data into a geographic information file of water bodies in the target disaster area; Based on the modulation channel of the No. 3 satellite, the geographic information files of the water body area are downloaded to the ground receiving center.

[0015] Preferably, the formula for calculating the preset adaptive segmentation threshold is:

[0016] Where T is the preset adaptive segmentation threshold, L is the number of quantization levels of the target disaster-stricken area, and i is the probability level index of the target disaster-stricken area. Let i be the frequency of the pixel with probability value i. This is the preset initial estimated threshold.

[0017] The present invention has the following beneficial effects: This invention constructs a dedicated training sample set and optimizes the model for a specific satellite model, significantly improving the accuracy and reliability of onboard water body identification. By performing radiometric normalization and pixel-level fine annotation on historical multispectral images from the Beijing-3 satellite, a dedicated training sample set highly adapted to the spectral characteristics of the satellite sensors was constructed. This allows the deep learning semantic segmentation network to fully learn the unique spectral and textural features of water bodies in the satellite imagery. Simultaneously, by performing weight-structured pruning, quantization compression, and operator optimization on the trained model, a lightweight inference model adapted to the onboard embedded processing platform is generated, effectively solving the problem of low efficiency of general-purpose models in the resource-constrained environment of spacecraft. This targeted sample construction and model optimization method significantly improves the accuracy of onboard water body identification, greatly reduces the false positive and false negative rates, and provides high-confidence disaster information for power emergency command.

[0018] This invention achieves real-time intelligent interpretation and lightweight downlinking of results from satellite, improving the timeliness of acquiring water information in power-affected areas and alleviating the pressure on satellite-to-ground data transmission. When the satellite detects the target disaster area, it immediately performs rapid radiometric and geometric correction on the captured real-time multispectral images in orbit, and then immediately calls the embedded inference model to perform pixel-by-pixel forward inference to generate a suspected water body identification map in real time, completely changing the traditional delayed processing mode of "on-board acquisition and storage, ground processing and interpretation". By performing adaptive threshold segmentation, connected component analysis, and vector contour extraction on the generated suspected water body identification map, the massive image data is transformed into a lightweight vector data package containing only the geographical coordinates of the water body boundaries, and then quickly downlinked to the ground receiving center through telemetry or data transmission channels. This combination of on-orbit processing and lightweight downlinking enables the power emergency command center to obtain water distribution information in the disaster area almost in real time after the satellite passes overhead, winning valuable golden rescue time for repair route planning, personnel and material scheduling, and disaster assessment. Attached Figure Description

[0019] Figure 1This is a schematic flowchart of the method of the present invention; Figure 2 This is a map showing the accuracy of water body identification in the disaster-stricken area using the method of the present invention. Figure 3 This is a diagram showing the missed detection rate of water bodies in disaster-stricken areas using the method of the present invention. Detailed Implementation

[0020] The embodiments of the present invention will be further described below with reference to the accompanying drawings: like Figure 1 As shown, a satellite-based identification method for water bodies in power-affected areas includes the following steps: S1. Acquire multispectral images of the selected area from the No. 3 satellite and perform image preprocessing to obtain the processed multispectral images of the No. 3 satellite. S2. Pixel-level annotation of the real water body areas in the processed multispectral image is performed to obtain the label data of the processed multispectral image, and the label data is used to construct a special training sample set for water body identification on satellite No. 3. S3. Input the dedicated training sample set for spaceborne water body identification into the preset deep learning semantic segmentation network for iterative training to build the spaceborne water body identification model of the No. 3 satellite. S4. Quantize and compress the spaceborne water body identification model, and optimize the quantization and compression results to obtain the embedded inference model of the No. 3 satellite. S5. When the third satellite detects the target disaster area, it acquires real-time multispectral images of the target disaster area and performs forward reasoning on each frame of the real-time multispectral images based on the embedded reasoning model to identify water areas in each frame of the images and generate a suspected water body identification map of the target disaster area. S6. Perform binarization compression and vector contour extraction on the suspected water body identification map to generate a lightweight vector data package of the target disaster area, and then download the lightweight vector data package to the ground receiving center of the No. 3 satellite.

[0021] The selected area refers to a pre-defined key area where routine power facility operation and maintenance and disaster prevention and control are required. Typical examples include transmission line corridors, areas surrounding substations, and riverbanks—areas vulnerable to flooding and other critical power infrastructure. In this embodiment, the selected area provides a data source for the offline training of the satellite-borne water body identification model: by acquiring historical multispectral images of the selected area taken by the No. 3 satellite, a dedicated training sample set adapted to the satellite's spectral and radiometric characteristics is constructed. This completes the customized training and optimization of the onboard embedded inference model, providing a model foundation for real-time water body identification in the target disaster-stricken area.

[0022] In S1, the processed multispectral imagery from satellite 3 is obtained, including: Radiometric calibration was performed on the multispectral images of the No. 3 satellite to obtain the radiance images of the No. 3 satellite; Atmospheric correction was performed on the radiance image to obtain the surface reflectance image from the No. 3 satellite; Geometric correction was performed on the surface reflectance image to obtain the processed multispectral image from the No. 3 satellite.

[0023] The system acquires multispectral images of the target disaster area taken by the Beijing-3 satellite, which contain digital quantization values ​​for multiple bands.

[0024] The acquired multispectral images are radiometrically calibrated. Specifically, the radiometric calibration coefficients provided by the satellite sensor's laboratory calibration and on-board calibration system before launch are used to convert the raw digital quantization value of each pixel in the image into a physically meaningful radiance value band by band, thereby generating the radiance image of the Beijing-3 satellite.

[0025] Atmospheric correction processing is performed on the generated radiance image. Specifically, based on the solar angle, observation geometry, and synchronously acquired atmospheric parameters at the time of image acquisition, the scattering and absorption effects of atmospheric molecules and aerosols are calculated pixel by pixel using a radiative transfer model. The atmospheric path radiation contribution is subtracted from the radiance image, and the radiance is converted into surface reflectance, thereby generating a surface reflectance image from the Beijing-3 satellite.

[0026] The obtained surface reflectance images are subjected to geometric fine correction processing. Specifically, the mapping relationship between image coordinates and geographic coordinates is established by using satellite precise orbital attitude data, digital elevation model and ground control point library. Each pixel of the original image is repositioned to the precise geographic grid by resampling method to eliminate geometric distortion caused by satellite platform jitter, terrain undulation and earth curvature, thereby generating a processed multispectral image with accurate geographic positioning information.

[0027] Radiometric calibration converts raw digital quantization values ​​into physically meaningful radiance values, eliminating the impact of sensor response differences on image quality. Atmospheric correction accurately removes the scattering and absorption contributions of atmospheric molecules and aerosols from the radiance image and converts it into a surface reflectance image, eliminating inconsistencies in radiance across different time phases and atmospheric conditions. Geometric precision correction uses satellite orbital attitude data and a digital elevation model to accurately register the surface reflectance image onto a geographic grid, eliminating geometric distortions caused by satellite platform jitter and terrain undulations. Finally, processed multispectral images with consistent radiometric characteristics and accurate geolocation are generated, providing a high-quality data foundation for the subsequent training and real-time inference of the spaceborne water body identification model, thus ensuring the accuracy and reliability of water body identification.

[0028] In S2, the labeled data of the processed multispectral image is obtained, and the labeled data is used to construct a dedicated training sample set for onboard water body identification on the No. 3 satellite, including: Multispectral images containing water areas were selected from historical multispectral images of the No. 3 satellite as images to be labeled. Radiometric normalization was performed on the image to be labeled to obtain the radiometric normalized image of the No. 3 satellite; Water bodies were labeled on the radiation-normalized image to obtain the initial label map of the radiation-normalized image; Morphological operations are performed on the initial label image to obtain the optimized label image; The optimized label map is paired with the radiometrically normalized image, and the pairing results are divided into training subset and validation subset according to a preset ratio; The training subset and the validation subset were used to construct a dedicated training sample set for onboard water body identification on the No. 3 satellite.

[0029] Multispectral images containing water bodies such as rivers, lakes, reservoirs, and flood-inundated areas were selected from the historical archived multispectral image database of the Beijing-3 satellite as images to be labeled.

[0030] The selected images to be labeled are subjected to radiometric normalization. Specifically, stable ground features such as bare land and deep water areas in the images are used as references. By statistically analyzing the mean and standard deviation of each band, a linear transformation method is used to adjust the images to be labeled at different time phases to a uniform radiometric scale. This eliminates the radiometric inconsistencies caused by differences in imaging time, solar altitude angle, and atmospheric conditions, thereby generating radiometric normalized images of the Beijing-3 satellite with consistent radiometric characteristics.

[0031] On the generated radiometrically normalized image, a combination of manual interactive interpretation and semi-automatic segmentation is used to perform pixel-level water body labeling. Professional interpreters accurately delineate the boundaries of each water body region on the image based on the spectral and morphological characteristics of the water body. Pixels belonging to the water body are labeled as foreground and assigned a value of one, while pixels not belonging to the water body are labeled as background and assigned a value of zero, thereby generating an initial label map corresponding to the radiometrically normalized image.

[0032] The generated initial label image is processed by morphological operations. Specifically, morphological opening operation (erosion followed by dilation) is first used to remove isolated noise pixels and small mislabeled areas in the initial label image. Then, morphological closing operation (dilation followed by erosion) is used to fill in the small holes in the water area caused by missing labels, making the boundary of the water area smoother and more continuous, thus generating the optimized label image.

[0033] The optimized label images are paired one-to-one with the corresponding radiometrically normalized images, so that each pair of data contains an optical image and a precisely corresponding label image. All paired data are then randomly divided into a training subset for model training and a validation subset for model validation according to a preset ratio. The divided training and validation subsets are then combined to construct a dedicated training sample set specifically for water body identification on the Beijing-3 satellite.

[0034] Radiometric normalization eliminates radiometric differences between images from different time phases, generating radiometrically normalized images that ensure spectral consistency in subsequent annotation data. Pixel-level fine annotation is then performed on the radiometrically normalized images to generate initial label maps, assigning each pixel realistic water body attribute information. Morphological opening and closing operations are used to optimize the initial label maps, removing isolated noise points and filling small holes to generate optimized label maps with smooth and continuous boundaries, significantly improving the quality and accuracy of the annotation data.

[0035] The optimized label map was paired with radiometrically normalized images and divided into training and validation subsets. This resulted in a dedicated training sample set for spaceborne water body identification, specifically designed for the imaging characteristics of the Beijing-3 satellite. This provided high-quality sample data that was highly adapted to the spectral characteristics of the satellite sensor for the subsequent training of the deep learning semantic segmentation network. This ensured that the spaceborne water body identification model could fully learn the unique spectral features of water bodies in the satellite images, laying a solid data foundation for high-precision water body identification in orbit.

[0036] In S3, a satellite-borne water body identification model for satellite 3 is constructed, including: Input a training subset from the dedicated training sample set for spaceborne water body identification into a pre-defined deep learning semantic segmentation network; The deep learning semantic segmentation network is forward propagated so that the training subset passes through the convolutional layer, pooling layer and upsampling layer of the deep learning semantic segmentation network in sequence to generate a predicted segmentation map. The predicted segmentation map is compared with the true label map in the training subset to obtain the loss value of the training subset; The loss value is backpropagated in the deep learning semantic segmentation network to obtain the gradient information of the weight parameters of the loss value. Based on the gradient information of the weight parameters, the weight parameters in the deep learning semantic segmentation network are iteratively updated to obtain the iteratively updated deep learning semantic segmentation network. Based on the validation subset, the segmentation accuracy of the iteratively updated deep learning semantic segmentation network is validated. If the segmentation accuracy no longer improves, the training of the deep learning semantic segmentation network is stopped, and the preliminary training model of the No. 3 satellite is obtained. The weight structure pruning of the initial trained model is performed to obtain the pruned sparse model. Based on the training samples of the training subset, the pruned sparse model is fine-tuned to obtain the onboard water body identification model of the No. 3 satellite.

[0037] A training subset from the dedicated training sample set for spaceborne water body identification is loaded and input into a pre-built deep learning semantic segmentation network. This network adopts an encoder-decoder structure. The encoder part contains multiple convolutional and pooling layers for extracting image features, and the decoder part contains upsampling layers for restoring image resolution.

[0038] The forward propagation computation is performed on the deep learning semantic segmentation network, so that the multispectral image data in the training subset are sequentially passed through the convolutional layer of the network encoder for feature extraction, then through the pooling layer to reduce the feature map size and retain the main feature information, and then through the upsampling layer of the decoder to gradually restore the feature map size to the original input size. Finally, the probability value of belonging to water body is output at each pixel position, generating a predicted segmentation map corresponding to the input image.

[0039] The generated predicted segmentation map is compared pixel-by-pixel with the pre-labeled ground truth map in the training subset. The consistency between the predicted result and the ground truth label at each pixel position is calculated, and the difference between the two is aggregated into a loss value that characterizes the current prediction accuracy of the network. The calculated loss value is then backpropagated in the deep learning semantic segmentation network. That is, starting from the output layer of the network, the partial derivatives of the loss function with respect to the weight parameters of each layer are calculated layer by layer according to the chain rule to obtain the direction and magnitude information of the weight parameters that need to be adjusted, i.e., the gradient information of the weight parameters.

[0040] Based on the gradient information of the weight parameters obtained through backpropagation, the gradient descent optimization method is used to update the weight parameters of all convolutional layers in the deep learning semantic segmentation network. This makes the predicted segmentation map generated by the updated network in the next round of forward propagation closer to the real label map, thus obtaining the deep learning semantic segmentation network after this iteration. After each training round, the multispectral images in the validation subset are input into the iteratively updated deep learning semantic segmentation network for forward propagation to generate the corresponding validation predicted segmentation map. The validation predicted segmentation map is compared with the real label map in the validation subset to calculate the validation accuracy. When the validation set accuracy no longer improves after several consecutive training rounds, it indicates that the network has fully converged. Training is then stopped, and the current network weight parameters are saved, resulting in the preliminary training model of the Beijing-3 satellite.

[0041] The initial trained model is subjected to weight structure pruning. Specifically, the weight parameter distribution of each convolutional kernel in the model is analyzed, the importance score of each convolutional kernel to the model output is calculated, and convolutional kernels with importance scores below a preset threshold and their connection weights are removed as a whole, while convolutional kernels with greater contributions are retained. This significantly reduces the number of model parameters and computational cost while maintaining model accuracy, resulting in a pruned sparse model.

[0042] The pruned sparse model is fine-tuned based on the training samples in the training subset. This involves continuing forward and backward propagation of the pruned sparse model with a small learning rate in a few training rounds to restore the accuracy temporarily lost due to pruning. At the same time, the remaining weight parameters are further optimized, ultimately resulting in a spaceborne water body recognition model that is adapted to the embedded processing platform of the Beijing-3 satellite and has high-precision water body recognition capabilities.

[0043] The deep learning semantic segmentation network was iteratively trained using forward and backward propagation. Predicted segmentation maps were generated through layer-by-layer feature extraction and compared with the real label maps to calculate the loss value. Backpropagation was then used to update the network weight parameters, enabling the network to fully learn the spectral and texture features of water bodies in multispectral imagery from the Beijing-3 satellite. The training process was monitored using a validation subset to ensure network convergence and obtain a preliminary trained model. The preliminary trained model underwent weight pruning to remove convolutional kernels with smaller contributions, significantly reducing model complexity and computational cost. Fine-tuning was then used to recover the accuracy loss caused by pruning. The final spaceborne water body recognition model possesses both high recognition accuracy and a lightweight structure, efficiently adapting to the limited hardware resources of the spaceborne embedded processing platform. This provides a high-performance and efficient model foundation for subsequent real-time on-orbit water body recognition.

[0044] In S4, the embedded inference model of satellite No. 3 is obtained, including: Analyze the network structure of the spaceborne water body identification model to obtain the model structure information of the spaceborne water body identification model; Based on the model structure information and the integer operation instruction set of the onboard embedded processing platform in the No. 3 satellite, the onboard water body identification model is quantified. The quantized spaceborne water body identification model is calibrated using representative samples from the target disaster-stricken area to obtain the calibrated quantized model. The standard convolution operator in the calibrated quantization model is replaced with the depth-separable convolution operator of the spaceborne embedded processing platform to obtain the operator-optimized model. The optimized model was converted into a binary instruction sequence, and an in-loop test was performed on the binary instruction sequence to obtain the embedded inference model of the No. 3 satellite.

[0045] The network structure of the trained spaceborne water body identification model is analyzed. Specifically, the network topology of the model is analyzed layer by layer. The kernel size, number of input and output channels, stride and padding method of each convolutional layer are counted. The type and downsampling factor of the pooling layer and the number of nodes of the fully connected layer are recorded. At the same time, the numerical distribution range and sparsity of the weight parameters of each layer are analyzed to obtain the model structure information containing model architecture information and parameter distribution characteristics.

[0046] Based on the obtained model structure information and the integer arithmetic instruction set supported by the embedded processing platform of Beijing-3 satellite, the quantization scheme of the model is determined. Specifically, the numerical range of the weight parameters and activation values ​​of each layer in the model is mapped to the representation space of an eight-bit integer. The 32-bit floating-point quantization is converted into an 8-bit integer by calculating the scaling factor and zero offset, and the quantization parameters of each layer are recorded, thereby obtaining the quantized satellite water body identification model.

[0047] The quantized spaceborne water body identification model is calibrated using representative multispectral image samples from the target disaster area. Specifically, the representative samples are input into the quantized model for forward inference. The difference between the probability map output by the quantized model and the probability map output by the original floating-point model is compared. By minimizing the quantization error, the quantization scaling factor and zero offset are finely adjusted and integrated into the model to make the output of the quantized model as close as possible to the output of the original floating-point model, thus obtaining the calibrated quantized model.

[0048] The standard convolution operator in the calibrated quantization model is replaced with a depthwise separable convolution operator. Specifically, each standard convolutional layer is decomposed into two steps: depthwise convolution and pointwise convolution. Depthwise convolution performs spatial convolution operations independently on each input channel, while pointwise convolution uses a 1x1 convolution kernel to perform feature fusion in the channel dimension. This significantly reduces the amount of computation and parameters while maintaining the model's representational ability, making the model more suitable for the parallel computing architecture of the spaceborne embedded processing platform, resulting in an operator-optimized model.

[0049] The optimized model is converted into a binary instruction sequence adapted to the runtime environment of the spaceborne embedded processing platform. Specifically, a dedicated model compiler translates the computational graph structure and weight parameters of each layer of the model into machine instructions that the processing platform can directly execute, generating a fixed binary file. The generated binary instruction sequence is then loaded into the hardware-in-the-loop simulation environment of the spaceborne embedded processing platform. Test image data is input for forward inference, and the inference output is compared with the expected results to verify that its functions and performance meet the design requirements. Finally, an embedded inference model that can run stably and efficiently on the Beijing-3 satellite's spaceborne embedded processing platform is obtained.

[0050] The final embedded inference model features small storage size, fast computing speed, and low power consumption, which can perfectly adapt to the hardware resource limitations of the embedded processing platform on the Beijing-3 satellite. It provides an efficient and reliable inference engine for real-time water body identification on the satellite, ensuring that the satellite can quickly and accurately complete the water body identification task in the resource-constrained on-board environment.

[0051] In S5, the specific disaster-stricken areas monitored by satellite 3 are as follows: Satellite 3 receives emergency observation instructions sent by the ground emergency command center. The emergency observation instructions include the geographical coordinate range of the target disaster area. Based on the geographical coordinate range and the orbital parameters of the No. 3 satellite, an imaging plan is generated, which includes the side-swing angle of the No. 3 satellite and the imaging time. At the imaging moment, the control satellite No. 3 takes pictures of the target disaster area according to the side tilt angle to obtain real-time multispectral images of the target disaster area.

[0052] During its operation in orbit, the Beijing-3 satellite receives emergency observation commands sent in real time by the ground emergency command center through the uplink injection channel. These commands carry the geographical coordinate range of the target disaster area as determined by the ground emergency command center based on the power facility distribution map and disaster forecast information. The geographical coordinate range is described by a rectangular boundary formed by the latitude and longitude of the upper left corner and the lower right corner.

[0053] Imaging mission planning is carried out based on the received geographic coordinate range and the precise orbital parameters of the Beijing-3 satellite currently in orbit. Specifically, the orbital prediction model onboard the satellite is used to calculate the satellite's trajectory over the target disaster area in the future. Combined with the satellite's attitude maneuverability, the required lateral tilt angle to achieve coverage of the target area is calculated. At the same time, the optimal imaging time is calculated based on the relative geometric relationship between the satellite's position and the center of the target area, ensuring that the satellite is in a position that can completely cover the target area with the minimum lateral tilt angle at the imaging time. This generates an imaging plan that includes the lateral tilt angle and the imaging time.

[0054] When the imaging time determined by the imaging plan arrives, the satellite attitude and orbit control system controls the satellite platform to perform attitude maneuvers according to the side-swing angle in the imaging plan, so that the optical axis of the satellite imaging sensor is precisely pointed to the center of the target disaster area. At the same time, the imaging system is triggered to start shooting. The imaging sensor scans and images the ground objects according to the set integration time and gain parameters, and acquires multispectral image data covering the entire target disaster area, and finally obtains real-time multispectral images of the target disaster area.

[0055] By receiving emergency observation commands containing the geographic coordinates of the target disaster area, the satellite can accurately acquire the observation needs of the ground emergency command center. Based on the geographic coordinates and satellite orbital parameters, it generates an imaging plan that includes the side-swing angle and imaging time, ensuring that the satellite can fully cover the target area under optimal observation geometry. At the precisely calculated imaging time, the satellite is controlled to take pictures at the predetermined side-swing angle, realizing rapid response and accurate imaging of the power disaster area. This provides high-quality, full-coverage raw image data for subsequent real-time water body identification on the satellite, ensuring the timeliness and accuracy of power emergency monitoring missions.

[0056] In S5, a suspected water body identification map of the target disaster area is generated, including: When the third satellite detects the target disaster area, it continuously takes pictures of the target disaster area to obtain multiple frames of real-time multispectral images of the target disaster area, and integrates the multiple frames of real-time multispectral images into an image sequence. Rapid radiometric and geometric corrections are performed on real-time multispectral images in an image sequence to obtain corrected real-time image frames. The corrected real-time image frames are sequentially input into the embedded inference model, which outputs the water body probability value of the corrected real-time image frame pixel by pixel and generates a probability map frame of the corrected real-time image frame. By stitching and fusing probabilistic image frames according to the shooting time sequence and geographical coordinates, a suspected water body identification map of the target disaster area is generated.

[0057] Once the Beijing-3 satellite reaches its designated orbital position and receives emergency observation instructions from the ground, it determines that the current imaging target is a power disaster area that requires key monitoring. It immediately triggers the satellite imaging system to continuously push-broom the disaster area, acquiring multiple frames of real-time multispectral images covering the entire disaster area within the limited time window of the satellite's transit. These multiple frames of real-time multispectral images acquired in chronological order are then integrated into a complete image sequence according to the order of their capture time.

[0058] Rapid radiometric and geometric correction processing is performed on each frame of the integrated image sequence in real time. Specifically, the pre-stored radiometric correction coefficients are called on the spaceborne embedded processing platform to quickly convert the digital quantization value of each pixel in the image, eliminating the radiometric distortion introduced by changes in sensor response state. At the same time, a simplified geometric positioning model is established using the precise orbit and attitude data transmitted by the satellite in real time, and the image is quickly repositioned to eliminate the geometric offset caused by the slight jitter of the satellite platform and the rotation of the Earth, thereby generating the corrected real-time image frame for each frame.

[0059] The generated corrected real-time image frames are sequentially input into the embedded inference model deployed on the spaceborne embedded processing platform according to the shooting order. The embedded inference model performs forward inference calculations on each input corrected real-time image frame. Through the layer-by-layer operation of each convolutional layer, pooling layer and activation function in the model, the probability value of the pixel belonging to the water body is output at each pixel position. The probability values ​​of all pixels are organized into a probability distribution map of the same size as the input image frame, thereby generating the probability map frame corresponding to each corrected real-time image frame.

[0060] All generated probability image frames are stitched and fused according to the shooting time sequence and their corresponding geographic coordinate information. Specifically, the location of each image frame in the overall map of the target disaster area is determined based on its geographic location information. The probability values ​​in the overlapping areas of adjacent frames are weighted and averaged to eliminate stitching seams and ensure smooth transition, ultimately generating a complete and continuous suspected water body identification map covering the entire target disaster area.

[0061] The system continuously captures multiple frames of real-time multispectral images of the target disaster area and integrates them into an image sequence to ensure complete coverage of the entire disaster area without omissions. Each frame in the image sequence undergoes rapid radiometric and geometric correction to generate corrected real-time image frames, eliminating errors introduced by sensors and the platform to ensure the accuracy of the input data. These corrected real-time image frames are then sequentially input into an embedded inference model for pixel-by-pixel forward inference to generate probability map frames. Intelligent interpretation from image to water body probability is completed in real-time on the satellite. The probability map frames are then stitched and fused according to time sequence and geographic coordinates to generate a suspected water body identification map covering the entire area. This enables the immediate acquisition of complete water body distribution information in the disaster area during satellite transit, completely changing the traditional process of waiting for the satellite to transmit raw data before ground processing. This allows the power emergency command center to acquire a high-precision water body distribution map almost simultaneously with satellite transit, gaining valuable golden rescue time for repair decisions.

[0062] In S6, lightweight vector data packets are generated for the target disaster area, including: Threshold segmentation is performed on the suspected water body identification image. Pixels with pixel values ​​greater than the preset adaptive segmentation threshold are marked as water bodies, and pixels with values ​​less than or equal to the preset adaptive segmentation threshold are marked as background, generating a binarized image. Connectivity analysis is performed on the binarized image to obtain the connected components of the binarized image; Boundary tracking is performed on connected components to obtain the water body contour of the binarized image, and the water body contour is converted into a vector polygon; Encapsulate the vector polygon and its attribute information into a lightweight vector data package of the target disaster area; The lightweight vector data packets are transmitted to the ground receiving center of satellite No. 3, including: Convert the vertex coordinates of the vector polygons in the lightweight vector data package to latitude and longitude coordinates to obtain the geographic coordinate vector data of the lightweight vector data package; Integrate geographic coordinate vector data into a geographic information file of water bodies in the target disaster area; Based on the modulation channel of the No. 3 satellite, the geographic information files of the water body area are downloaded to the ground receiving center.

[0063] The generated suspected water body identification image is subjected to threshold segmentation processing. Specifically, the water body probability value of each pixel in the suspected water body identification image is compared with a preset adaptive segmentation threshold. Pixels with a value greater than the preset adaptive segmentation threshold are determined to belong to the water body region and are assigned a value of 1. Pixels with a value less than or equal to the preset adaptive segmentation threshold are determined to belong to the background region and are assigned a value of 0. In this way, the grayscale probability distribution map is transformed into a binary image with only two values: foreground and background.

[0064] The generated binarized image is subjected to connected component analysis. Specifically, the eight-neighbor labeling algorithm is used to traverse all pixels in the binarized image. Adjacent pixels with the same value of 1 are grouped into the same connected region. Each independent connected region is assigned a unique identifier number, and the geometric attributes such as the area, circumscribed rectangle, and centroid position of each connected region are counted. Thus, all connected components and their attribute information contained in the binarized image are obtained.

[0065] For each connected component obtained, boundary tracking is performed. Specifically, starting from the boundary point of each connected component, the outer contour pixels are tracked point by point in a counterclockwise or clockwise direction. The image coordinate position of each pixel point passed by the contour is recorded to form a closed contour line composed of a series of ordered coordinate points. These contour lines are simplified and compressed using the Douglas-Puk algorithm to remove redundant points and retain feature points. Finally, the water body contour of each connected component is transformed into a vector polygon composed of a vertex coordinate sequence.

[0066] The generated vector polygons, along with their corresponding area, perimeter, circumscribed rectangle, and other attribute information, are encapsulated and organized into a complete data packet containing geometric shapes and attribute data according to a preset data format. This generates a lightweight vector data packet for the target disaster area. The generated lightweight vector data packet is then forwarded to the data transmission channel via the Beijing-3 satellite's satellite service system. After modulation, it is transmitted directionally to the ground receiving center in the form of radio frequency signals, completing the downlink of the on-board processing results.

[0067] The coordinate transformation process is performed on the vertex coordinates of the vector polygons stored in the lightweight vector data package. Specifically, based on the rigorous geometric imaging model of the image and the precise orbital attitude data of the satellite, a precise transformation relationship between the image coordinate system and the geographic coordinate system is established. The row and column coordinates of each vertex on the image are converted into latitude and longitude coordinates in the WGS-84 geographic coordinate system, so that the vector polygons have accurate geospatial location information, thereby obtaining the geographic coordinate vector data of the lightweight vector data package.

[0068] The obtained geographic coordinate vector data is integrated and organized according to the standard geographic information data format. Specifically, all converted vector polygons and their attribute information are stored in a structured manner according to the common data exchange format to generate an independent file containing the precise geographic boundaries and attribute descriptions of the water body area, thereby obtaining the geographic information file of the water body area of ​​the target disaster area.

[0069] The Beijing-3 satellite's modulation channel is used to transmit geographic information files of water bodies. Specifically, the geographic information files of water bodies are channel-coded and modulated to convert them into radio frequency signals suitable for satellite-to-ground link transmission. These signals are then transmitted directionally to the ground receiving center through the satellite's data transmission antenna, ensuring that the ground receiving center can accurately receive and parse the complete geographic information of the water bodies in the disaster-stricken area.

[0070] By adaptively thresholding suspected water body identification images to generate binarized images, and through connected component analysis and boundary tracking, raster-style water body regions are transformed into vector polygons and encapsulated into lightweight vector data packets. This compresses the massive image data obtained from on-board processing into a minimal amount of data containing only the geometric information of water body boundaries, fundamentally resolving the contradiction between the limited bandwidth of the satellite-to-ground data transmission channel and the enormous amount of original image data. The vertex coordinates of the vector polygons are converted into latitude and longitude coordinates to generate geographic information files of water body regions with accurate geospatial information. This allows ground emergency command centers to receive the data and directly overlay it onto electronic maps without complex geometric corrections and georegistration, shortening the data processing chain from satellite imaging to emergency decision-making and providing power repair teams with accurate and reliable information on the distribution of water bodies in disaster-stricken areas in the first instance.

[0071] The formula for calculating the preset adaptive segmentation threshold is:

[0072] Where T is the preset adaptive segmentation threshold, L is the number of quantization levels of the target disaster-stricken area, and i is the probability level index of the target disaster-stricken area. Let i be the frequency of the pixel with probability value i. This is the preset initial estimated threshold.

[0073] T is a preset adaptive segmentation threshold that divides continuous probability values ​​in a suspected water body identification image into two categories: water body and background. If a pixel's probability value is greater than T, it is identified as water body; if it is less than or equal to T, it is identified as background.

[0074] L represents the number of quantization levels for the target disaster area, referring to the total number of levels after the probability values ​​of water bodies in the suspected water body identification map are quantized. For example, if the probability values ​​are originally continuously distributed between 0 and 1, they are discretized into L levels, each level corresponding to a probability interval. Typically, L can be 256, corresponding to 8-bit quantized grayscale levels.

[0075] i represents the probability level index of the target disaster area, which refers to the number of the quantification level. Each index corresponds to a specific probability value or probability range.

[0076] Let be the pixel frequency with probability value i, representing the probability level index equal to in the suspected water body identification map. The number of pixels. By statistically analyzing the entire suspected water body identification image, a frequency distribution histogram of i can be obtained. It refers to the height corresponding to the i-th bar in this histogram.

[0077] The preset initial estimation threshold refers to the preliminary estimated boundary between the water body and the background before any statistical information is available. Pixels that are close to this value are given higher weight in subsequent calculations.

[0078] The weighting function refers to calculating the corresponding weight value for each probability level index i, where the weight value increases as the index approaches the initial estimated threshold. The higher the probability level, the greater its weight; for those far from... The lower the probability level, the smaller its weight. Pixels near the segmentation threshold are more prone to misjudgment, so they are given higher weights to influence the final threshold determination.

[0079] The adaptive threshold calculation formula utilizes histogram statistics combined with a weighting function based on prior knowledge to dynamically calculate a segmentation threshold that reflects the overall probability distribution characteristics of the current image through a weighted average. The introduction of the weighting function makes the final threshold more robust, considering both the global pixel distribution and prior knowledge. This avoids segmentation failures that may occur when relying solely on global averaging or fixed thresholds under different lighting and water quality conditions, thereby improving the accuracy and robustness of water body identification.

[0080] Figure 2 This is a map showing the accuracy of water body identification in the disaster area using the on-board identification method in this embodiment. Figure 2The horizontal axis represents the image signal-to-noise ratio (SNR) in dB; the vertical axis on the left represents the recognition accuracy (IOU), which is dimensionless (range 0-1) and represents the intersection-union ratio between the water area recognized by the spaceborne water recognition model and the real water area; the vertical axis on the right represents the inference time in ms, which represents the computation time required for the embedded inference model to process a single frame of image. Figure 2 The blue curve shows the trend of recognition accuracy as a function of image signal-to-noise ratio (SNR), while the purple-red curve shows the trend of inference time as a function of SNR. From... Figure 2 As can be seen, as the image signal-to-noise ratio increases from 5dB to 30dB, the water body recognition accuracy gradually increases from approximately 0.65 and stabilizes at around 0.9; the inference time remains within the range of 18ms to 21ms, with only slight fluctuations and no obvious upward or downward trend. This trend indicates that the method in this embodiment can maintain stable inference efficiency under different image quality conditions, while the recognition accuracy steadily improves with the improvement of image quality, fully verifying the real-time performance and robustness of the spaceborne water body recognition model. Furthermore, the embedded inference model after quantization compression and operator optimization has good computational efficiency stability and is unaffected by changes in input image quality. Figure 2 The recognition accuracy and real-time performance of the method in this embodiment were verified when running on a spaceborne embedded processing platform, providing reliable data support for subsequent power emergency applications.

[0081] Figure 3 This is a map showing the false negative rate of water body identification in disaster-stricken areas using the on-board identification method in this embodiment. Figure 3 The horizontal axis represents the image signal-to-noise ratio (SNR) in dB, and the vertical axis represents the false negative rate in percentage, indicating the proportion of actual water bodies that were not identified by the spaceborne water body recognition model. Figure 3 The orange curve illustrates the trend of the false negative rate as a function of the signal-to-noise ratio. From Figure 3 As can be seen, when the image signal-to-noise ratio (SNR) is 5dB, the false negative rate is approximately 14.7%; as the SNR increases to 15dB, the false negative rate rapidly decreases to approximately 10%; when the SNR reaches 25dB or higher, the false negative rate stabilizes in the low range of 2.5% to 5%. This trend indicates that the method in this embodiment can still control the false negative rate below 15% even under poor image quality conditions, and the false negative rate is below 5% under normal imaging conditions, fully verifying the adaptability and reliability of the spaceborne water body recognition model in complex imaging environments. A low false negative rate is crucial for power emergency repairs, preventing repair teams from mistakenly entering dangerous areas or missing critical disaster information due to overlooked water bodies, thus providing a solid guarantee for the safe operation and rapid recovery of the power grid.

[0082] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A satellite-based identification method for water bodies in power-affected areas, characterized by the following steps: include: S1. Acquire multispectral images of the selected area from the No. 3 satellite and perform image preprocessing to obtain the processed multispectral images of the No. 3 satellite. S2. Pixel-level annotation of the real water body areas in the processed multispectral image is performed to obtain the label data of the processed multispectral image, and the label data is used to construct a special training sample set for water body identification on satellite No.

3. S3. Input the dedicated training sample set for spaceborne water body identification into the preset deep learning semantic segmentation network for iterative training to build the spaceborne water body identification model of the No. 3 satellite. S4. Quantize and compress the spaceborne water body identification model, and optimize the quantization and compression results to obtain the embedded inference model of the No. 3 satellite. S5. When the third satellite detects the target disaster area, it acquires real-time multispectral images of the target disaster area and performs forward reasoning on each frame of the real-time multispectral images based on the embedded reasoning model to identify water areas in each frame of the images and generate a suspected water body identification map of the target disaster area. S6. Perform binarization compression and vector contour extraction on the suspected water body identification map to generate a lightweight vector data package of the target disaster area, and then download the lightweight vector data package to the ground receiving center of the No. 3 satellite.

2. The satellite-based identification method for water bodies in power-affected areas as described in claim 1, characterized in that, In S1, the processed multispectral image of satellite No. 3 is obtained, including: Radiometric calibration was performed on the multispectral images of the No. 3 satellite to obtain the radiance images of the No. 3 satellite; Atmospheric correction was performed on the radiance image to obtain the surface reflectance image from the No. 3 satellite; Geometric correction was performed on the surface reflectance image to obtain the processed multispectral image from the No. 3 satellite.

3. The satellite-based identification method for water bodies in power-affected areas as described in claim 1, characterized in that, In step S2, the labeled data of the processed multispectral image is obtained, and the labeled data is used to construct a dedicated training sample set for water body identification on satellite-3, including: Multispectral images containing water areas were selected from historical multispectral images of the No. 3 satellite as images to be labeled. Radiometric normalization was performed on the image to be labeled to obtain the radiometric normalized image of the No. 3 satellite; Water bodies were labeled on the radiation-normalized image to obtain the initial label map of the radiation-normalized image; Morphological operations are performed on the initial label image to obtain the optimized label image; The optimized label map is paired with the radiometrically normalized image, and the pairing results are divided into training subset and validation subset according to a preset ratio; The training subset and the validation subset were used to construct a dedicated training sample set for onboard water body identification on the No. 3 satellite.

4. The satellite identification method for water bodies in power disaster areas as described in claim 3, characterized in that, In S3, the onboard water body identification model for satellite No. 3 is constructed, including: Input a training subset from the dedicated training sample set for spaceborne water body identification into a pre-defined deep learning semantic segmentation network; The deep learning semantic segmentation network is forward propagated so that the training subset passes through the convolutional layer, pooling layer and upsampling layer of the deep learning semantic segmentation network in sequence to generate a predicted segmentation map. The predicted segmentation map is compared with the true label map in the training subset to obtain the loss value of the training subset; The loss value is backpropagated in the deep learning semantic segmentation network to obtain the gradient information of the weight parameters of the loss value. Based on the gradient information of the weight parameters, the weight parameters in the deep learning semantic segmentation network are iteratively updated to obtain the iteratively updated deep learning semantic segmentation network. Based on the validation subset, the segmentation accuracy of the iteratively updated deep learning semantic segmentation network is validated. If the segmentation accuracy no longer improves, the training of the deep learning semantic segmentation network is stopped, and the preliminary training model of the No. 3 satellite is obtained. The weight structure pruning of the initial trained model is performed to obtain the pruned sparse model. Based on the training samples of the training subset, the pruned sparse model is fine-tuned to obtain the onboard water body identification model of the No. 3 satellite.

5. The satellite-based identification method for water bodies in power-affected areas as described in claim 1, characterized in that, In S4, the embedded inference model of satellite No. 3 is obtained, including: Analyze the network structure of the spaceborne water body identification model to obtain the model structure information of the spaceborne water body identification model; Based on the model structure information and the integer operation instruction set of the onboard embedded processing platform in the No. 3 satellite, the onboard water body identification model is quantified. The quantized spaceborne water body identification model is calibrated using representative samples from the target disaster-stricken area to obtain the calibrated quantized model. The standard convolution operator in the calibrated quantization model is replaced with the depth-separable convolution operator of the spaceborne embedded processing platform to obtain the operator-optimized model. The optimized model was converted into a binary instruction sequence, and an in-loop test was performed on the binary instruction sequence to obtain the embedded inference model of the No. 3 satellite.

6. The satellite identification method for water bodies in power-affected areas as described in claim 1, characterized in that, In S5, the disaster-stricken area monitored by satellite 3 is specifically as follows: Satellite 3 receives emergency observation instructions sent by the ground emergency command center. The emergency observation instructions include the geographical coordinate range of the target disaster area. Based on the geographical coordinate range and the orbital parameters of the No. 3 satellite, an imaging plan is generated, which includes the side-swing angle of the No. 3 satellite and the imaging time. At the imaging moment, the control satellite No. 3 takes pictures of the target disaster area according to the side tilt angle to obtain real-time multispectral images of the target disaster area.

7. The satellite-based identification method for water bodies in power-affected areas as described in claim 1, characterized in that, In step S5, generating a suspected water body identification map of the target disaster area includes: When the third satellite detects the target disaster area, it continuously takes pictures of the target disaster area to obtain multiple frames of real-time multispectral images of the target disaster area, and integrates the multiple frames of real-time multispectral images into an image sequence. Rapid radiometric and geometric corrections are performed on real-time multispectral images in an image sequence to obtain corrected real-time image frames. The corrected real-time image frames are sequentially input into the embedded inference model, which outputs the water body probability value of the corrected real-time image frame pixel by pixel and generates a probability map frame of the corrected real-time image frame. By stitching and fusing probabilistic image frames according to the shooting time sequence and geographical coordinates, a suspected water body identification map of the target disaster area is generated.

8. The satellite identification method for water bodies in power disaster areas as described in claim 1, characterized in that, In step S6, generating a lightweight vector data packet for the target disaster area includes: Threshold segmentation is performed on the suspected water body identification image. Pixels with pixel values ​​greater than the preset adaptive segmentation threshold are marked as water bodies, and pixels with values ​​less than or equal to the preset adaptive segmentation threshold are marked as background, generating a binarized image. Connectivity analysis is performed on the binarized image to obtain the connected components of the binarized image; Boundary tracking is performed on connected components to obtain the water body contour of the binarized image, and the water body contour is converted into a vector polygon; The vector polygon and its attribute information are encapsulated into a lightweight vector data package for the target disaster area.

9. The satellite identification method for water bodies in power disaster areas as described in claim 8, characterized in that, The lightweight vector data packets are transmitted to the ground receiving center of satellite No. 3, including: Convert the vertex coordinates of the vector polygons in the lightweight vector data package to latitude and longitude coordinates to obtain the geographic coordinate vector data of the lightweight vector data package; Integrate geographic coordinate vector data into a geographic information file of water bodies in the target disaster area; Based on the modulation channel of the No. 3 satellite, the geographic information files of the water body area are downloaded to the ground receiving center.

10. The satellite identification method for water bodies in power disaster areas as described in claim 8, characterized in that, The formula for calculating the preset adaptive segmentation threshold is: Where T is the preset adaptive segmentation threshold, L is the number of quantization levels of the target disaster-stricken area, and i is the probability level index of the target disaster-stricken area. Let i be the frequency of the pixel with probability value i. This is the preset initial estimated threshold.

Images