Water body-non-water body classification method and system based on physical and data-driven dual mechanisms
By combining physical and data-driven methods, using the optical characteristics difference of SAR images and the random forest model, the problems of cloud occlusion, empirical dependence and high computing resources for water body extraction in the prior art are solved, and efficient and accurate water body extraction and explainable decision-making process are achieved.
Patent Information
- Application Number
- CN202510445025.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-10
- Publication Date
- 2025-07-18
AI Technical Summary
The prior art has problems such as cloud occlusion in flood monitoring and water extraction, which leads to monitoring failure, strong dependence on artificial experience, high requirements for computing resources, insufficient interpretability and limited generalization capabilities. It is especially difficult to achieve efficient and accurate water extraction in large-scale flood events.
Using a dual mechanism based on physical and data-driven method, multi-time phase SAR images are acquired, pre-processed and optical feature enhancement are performed, and a random forest model is used for training. Using the differences in optical characteristics between water bodies and ground objects, a visual water body-non-water body classification model is constructed to reduce dependence on manual annotation and improve the generalization ability of the model.
Fast and accurate water extraction in large-scale flood events in river basin was achieved, with an overall accuracy of 99.55%, and a Kappa coefficient of 0.9724, reducing human operation errors and computing resource requirements, and improving the interpretability and generalization capabilities of the model.
Smart Images
Figure CN120339706A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data classification, and more specifically, to a water body - non - water body classification method and system based on a dual mechanism of physics and data driving. Background Art
[0002] In the field of hydrological data acquisition and flood monitoring, remote sensing technology, by virtue of the combination of multi - sensors and platforms, breaks through geographical and climatic limitations and realizes large - scale data collection, with significant advantages. Classic optical satellites, such as Landsat and Sentinel - 2, observe the earth's surface by means of visible and near - infrared bands. However, Mie scattering caused by liquid water droplets and ice crystals in clouds will cause the reflectivity of remote sensing signals to increase abnormally when they are emitted to clouds. Given the inevitable connection between rainfall and clouds, in sudden short - duration, basin - wide emergency flood events, optical satellites are often unable to effectively carry out earth observation due to cloud cover, resulting in the failure of emergency monitoring.
[0003] Synthetic Aperture Radar (SAR), as an advanced remote sensing product in the late 20th century, uses centimeter - level microwave bands (C - band 5.6 cm, L - band 23 cm), and its wavelength is much larger than the cloud droplet diameter (5 - 100 μm). According to Rayleigh scattering theory, when the particle size is much smaller than the wavelength, the scattering can be ignored, which enables SAR to penetrate clouds several kilometers thick and achieve all - weather observation.
[0004] Traditional water body extraction methods mainly include threshold method, full - polarization processing, and change detection. The threshold method realizes water body segmentation by setting a global threshold of the backscattering coefficient. The Otsu algorithm automatically finds the optimal threshold by maximizing the between - class variance, and the KI algorithm assumes a Gaussian mixture distribution and minimizes the classification error. However, these methods are highly dependent on artificial experience, and subtle errors may lead to significant classification errors. Full - polarization SAR converts four polarization channels into an RGB image and maps it to the CIELab space to enhance water body identification by decomposing the surface scattering mechanism. However, the complex polarization decomposition process requires high computing resources. The change detection method identifies the change in backscattering by comparing pre - disaster and post - disaster images, and commonly uses a ratio operator to generate a difference map. However, it depends on accurate image registration, and the registration error will directly affect the result accuracy.
[0005] In recent years, artificial intelligence technology has opened up new paths for water body extraction. Machine learning models such as Random Forest and XGBoost have performed excellently in the monitoring of small and medium-sized river basins, being able to process multi-source data and suppress noise. However, whether it is supervised learning or unsupervised learning in machine learning, currently it generally faces the "black box dilemma". Due to insufficient interpretability, key information such as feature importance and decision-making paths is missing, making it difficult to trace misjudgments and researchers hard to understand the decision-making process of machine learning. In addition, the feature construction of mainstream machine learning models often lacks the guidance of physical mechanisms, resulting in limited generalization ability. Moreover, deep learning relies on large-scale labeled data for training, making it difficult to be promoted in areas with scarce data, and the labeling process also consumes a large amount of manpower and material resources.
[0006] Currently, in flood monitoring and water body extraction, whether it is optical satellite observation, traditional SAR algorithms, or machine learning technologies, there are certain defects. Therefore, how to provide a new technical solution to improve the accuracy, efficiency, interpretability, and generalization ability of water body extraction is an urgent problem for those skilled in the art. Summary of the Invention
[0007] In view of this, the present invention provides a water body - non-water body classification method and system based on a dual mechanism of physics and data driving, which can quickly and accurately extract water bodies when large-scale flood events occur in river basins by combining remote sensing and image processing technologies, thereby optimizing flood control and scheduling decisions at the river basin scale.
[0008] To achieve the above object, the present invention adopts the following technical solutions:
[0009] On the one hand, the present invention provides a water body - non-water body classification method based on a dual mechanism of physics and data driving, including:
[0010] Obtain multi-temporal SAR images;
[0011] Preprocess the SAR images;
[0012] Perform physical mechanism enhancement by fusing water body optical features on the preprocessed SAR images to obtain enhanced images;
[0013] Construct a classification model based on the Random Forest model, and train the classification model based on the enhanced images to obtain a trained classification model;
[0014] Input the enhanced images of the target area into the trained classification model to obtain the water body classification result.
[0015] Preferably, preprocessing the SAR images includes:
[0016] Perform filtering processing on the SAR images using the Frost filtering method;
[0017] Geocode and radiometric calibration are performed on the filtered SAR image.
[0018] Preferably, the pixel value calculation formula for filtering the SAR image using the Frost filtering method is:
[0019]
[0020] Among them, f(i,j) is the original pixel value in the SAR image; m is the neighborhood mean centered on (i,j); σ 2 is the neighborhood variance; γ is the control parameter.
[0021] Preferably, geocoding and radiometric calibration of the filtered SAR image specifically include:
[0022] Calculate the geographic coordinates of each pixel in the filtered SAR image through satellite orbit parameters and radar parameters, combined with the range-Doppler SAR image geometric positioning model;
[0023] Convert the digital quantization value of the geocoded SAR image into a physical backscattering coefficient, and the formula is:
[0024]
[0025] Among them, DN is the digital quantization value; C is the calibration constant; R is the slant range from the radar antenna to the target; θ is the radar incident angle.
[0026] Preferably, enhance the physical mechanism of fusing the optical characteristics of water bodies in the preprocessed SAR image to obtain an enhanced image, including:
[0027] Perform feature-enhanced band calculation on the preprocessed SAR image, and the formula is as follows:
[0028]
[0029] Among them, HH is the radar backscattering coefficient in the HH polarization mode; HV is the radar backscattering coefficient in the HV polarization mode.
[0030] On the other hand, the present invention provides a water-non-water classification system based on a dual mechanism of physics and data-driven, including:
[0031] An acquisition module for acquiring multi-temporal SAR images;
[0032] A preprocessing module for preprocessing the SAR image;
[0033] A feature enhancement module, which is used to enhance the physical mechanism of fusing water body optical features for the pre - processed SAR image to obtain an enhanced image;
[0034] A model construction module, which is used to construct a classification model based on a random forest model and train the classification model based on the enhanced image to obtain a trained classification model;
[0035] A classification module, which is used to input the enhanced image of the target area into the trained classification model to obtain a water body classification result.
[0036] As can be seen from the above - mentioned technical solutions, compared with the prior art, the present invention discloses a water body - non - water body classification method and system based on a dual mechanism of physics and data - driven. By using the physical mechanism of the optical property difference between water bodies and ground objects, the difference between water bodies and other ground objects in the image is enlarged through mathematical methods to enhance the representativeness of the input feature parameters of the random forest; combining the above - mentioned physical mechanism with the machine learning mechanism, integrating various feature parameters such as gray - scale features and texture features in the SAR image, enhancing the learning ability and interpretability of the random forest classifier to visualize the model decision - making process, and further providing a reasonable reference for flood control decision - making. Description of the Drawings
[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.
[0038] Figure 1 It is the technical roadmap of the classification method provided by the present invention;
[0039] Figure 2 It is a schematic diagram of the signal reception and scattering phenomenon of water bodies / other ground objects;
[0040] Figure 3 It is the input feature map of the random forest;
[0041] Figure 4 It is the water body extraction result of the radar image of a certain flood detention area;
[0042] Figure 5 It is the structural framework diagram of the classification system provided by the present invention. Detailed Embodiments
[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0044] An embodiment of the present invention discloses a water body - non - water body classification method based on a dual - mechanism of physics and data - driven, as Figure 1 shown, including:
[0045] Obtain multi - temporal SAR images;
[0046] Pre - process the SAR images;
[0047] Enhance the physical mechanism of the SAR images after pre - processing by fusing the optical features of water bodies to obtain enhanced images. The smooth and flat characteristics of the water surface make its backscattering coefficient small, while the rough characteristics of the ground object surface make its backscattering coefficient large. The difference between water bodies and ground objects is enlarged by calculating the waveform of the SAR images.
[0048] Construct a classification model based on the random forest model, and train the classification model based on the enhanced images to obtain a trained classification model. In the process of establishing the model, it is necessary to manually make water body and non - water body labels for classification in the images. At this time, the random forest will start the image processing module to extract the gray - scale features and texture features of the images. The visualization method of gray - scale features is to calculate the co - occurrence gray - scale matrix of the matrix - type image and obtain the target features of the model: homogeneity, energy, and entropy. The calculation formulas of the three indicators are as follows:
[0049]
[0050] where p(i,j) is the probability of the pair of gray - scale values i and j appearing in the co - occurrence gray - scale matrix; N is the number of gray - scale levels of the image.
[0051] The visualization method of texture features is to calculate the local binary pattern and the frequency - domain features based on the Haar wavelet transform at the pixel scale in the image. The local binary pattern is a classic method for describing local texture features of images. Its core idea is to generate a binary pattern by comparing the gray - scale difference between the central pixel and its neighboring pixels. The standard local binary formula is as follows:
[0052]
[0053] where P is the number of neighboring pixels; R is the neighborhood radius; s(x) is the sign function; g p is the gray - scale value of the central pixel; g cis the gray value of neighborhood pixels. The Haar wavelet transform decomposes the image into components of different frequencies. Water bodies usually appear as low backscatter regions in radar images, with clear edges from surrounding ground features (such as vegetation and buildings), and the high-frequency subbands of the Haar wavelet can effectively capture these edges.
[0054] After the training of the random forest model is completed, the model will enable the validation set automatically divided from the manually created water-non-water labels to further carry out the model validation work. In this example, the inversion work of the "23·7" extreme flood event in the Haihe River Basin was carried out. According to the flood description and measured data in the "Analysis Report on the Simulation and Replay of the "23.7" Extreme Flood in the Haihe River Basin" jointly published by the Haihe River Water Conservancy Commission of the Ministry of Water Resources, China Institute of Water Resources and Hydropower Research, and Hohai University in December 2023, the water and non-water labels of the radar image located in the Dongdian Flood Detention and Retention Area on August 12, 2023 were made and used for model training and validation. Figure 4 is the water body extraction result obtained by using the classification method provided by the present invention. The extraction result of the water body is very close to the actual measurement, with an overall accuracy reaching 99.55% and a Kappa coefficient reaching 0.9724. After verification, the performance of the model can reach the level of real-time flood monitoring.
[0055] 1) Different from the classical "black box" model, in the embodiment of the present invention, the features input to the classifier are manually selected, visualizing the decision-making process of machine learning.
[0056] 2) A high-performance water body extraction model is trained with a small-scale data set, reducing the burden of manually making labels and reducing errors caused by human operation mistakes or lack of experience.
[0057] 3) The random forest model with strong generalization ability and expression ability is selected, reducing the usage threshold of the model for users to a certain extent.
[0058] The enhanced image of the target area is input into the trained classification model to obtain the water body classification result. The SAR image in raster TIF format after enhanced features is used as the base map, and after being matrixed, it is input into the classification model in the format of a two-dimensional array. The input image is as Figure 3 shown. The classification model first performs automated cutting and tiling parallel processing on the input enhanced image, dividing the enhanced image into multiple sub-images to improve the model performance and calculation efficiency; then, multiple sub-images respectively pass through the image processing module in the classification model to extract the gray features and texture features of the image; finally, based on the gray features and texture features, a decision tree is used to classify the water body-non-water body in the target area.
[0059] Furthermore, preprocess the SAR image, including:
[0060] The Frost filtering method is used to filter the SAR image, effectively ensuring the edge features of the image;
[0061] Geocoding and radiometric calibration are performed on the filtered SAR image to convert the pixel information of the radar image into the backscattering coefficient.
[0062] Specifically, for the filtering of the synthetic aperture radar data image, the Frost filtering method of adaptive speckle noise suppression is used for filtering, while suppressing noise, the edge information of the image is retained; the calculation formula of the pixel value after filtering is:
[0063]
[0064] Among them, f(i,j) is the original pixel value in the SAR image; m is the neighborhood mean centered on (i,j); σ 2 is the neighborhood variance; γ is the control parameter. The exponential term is the adaptive weight. When is larger (such as the edge or strong scattering area), the weight approaches 1, retaining more original information; is smaller (such as the uniform area), the weight approaches 0, enhancing the smoothing effect.
[0065] Specifically, geocoding and radiometric calibration are performed on the filtered SAR image, specifically including:
[0066] Through the satellite orbit parameters (position and speed when taking the image) and radar parameters (wavelength, pulse repetition frequency), combined with the range-Doppler SAR image geometric positioning model, calculate the geographic coordinates of each pixel in the filtered SAR image;
[0067] Convert the digital quantization value of the geocoded SAR image into the physical backscattering coefficient, eliminating the influence of factors such as sensors and the atmosphere. The radiometric calibration formula is:
[0068]
[0069] Among them, DN is the digital quantization value; C is the calibration constant; R is the slant range from the radar antenna to the target, that is, the straight-line distance between the radar and the ground target; θ is the radar incident angle.
[0070] In another embodiment, as Figure 2 shown, while the water body absorbs the incident wave, due to its smooth surface, it will produce near-specular reflection, and most of the ground objects will produce diffuse reflection due to their rough surfaces. Therefore, it is necessary to perform band calculation for feature enhancement on the preprocessed SAR image to adjust the distribution of larger and smaller backscattering coefficients. The band calculation formula is as follows:
[0071]
[0072] Among them, HH is the radar backscattering coefficient under the HH polarization mode; HV is the radar backscattering coefficient under the HV polarization mode. The HH and HV polarization modes are the characteristic polarization modes selected for the GF-3 synthetic aperture radar images in this embodiment.
[0073] On the other hand, the present invention provides a water body - non - water body classification system based on a dual mechanism of physics and data - driven, as Figure 5 shown, including:
[0074] An acquisition module, configured to obtain multi - temporal SAR images;
[0075] A pre - processing module, configured to pre - process the SAR images;
[0076] A feature enhancement module, configured to perform physical mechanism enhancement by fusing water body optical features on the pre - processed SAR images to obtain enhanced images;
[0077] A model construction module, configured to construct a classification model based on a random forest model and train the classification model based on the enhanced images to obtain a trained classification model;
[0078] A classification module, configured to input the enhanced images of the target area into the trained classification model to obtain water body classification results.
[0079] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.
[0080] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A water body - non - water body classification method based on a dual - mechanism of physical and data - driven, characterized in that, Including: Obtain multi-temporal SAR images; Preprocess the SAR images; Perform physical mechanism enhancement by fusing the optical characteristics of water bodies on the preprocessed SAR images to obtain enhanced images; Construct a classification model based on the random forest model and train the classification model based on the enhanced images to obtain a trained classification model; Input the enhanced images of the target area into the trained classification model to obtain the water body classification result.
2. The water body - non - water body classification method based on the dual mechanisms of physics and data - driven according to claim 1, characterized in that, Preprocessing the SAR images includes: Filter the SAR images using the Frost filtering method; Perform geocoding and radiometric calibration on the filtered SAR images.
3. A water-non-water classification method based on a dual mechanism of physics and data-driven according to claim 2, characterized in that, The pixel value calculation formula after filtering the SAR images using the Frost filtering method is: Among them, f(i, j) is the original pixel value in the SAR image; m is the neighborhood mean centered on (i, j); σ 2 is the neighborhood variance; γ is the control parameter.
4. A water-non-water classification method based on a dual mechanism of physics and data-driven according to claim 2, characterized in that Performing geocoding and radiometric calibration on the filtered SAR images specifically includes: Calculate the geographic coordinates of each pixel in the filtered SAR image through satellite orbit parameters and radar parameters, combined with the range-Doppler SAR image geometric positioning model; Convert the digital quantization value of the geocoded SAR image into the physical backscattering coefficient, and the formula is: Where, DN is the digital quantization value; C is the calibration constant; R is the slant range from the radar antenna to the target; θ is the radar incident angle.
5. A water-non-water classification method based on a dual mechanism of physics and data-driven according to claim 1, characterized in that Performing physical mechanism enhancement by fusing the optical characteristics of water bodies on the preprocessed SAR images to obtain enhanced images includes: Perform feature-enhanced band calculation on the preprocessed SAR images, and the formula is as follows: Where, HH is the radar backscattering coefficient in the HH polarization mode; HV is the radar backscattering coefficient in the HV polarization mode.
6. A water-non-water classification system based on a dual mechanism of physics and data-driven, characterized in that, Including: An acquisition module for obtaining multi-temporal SAR images; A preprocessing module for preprocessing the SAR images; A feature enhancement module for performing physical mechanism enhancement by fusing the optical characteristics of water bodies on the preprocessed SAR images to obtain enhanced images; A model construction module for constructing a classification model based on the random forest model and training the classification model based on the enhanced images to obtain a trained classification model; A classification module for inputting the enhanced images of the target area into the trained classification model to obtain the water body classification result.