Water body automatic extraction method and device fusing multi-source remote sensing data

Abstract

The invention relates to a water body automatic extraction method and device fusing optical remote sensing data, synthetic aperture radar (SAR) data and unmanned aerial vehicle images. Preprocessing such as radiometric calibration, atmospheric correction and geometric correction is carried out on multi-source remote sensing data, an improved DeepLabv3 + model is used for processing an optical image, a UNet model is used for processing an SAR image, and accurate extraction of a water body area is achieved. And the system fuses optical and SAR data extraction results to generate a high-precision water body recognition result. The method is suitable for flood disaster monitoring and emergency response, and the recognition precision and timeliness of the flood inundation range are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of large-scale, high-precision automatic extraction and dynamic monitoring of water bodies, and in particular relates to a method and device for automatic extraction of water bodies that integrates multi-source remote sensing data. Background Art

[0002] As global climate change intensifies and extreme weather occurs more frequently, floods have become one of the most destructive and widespread natural disasters. Traditional flood monitoring methods, such as manual inspections and hydrological monitoring, can provide accurate water information, but they often face problems such as operational difficulties, high time costs, and limited coverage during disasters.

[0003] With the rapid development of remote sensing technology, optical satellite images, SAR data and drone images are gradually applied to flood disaster monitoring. Remote sensing satellites have the advantages of large range, rapid acquisition, continuous observation and low cost, and can be used to dynamically monitor the flood inundation range. However, there are limitations to a single data source. Although optical remote sensing satellites (such as Landsat, SPOT, etc.) have high spatial resolution and rich spectral information, they are easily affected by clouds and weather conditions, especially when observed under cloud cover or at night, a large amount of flood information may be lost. Synthetic aperture radar (SAR) can overcome these weather obstacles and has all-weather and all-day observation capabilities. It can penetrate clouds and obtain surface images. SAR has been widely used in flood mapping, wetland monitoring and water research. Commonly used satellites include the European Space Agency's Sentinel-1 radar satellite and my country's Gaofen-3 radar satellite. However, the number of existing SAR satellites is limited, making it difficult to achieve global daily coverage and timely and comprehensive monitoring of the maximum flood inundation range. With the advantages of high resolution and flexible maneuverability, drone aerial survey has become an important means of flood emergency monitoring. UAV technology combined with oblique photogrammetry and GIS analysis can efficiently extract flood inundation depth and range. By integrating multiple data sources such as ground observations, satellite remote sensing and drone aerial surveys, an integrated "sky-ground" monitoring system is formed, which helps to quickly provide accurate flood disaster information.

[0004] In order to solve the limitation of a single data source, the present invention proposes a method and device for automatic water extraction by integrating optical remote sensing satellite data, SAR data and drone images, so as to improve the monitoring accuracy and timeliness of flood inundation range. Summary of the invention

[0005] The present invention proposes a method and device for automatically extracting water bodies by fusing multi-source remote sensing data, aiming to solve the limitations of traditional flood disaster monitoring methods under severe meteorological conditions and improve the accuracy and timeliness of water body identification during flood disasters. By fusing optical remote sensing data, SAR data, and high-resolution data obtained by drone aerial survey, the present invention uses a deep learning model to perform comprehensive preprocessing and deep learning training on different types of data to achieve accurate extraction of flood inundation range. This method overcomes the limitations of a single remote sensing data source, significantly improves the accuracy of flood monitoring, and provides strong technical support for disaster assessment and emergency response.

[0006] Step 1: Multi-source data acquisition and preprocessing

[0007] S11, Optical satellite imagery and drone imagery data acquisition

[0008] The present invention first obtains optical satellite images, mainly from the Gaofen-2 (GF2) satellite independently developed by my country. The satellite is equipped with a 1-meter panchromatic and 4-meter multispectral camera, with sub-meter spatial resolution and high positioning accuracy, and can provide high-resolution observation data for flood monitoring. The optical images obtained by it include R, G, B, and NIR bands, which are used for subsequent water extraction tasks.

[0009] The acquisition of drone images is completed through drone aerial survey technology, ensuring that drone images and optical satellite images have consistent band combinations. The acquisition of these data ensures the integrity of surface information and is suitable for subsequent data fusion and deep learning model training.

[0010] S12, optical satellite imagery and drone imagery preprocessing

[0011] Before water extraction, a series of preprocessing steps are performed on optical remote sensing images and drone images to ensure image quality and accuracy:

[0012] Radiometric calibration: Converts the original digital value (DN value) into physical radiation brightness or reflectivity to ensure that the image data truly reflects the spectral characteristics of the object. Radiometric calibration is achieved by using sensor calibration coefficients.

[0013] Atmospheric correction: To eliminate the effects of aerosol, gas absorption and scattering in the atmosphere and restore the true reflectivity of the ground objects. Atmospheric correction is usually combined with radiation calibration to improve the accuracy of quantitative image analysis.

[0014] Orthorectification: Through geometric correction processing, the geometric distortion of the image caused by terrain undulation and sensor tilt is eliminated to ensure the consistency of the geographic coordinates of the image.

[0015] Image fusion: Fusion of multi-band or multi-sensor image data to generate high-quality images. Commonly used image fusion methods include principal component analysis (PCA), intensity-hue-saturation transform (IHS), and wavelet transform to ensure that the image contains rich spectral and spatial information.

[0016] Image cropping: crop images based on the region of interest (ROI) to reduce the amount of data processing and focus on processing specific areas. Cropping methods include cropping based on image space coordinates and user-specified ROI cropping.

[0017] S13, SAR image data acquisition

[0018] SAR data acquisition relies on SAR satellites such as Sentinel1 and Gaofen-3. SAR has the ability to penetrate clouds and can work in any weather conditions, and is particularly suitable for water monitoring under adverse weather conditions. The VV and VH dual-polarization data of SAR images provide rich surface backscatter information and can effectively reflect the characteristics of flooded areas.

[0019] S14, SAR image preprocessing

[0020] The preprocessing steps of SAR data include:

[0021] Orbit correction: Due to the low accuracy of satellite orbit status, the original image needs to be corrected using accurate orbit files to ensure the accurate correspondence between the SAR image and the geographic coordinate system.

[0022] Noise removal: SAR images usually contain thermal noise, and RefinedLee filter is used to denoise them, retaining edge details and eliminating coherent speckle noise.

[0023] Radiation calibration: Convert the received backscatter signal into a physical quantity with units, the backscatter coefficient σ o , ensuring the physical meaning of SAR data.

[0024]

[0025] In the formula, σ o is the backscatter coefficient of each pixel in db; A is the DN value on the original image; K is the absolute calibration factor of the product; θ is the incident angle.

[0026] Terrain correction: Processes geometric distortion caused by terrain, especially in mountainous areas or areas with complex terrain, to give images higher geographic accuracy.

[0027] Decibel processing: Convert the backscatter coefficient into dB value to enhance the water characteristics and ensure the convenience of data processing and visualization.

[0028] σ0(dB)=10·log 10 σ0

[0029] Where σ0 represents the radar backscatter coefficient, and σ0(dB) represents the backscatter coefficient after decibel processing.

[0030] S15, water index calculation and threshold selection

[0031] In optical images, the Normalized Difference Water Index (NDWI) is calculated to highlight water features. The calculation formula is:

[0032]

[0033] Among them, G is the reflectance value of the green band, NIR is the reflectance value of the near-infrared band, and the calculated NDWI value can effectively highlight the characteristics of the water body.

[0034] In SAR images, the water body index (SDWI) is calculated to highlight the difference between water and land using the following formula:

[0035] SDWI=ln(10*VV*VH)-8

[0036] According to the distribution characteristics of NDWI and SDWI values, geographic information software is used to select thresholds, divide water and non-water areas, and generate binary annotation data. Morphological operations such as corrosion and expansion are used to further optimize the annotated image, remove noise, and ensure annotation accuracy.

[0037] Step 2: Deep learning model training and application

[0038] S21, Water Extraction Model from Optical Remote Sensing and UAV Imagery

[0039] The DeepLabv3+ model is mainly used for water extraction tasks from optical images and drone images. In order to meet the needs of multi-source remote sensing data, this paper makes structural improvements to DeepLabv3+, enabling it to effectively process multi-channel data containing 4 bands (R, G, B, NIR) and improve the accuracy of the model in water extraction. The specific improvements and processing procedures are as follows:

[0040] 1. Improvements to the DeepLabv3+ model:

[0041] Input layer adjustment: The original DeepLabv3+ model is usually used to process three-channel RGB images. To adapt to the input of 4-channel optical images (R, G, B, NIR), the number of input channels of the first layer of convolution kernel of the model is adjusted from 3 to 4. In this way, the model can process multi-spectral information at the same time, especially the added near-infrared (NIR) band, which significantly improves the recognition ability of water bodies.

[0042] Encoder adjustment: The encoder of DeepLabv3+ uses dilated convolution to expand the receptive field, but for the high-resolution areas of optical images, the original sampling rate may cause excessive smoothing of features. Therefore, in this invention, we adjust the sampling rate of the dilated convolution in the encoder so that the model can better retain detail information while maintaining a large receptive field.

[0043] Decoder optimization: In order to solve the more complex problem of water edge areas, the upsampling part of the decoder introduced a more detailed edge recovery strategy, combined with multi-scale features to reduce the loss of water edge information and improve segmentation accuracy.

[0044] 2. Image processing steps:

[0045] Input preprocessing: Before processing optical satellite images or drone images, the images are first standardized and the pixel values ​​are normalized to the range of [0,1] to ensure that data from different bands can be processed uniformly by the model.

[0046] Model reasoning: The preprocessed multi-channel image is input into the improved DeepLabv3+ model. The encoder of the model first extracts high-level semantic features, and the decoder then upsamples these features and restores the image resolution.

[0047] Water mask generation: Through layer-by-layer extraction of the convolutional layer, the model finally outputs a binary mask image with the same resolution as the input image. In this mask, the model marks the water area as 1 and the non-water area as 0.

[0048] Post-processing: The generated water mask may contain some noise points, especially in areas with complex terrain or cloud interference. The isolated noise points are removed through subsequent morphological operations (such as dilation and erosion), and the edges of the water area are smoothed.

[0049] 3. Model training and optimization:

[0050] Data enhancement: Due to the high resolution of optical images and drone images, directly using the original data for training can easily lead to insufficient generalization of the model. Therefore, we performed a variety of data enhancement operations on the training data, including random rotation, cropping, horizontal flipping, and color jittering, to increase the diversity of samples.

[0051] Loss function: A binary cross entropy loss function is used to calculate the difference between the model prediction results and the true labels. The model parameters are adjusted through back propagation to gradually optimize the model weights.

[0052] Hyperparameter adjustment: During the training process, multiple experiments are performed on hyperparameters such as learning rate and batch size to ultimately determine the appropriate model configuration to achieve the best training results.

[0053] S22, SAR image water extraction model

[0054] The complexity of SAR image processing and the small sample size determine the selection of the UNet model as the main framework for water extraction. The UNet model has a simple structure, but performs well for small sample training, and is particularly suitable for processing images with complex textures and noise in segmentation tasks, such as water and ground features in SAR images.

[0055] 1. Reasons for choosing UNet:

[0056] Processing complex SAR data: SAR images are highly complex due to the influence of coherent speckle noise and geometric distortion. The UNet structure can effectively combine low-level detail features and high-level semantic information through its skip connection, ensuring that the edge and texture features of water bodies are preserved in a complex noise background.

[0057] Adapt to small sample training: Due to the limited amount of SAR image data, choosing a UNet model with a small number of parameters and a fast convergence speed is conducive to efficient training under small sample conditions.

[0058] 2. SAR image processing steps:

[0059] Preprocessing: First, the SAR images are subjected to orbit correction, denoising and decibel processing to ensure that the SAR data has stable geographic accuracy and backscatter signal strength. Then, the VV and VH polarization data are synthesized and stacked with the water body index (SDWI) to generate 3-channel input data.

[0060] Model input: The processed 3-channel SAR image is input into the UNet model. The UNet encoder gradually extracts the features of the SAR image through multi-layer convolution and pooling operations. The feature maps in the encoding stage are downsampled layer by layer to capture semantic information at different scales.

[0061] Skip connection: The shallower features in the encoder are directly transferred to the decoder through skip connection, so that high-level semantic information is combined with low-level detail information, which can better restore the edge features of water bodies.

[0062] Decoding and output: The decoder performs upsampling on the extracted feature map, gradually restores the original image resolution, and finally outputs a binary mask of the same size as the input image, with the water area as 1 and the non-water area as 0.

[0063] Post-processing: Since SAR images are often accompanied by noise, especially in the water boundary area, morphological operations such as median filtering and closing operations are applied to the predicted mask map to further eliminate noise and smooth the boundaries.

[0064] 3. Model training and optimization:

[0065] Loss function and optimizer: The binary cross entropy loss function is also used to measure the difference between the model prediction result and the true label. At the same time, the Adam optimizer is selected to optimize the model parameters and accelerate convergence through its adaptive learning rate mechanism.

[0066] Data enhancement: In order to overcome the problem of small sample size of SAR data, data enhancement is performed using random rotation, translation, cropping, and flipping to increase data diversity and improve the robustness of the model.

[0067] Hyperparameter adjustment: Adjust the learning rate, batch size, etc. of the model, and use the performance of the validation set to control the early stopping of training to avoid overfitting of the model.

[0068] S23, model fusion and water extraction process

[0069] In practical applications, the extraction results of optical images and SAR images are processed in stages, predicted by DeepLabv3+ and UNet respectively, and then the results are fused to obtain a more comprehensive water mask. The fusion steps are as follows:

[0070] 1. Model-independent reasoning:

[0071] The optical image is input into DeepLabv3+, and the SAR image is input into UNet, and two binary mask images are generated, representing the water body extraction results under the two data sources respectively.

[0072] 2. Result fusion:

[0073] The two mask images are superimposed to generate a fused mask. For overlapping areas, if both mask images are identified as water bodies, the final output is the water body area; if one mask is identified as a water body and the other is not, the fusion rules are set to obtain the best results by combining geographic information and expert judgment.

[0074] 3. Post-processing:

[0075] Post-processing operations such as closing operation and median filtering are applied on the final fused mask to further remove noise and ensure smooth and continuous water body boundaries. Finally, the water body mask data is exported as vector data for subsequent GIS analysis.

[0076] The present invention effectively handles the task of water extraction in multi-source remote sensing images through the improved DeepLabv3+ and UNet models, and significantly improves the accuracy and timeliness of water body identification during flood disasters by integrating the advantages of optical and SAR data.

[0077] Step 3: Water body extraction and result fusion

[0078] S31, water extraction from optical images

[0079] The trained DeepLabv3+ model is used to extract water from optical images and generate water masks.

[0080] S32, SAR image water extraction

[0081] Convert the VV, VH and SDWI band data to uint8 format, ensure that their value range is between 0 and 255, and stack the VV, VH and SDWI band data in uint8 format together to form three-channel image data. In this way, each pixel has three values, corresponding to VV, VH and SDWI respectively. The VV, VH and SDWI channels of the SAR image are processed through the trained UNet model to generate a water mask.

[0082] S33, high-precision water body identification and data fusion

[0083] In this step, the trained deep learning model is used to extract water bodies from multi-temporal remote sensing data before and after the flood disaster. The specific process is as follows:

[0084] The data of the input optical remote sensing image and SAR image after preprocessing will enter the water body extraction device for further processing.

[0085] The device makes a judgment based on the type of input data. If the input is a four-channel optical image (R, G, B, NIR), the DeepLabv3+ model is used for water extraction; if the input is a SAR image, the UNet model is used for processing.

[0086] During the data extraction process, the DeepLabv3+ model generates water masks for optical images by extracting multi-channel features and restoring edges of images; the UNet model processes the noise and complex terrain features in SAR images to generate corresponding water masks.

[0087] When optical images and SAR images exist at the same time, the device will fuse the water body extraction results processed separately. The fusion process uses the water body mask map superimposed by the two data sources, combined with the high-resolution spectral information provided by the optical image and the penetration ability of the SAR image to improve the extraction accuracy. Through multi-source data fusion technology, the limitations of a single data source under different meteorological conditions and observation times are effectively overcome.

[0088] The fused result will generate a comprehensive water body mask covering the flooded area. This method can not only realize the identification of water bodies in a large area, but also provide accurate information on the dynamic changes of floods, providing an accurate and reliable basis for emergency response and disaster assessment.

[0089] Step 4: Post-processing and accuracy evaluation

[0090] S41, post-processing

[0091] The generated water body mask is subjected to median filtering and closing operations to remove noise, fill holes, eliminate the jagged effect of water body boundaries, and improve the quality of water body identification results. The processed water body mask can be exported to vector data format (shp) for further geographic information analysis.

[0092] S42, Accuracy Evaluation

[0093] The water extraction accuracy of the model is evaluated by calculating indicators such as accuracy, mean intersection over union (MIoU) and recall. Accuracy measures the prediction accuracy of the model for all pixels, MIoU reflects the overlap between the model's predicted area and the actual area, and recall evaluates the model's ability to identify water areas. Through the comprehensive evaluation of these indicators, the generalization ability of the model and its effectiveness in practical applications are verified.

[0094] Accuracy indicates the ratio of the number of samples correctly predicted by the model to the total number of samples. It reflects the overall accuracy of the model. The formula is as follows.

[0095]

[0096] Among them, TP stands for True Positive, TN stands for True Negative, FP stands for False Positive, and FN stands for False Negative.

[0097] MIoU is a commonly used evaluation indicator in semantic segmentation tasks. Iou represents the degree of overlap between the predicted result and the true value. MIoU is the average value of Iou of all categories, reflecting the comprehensive performance of the model in all categories.

[0098]

[0099] Recall focuses on the coverage of true positive samples by the algorithm in the segmentation task. It represents the proportion of truly positive samples that are correctly identified and is an important indicator for evaluating whether the model misses important information.

[0100]

[0101] Through the above detailed steps, the present invention realizes the fusion processing of optical remote sensing data, SAR data and UAV images, provides a complete and effective solution for water body extraction in flood disasters, and significantly improves the accuracy and timeliness of water body identification.

[0102] Compared with the prior art, the main advantages of the present invention are:

[0103] 1. Multi-source data: The present invention makes comprehensive use of optical remote sensing data, SAR data and drone images, overcoming the limitations of a single data source and enabling accurate water monitoring under different meteorological conditions.

[0104] 2. High precision: With the help of deep learning models, water bodies can be extracted with high precision, especially under extreme weather conditions of flood disasters. The model of the present invention can realize accurate water body boundary identification through the fusion of multi-source data.

[0105] 3. Full-process processing: From data acquisition, preprocessing, deep learning training to result fusion and accuracy evaluation, the present invention provides a complete processing flow, which greatly improves the efficiency of emergency response and significantly enhances the scientificity and reliability of disaster assessment.

[0106] The advantage of this system is that it realizes automation and efficiency in water extraction, which not only improves the monitoring accuracy of flood inundation range, but also provides reliable technical support for disaster emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] In order to more clearly illustrate the specific implementation methods of the present application or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0108] Figure 1 A flow chart of a method for extracting water from multi-source remote sensing image data provided in an embodiment of the present application;

[0109] Figure 2 A schematic diagram of the structure of the water body segmentation model provided in the embodiment of the present application;

[0110] Figure 3 A functional structural diagram of a multi-source remote sensing image data water body extraction device provided in an embodiment of the present application;

[0111] Figure 4 This is a schematic diagram of the results of the main processing steps in the Hudong flood detention area of ​​the embodiment of the present application. DETAILED DESCRIPTION

[0112] The present invention is further described below in conjunction with specific embodiments.

[0113] Example 1: A method for automatically extracting water bodies by integrating multi-source remote sensing data

[0114] This embodiment introduces in detail a method for automatically extracting water bodies by fusing optical remote sensing data, SAR data, and drone image data, which is specifically described in four steps.

[0115] Step 1: Acquisition and preprocessing of multi-source remote sensing data

[0116] S11, Optical image data acquisition and preprocessing

[0117] Acquire multispectral images from optical remote sensing satellites (such as GF2), usually including four bands: R, G, B, and NIR.

[0118] Radiometric calibration: Convert the original digital value (DN value) into reflectance or radiation brightness value to ensure that the image can truly reflect the spectral characteristics of the object.

[0119] Atmospheric correction: remove the influence of aerosol and water vapor in the atmosphere and restore the true reflectivity of the ground.

[0120] Orthorectification: Eliminate geometric distortion caused by terrain undulations or sensor tilt to ensure that images accurately correspond to geographic coordinates.

[0121] Image cropping and fusion: crop large images according to the region of interest (ROI) and fuse different bands or images using methods such as principal component analysis (PCA) or intensity hue saturation (IHS).

[0122] S12, SAR image data acquisition and preprocessing

[0123] Obtain C-band dual-polarization data (VV, VH) from SAR satellites such as Sentinel1 or Gaofen-3.

[0124] Track Correction: Use accurate track files to correct the trajectory of SAR images to ensure their geometric accuracy.

[0125] Radiation calibration: converting the backscatter signal into physically meaningful backscatter coefficients.

[0126] Denoising and filtering: Use the RefinedLee filter to remove coherent speckle noise and retain the geometric structure and edge details of the objects.

[0127] Decibel conversion: Convert backscatter coefficients into decibel values ​​(dB) to highlight water characteristics and reduce data storage requirements.

[0128] S13, UAV image data acquisition and preprocessing

[0129] A drone was used to obtain high-resolution images of the study area, and the data included four bands: R, G, B, and NIR.

[0130] After radiation correction and orthorectification, the image data can be directly fused with satellite images to improve the spatial resolution of water body extraction.

[0131] Step 2: Deep learning model training

[0132] S21, optical image processing model (DeepLabv3+)

[0133] For optical images, this method uses an improved DeepLabv3+ model. The input layer of the model has been improved to process four-channel (R, G, B, NIR) data. Accurate water boundary segmentation is achieved by extracting multi-scale features from the high-resolution spectral features of the image.

[0134] Data augmentation: During the training process, the input data is randomly rotated, cropped, flipped, and so on to expand the training data set and improve the model's generalization ability.

[0135] Model optimization: Adjust the model's hyperparameters (such as learning rate, batch size, etc.) and use cross-validation to improve model performance to ensure that water areas in optical images are accurately identified.

[0136] S22, SAR image processing model (UNet)

[0137] The UNet model is used to process SAR images. The UNet model combines low-level and high-level features through skip connections, which can effectively segment water areas in complex backgrounds.

[0138] Sample expansion: Due to the small sample size of SAR data, data enhancement techniques (such as random cropping and rotation) are used to expand the sample size and improve the performance of the model in noisy and complex data.

[0139] Loss function: Use the binary cross entropy loss function and use optimization algorithms (such as the Adam optimizer) to improve the convergence speed and accuracy of the model.

[0140] Step 3: Water Extraction

[0141] S31, water extraction from optical images

[0142] The preprocessed optical image is input into the trained DeepLabv3+ model. The model identifies the water area in the optical image and generates a water mask.

[0143] S32, SAR image water extraction

[0144] Similarly, the preprocessed SAR image is input into the UNet model, and the model can segment the water area in the SAR image and generate a binary water mask.

[0145] Step 4: Result fusion and water body identification

[0146] S41, Data Fusion

[0147] The water masks of optical images and SAR images are superimposed and fused. The high-resolution spectral information provided by optical images can identify small water bodies, while SAR images have all-weather capabilities, which can make up for the shortcomings of optical images in bad weather.

[0148] S42, post-processing

[0149] The fused water body mask is closed and smoothed to remove isolated noise points and ensure the continuity and smoothness of the water body boundary.

[0150] S43, Accuracy Assessment

[0151] The extraction results are evaluated using indicators such as accuracy, mean intersection over union (MIoU) and recall to ensure that the model has stable extraction effects under different data source conditions.

[0152] Example 2: A water body automatic extraction device integrating multi-source remote sensing data

[0153] This embodiment introduces a water body extraction device based on multi-source remote sensing data, which aims to automatically process different types of remote sensing data and achieve high-precision water body extraction and result fusion.

[0154] Step 1: Data acquisition and preprocessing unit

[0155] S11, acquisition unit

[0156] Responsible for collecting optical images, SAR images and drone image data. The unit can connect with different remote sensing satellites or drone systems to ensure the integrity and timeliness of data acquisition.

[0157] S12, pre-processing unit

[0158] The received image data is standardized, including radiation correction, atmospheric correction and geometric correction. Different processing methods are used for optical images and SAR images to ensure that each image has accurate physical characteristics and geographic accuracy.

[0159] Step 2: Data type determination and processing unit

[0160] S21, data type determination unit

[0161] According to the type of input data, it automatically determines whether the data belongs to optical images or SAR images. If it is a four-channel optical image, the DeepLabv3+ model is selected for processing; if it is a SAR image, the UNet model is selected for processing.

[0162] S22, optical image processing module

[0163] Use the DeepLabv3+ model to extract water from four-channel optical image data. This module is responsible for loading the trained model and performing inference operations to generate water masks in optical images.

[0164] S23, SAR image processing module

[0165] The UNet model is used to extract water from SAR images. By extracting features and preserving edges from SAR images, the module can identify water areas in complex noise environments.

[0166] Step 3: Result fusion and output unit

[0167] S31, result fusion unit

[0168] When both optical and SAR image data are available, the water masks processed by DeepLabv3+ and UNet models will be transmitted to the fusion unit, which superimposes and fuses the results of the two data sources to generate a high-precision water range map by combining the high resolution of optical images and the all-weather capabilities of SAR images.

[0169] S32, post-processing and output unit

[0170] The fused water mask undergoes morphological operations, such as closing and smoothing, to eliminate noise and ensure the continuity and accuracy of the water body boundary.

[0171] The final water body extraction results will be exported into vector format (such as SHP) for geographic information system (GIS) analysis and further disaster assessment.

[0172] Step 4: Model training and updating unit

[0173] S41, model training unit

[0174] The device’s built-in model training unit is used to update the model based on the latest remote sensing data. By continuously training and updating the DeepLabv3+ and UNet models, the device can be adapted to water extraction tasks in different regions and under different meteorological conditions.

[0175] S42, Model Optimization and Tuning

[0176] According to user needs, the device can tune the trained model and adjust hyperparameters such as learning rate and batch size to ensure that in actual applications, the device can always provide high-precision water extraction results.

[0177] The above embodiments describe in detail the automatic water extraction method and device of the present invention, covering various aspects of data acquisition, preprocessing, model training and result fusion, and can be widely used in flood disaster monitoring, emergency response and disaster assessment and other fields.

[0178] The actual example verifies that the water body automatic extraction method and device of the present invention have been successfully applied to the actual flood monitoring in the Hudong flood detention area. The verification results are as follows: Figure 4 By using the multi-source remote sensing data fusion water body extraction model of the present invention, the system achieved an accuracy rate of 99% and a mean intersection over union (MIoU) ratio of 98% in the lake east flood detention area, greatly improving the accuracy of flood inundation range estimation and significantly enhancing the flood disaster warning capability in the region.

[0179] Compared with traditional monitoring methods, this invention significantly shortens the time of flood monitoring and early warning, and improves the efficiency of emergency response. According to the project report, the system can predict floods 7 days in advance and issue early warnings 3 days before the flood, providing the local government with sufficient emergency response time. This early warning significantly reduces the economic losses caused by the disaster and effectively protects the lives and property of local residents.

[0180] Taking the existing population of 285,000 in the Hudong flood detention area as an example, based on the early warning measures of the system, it is estimated that the disaster losses can be reduced by about 140 million yuan. This fully reflects the economic and social benefits of the invention, which not only improves the accuracy and timeliness of flood disaster monitoring, but also provides a scientific decision-making basis for disaster emergency response, and plays an important role in promoting disaster prevention and mitigation.

Claims

1. A method for automatic water extraction by integrating multi-source remote sensing data, characterized in that The following steps are involved: Step 1. Data acquisition Acquire multi-source remote sensing data, including optical remote sensing satellite images, synthetic aperture radar (SAR) data, and drone image data; optical remote sensing data includes R, G, B, and NIR bands, SAR data is C-band dual-polarization data (VV and VH), and drone images provide high-resolution optical images; Step 2. Data preprocessing Preprocess the optical remote sensing images and SAR images separately, including: Preprocessing of optical remote sensing data: radiometric calibration, orthorectification, atmospheric correction and image fusion; Preprocessing of SAR data: orbit correction, radiation calibration, denoising (using RefinedLee filter) and decibel processing to make it have physical and geometric accuracy for subsequent analysis; Step 3. Water index calculation For the preprocessed optical image, the normalized difference water index (NDWI) is calculated, and the formula is: The water index is used to distinguish between water and non-water areas; For SAR images, the water body index (SDWI) is calculated using the formula: SDWI=ln(10*VV*VH)-8 Then, the VV and VH polarization data were synthesized and stacked with the water body index (SDWI) to generate 3-channel image data for model training; Step 4. Model selection and water extraction Select the appropriate model according to the data type: if it is a four-channel optical image, use the improved DeepLabv3+ model for water extraction; if it is a SAR image, use the UNet model for water extraction; The DeepLabv3+ model can process multispectral information in optical images and is suitable for accurately segmenting water boundaries; The UNet model combines multi-layer features of SAR data through skip connections and is suitable for identifying water bodies in noisy and complex images. Step 5. Result fusion When both optical image and SAR image data exist, two water body masks are generated respectively, and the results of the two data sources are superimposed and fused. Combining the high resolution of optical images and the all-weather capability of SAR images, large-scale and high-precision water body extraction results are generated. Step 6. Post-processing and optimization Perform morphological processing on the fused water body mask, such as closing operation and smoothing, to remove noise, optimize the water body boundary, and ensure the accuracy and continuity of the water body extraction results; Step 7. Result output The processed water body identification results are exported into vector format (such as SHP file) for further analysis and application in geographic information system (GIS).

2. The automatic water extraction method according to claim 1, wherein: The DeepLabv3+ model has been improved to adapt to four-channel optical image input, especially in terms of edge extraction and multi-scale feature processing, to ensure accurate recognition of water body boundaries.

3. The automatic water extraction method according to claim 1, wherein: The UNet model is suitable for water body extraction in SAR images. It can effectively segment water areas in images with complex noises and retain the edge features of water areas through jump connections.

4. The automatic water extraction method according to claim 1, wherein: The post-processing of the water body mask adopts closing operation and median filtering technology to further eliminate isolated noise points and smooth the edge of the water body to ensure the accuracy of the final water body recognition result.

5. A water body automatic extraction device for realizing multi-source remote sensing data fusion, comprising: Step 1. A data acquisition unit is used to acquire optical remote sensing data, SAR data and drone image data; Step 2. A preprocessing unit is used to perform radiometric calibration, atmospheric correction, orthorectification, denoising and decibel conversion on the acquired data; Step 3. A model selection unit is used to select a suitable deep learning model according to the data type to process optical images and SAR images respectively; Step 4. Water extraction unit, including DeepLabv3+ model and UNet model, which are used for water extraction from optical images and SAR images respectively; Step 5. A result fusion unit is used to fuse the water body extraction results of the optical image and the SAR image to generate a high-precision water body mask; Step 6. A post-processing unit is used to perform morphological processing and optimization on the fused water body mask to ensure the accuracy and continuity of the water body recognition results; Step 7. Output unit, used to export the water body identification results into vector format for use in geographic information system analysis.

Images