A water body extraction method fusing cygnss data and sar data

Abstract

The application provides a water body extraction method fusing CYGNSS data and SAR data, and belongs to the technical field of remote sensing disaster monitoring. The method utilizes the high sensitivity of CYGNSS data to the change of ground roughness and the high spatial resolution advantage of SAR data, realizes deep fusion and resolution improvement of the two types of data through a random forest model, solves the resolution adaptation problem, and can enhance the separability of water body and non-water body, thereby improving the water body recognition accuracy. The application also introduces a DEM perception superpixel improvement algorithm, and is no longer limited to single slope threshold rejection, but constructs an effective water body elevation range by analyzing the elevation distribution of high-confidence water bodies in the whole image. The range is used to constrain the superpixel, which can effectively eliminate the shadow of high-altitude mountains, and can recall the weak signal water bodies missed through the elevation feature, thereby significantly improving the water body extraction accuracy under complex terrain.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing disaster monitoring technology, specifically relating to a method for water body extraction that integrates CYGNSS data and SAR data. Background Technology

[0002] As one of the most widespread and frequent natural disasters globally, timely and accurate information on the extent of water inundation is crucial for disaster emergency response, disaster assessment, and post-disaster reconstruction.

[0003] Existing water body extraction technologies have many shortcomings. Most of them rely on a single data source, which has limitations. Optical remote sensing data is limited by clouds and lighting conditions, making it difficult to obtain effective observations during floods. Although single SAR data has all-weather observation capabilities, it is sensitive to terrain roughness and water bodies and non-water bodies are easily mixed in mountain shadows and vegetated areas. Although single CYGNSS (Cyclone Global Navigation Satellite System) data can identify water bodies through differences in scattering mechanisms, the original data has insufficient spatial resolution, and it is difficult to achieve the 10-meter level fine detection requirements in complex scenarios when used alone.

[0004] Traditional data fusion is inefficient. Existing fusion methods often use simple superposition or linear combination, without establishing an effective mapping relationship between high-resolution features (such as SAR polarization features) and low-resolution features (such as CYGNSS reflectivity). This results in the resolution compatibility problem between CYGNSS and SAR data not being solved, and the complementary characteristics of the two are not fully explored. As a result, the accuracy of water body extraction is limited, and information redundancy is easily generated.

[0005] In addition, traditional methods rely solely on a single slope threshold to eliminate false water bodies, without combining elevation distribution characteristics to construct global constraints, which easily leads to the omission of weak signal water bodies caused by vegetation obstruction and the influence of wind and waves.

[0006] In summary, existing water body extraction methods are insufficient to meet the high-precision requirements of flood disaster emergency monitoring in terms of resolution, anti-interference ability, and terrain adaptability. There is an urgent need for a water body extraction technology that can fully leverage the complementary advantages of CYGNSS and SAR data and balance high resolution and high accuracy, thereby achieving high-precision extraction of flood-affected areas. Summary of the Invention

[0007] In view of this, the present invention provides a water body extraction method that integrates CYGNSS data and SAR data. By coordinating the reflectivity of CYGNSS data and the backscattering characteristics of SAR data, as well as an improved DEM-sensing superpixel algorithm, the accuracy and reliability of water body extraction can be improved, thereby achieving high-precision extraction of flood-affected areas. This invention is particularly suitable for the identification and extraction of water body extents in complex terrain areas under flood disaster scenarios.

[0008] To solve the above-mentioned technical problems, the present invention is implemented as follows.

[0009] A method for water body extraction that integrates CYGNSS and SAR data includes:

[0010] Step 1: Acquire CYGNSS and SAR data of the area to be water body extracted; calculate CYGNSS signal-to-noise ratio reflectance based on CYGNSS data; extract polarization features from SAR data; obtain slope map using digital elevation model (DEM);

[0011] Step 2: Perform data fusion of CYGNSS signal-to-noise ratio reflectance and polarization features based on the random forest model to obtain high-resolution CYGNSS signal-to-noise ratio reflectance that incorporates SAR data and has a resolution improved to the same level as SAR data;

[0012] Step 3: Based on the high-resolution CYGNSS signal-to-noise ratio reflectance, water body extraction is performed to generate initial water body prediction results;

[0013] Step 4: Based on the slope map, filter the areas with gentle slopes from the initial water body prediction results to obtain high-confidence water bodies; statistically analyze the elevation range of the high-confidence water bodies as a global geographic feature constraint;

[0014] Step 5: Perform superpixel segmentation on the high-resolution CYGNSS signal-to-noise ratio reflectance, and determine the water proportion of each superpixel based on the initial water body prediction results; if the water proportion of a superpixel is greater than the first threshold and the slope meets the water body requirements, it is determined to be a water body area; then perform DEM-sensing weak signal recall: for superpixels that do not meet the first threshold, if the water proportion is greater than the second threshold, the second threshold is less than the first threshold, and the elevation meets the global geographic feature constraints, and the slope meets the water body requirements, then the superpixel is recalled and marked as a water body using the elevation consistency principle; finally, output a high-precision flood inundation map optimized by DEM-sensing weak signal recall.

[0015] Preferably, step 1 further includes a preprocessing operation, including:

[0016] Multidimensional constraint quality control was performed on the acquired CYGNSS data: antenna gain, incident angle, signal-to-noise ratio, and quality flag bits were read from the CYGNSS data; a physical parameter filter was constructed to remove invalid observation data with abnormal antenna gain, large incident angle, and low signal-to-noise ratio; a window for the peak position of the delay Doppler image was set to remove edge-truncated data; and bit masking technology was used to decode and filter the quality flag bits to remove noise data containing attitude instability, instrument calibration anomalies, and invalid specular reflection points.

[0017] The acquired SAR data is subjected to radiometric correction, geometric correction, and denoising.

[0018] Preferably, the specific data fusion method based on the random forest model in step 2 is as follows:

[0019] During the training phase, the latitude and longitude coordinates inherent in the CYGNSS data are used as sample locations. Polarization features with the same resolution as the CYGNSS data are extracted based on the sample locations, referred to as low-resolution polarization features. Similarly, the CYGNSS signal-to-noise ratio reflectance and low-resolution polarization features at the sample locations constitute training samples to train a random forest model, enabling the random forest model to learn the mapping relationship between polarization features and CYGNSS signal-to-noise ratio reflectance. During the inference phase, the high-resolution polarization features extracted from the SAR data in step 1 are input into the random forest model to obtain the CYGNSS signal-to-noise ratio reflectance that fuses the SAR data and is of the same high resolution as the SAR data, which is the high-resolution CYGNSS signal-to-noise ratio reflectance.

[0020] Preferably, in the random forest model, the number of decision trees is set to 100, the maximum tree depth is 10, the minimum number of sample splits is 5, and the training samples are divided into a training set and a validation set in a 7:3 ratio. The model parameters are optimized through the training set, and the model performance is evaluated through the validation set.

[0021] Preferably, step 3 specifically includes:

[0022] Based on the high-resolution CYGNSS signal-to-noise ratio reflectance obtained in step 2, water bodies are extracted using at least two classification algorithms. The water body extraction results of the two classification algorithms are then fused through a voting mechanism to generate initial water body prediction results.

[0023] Preferably, the step of using at least two classification algorithms to extract water bodies and fusing the water body extraction results of the two classification algorithms through a voting mechanism to generate an initial water body prediction result specifically includes:

[0024] Three algorithms—K-means clustering, Otsu thresholding, and Gaussian mixture model (GMM)—are used for water body extraction. If a pixel is identified as a water body by at least two of the algorithms, then that pixel is included in the initial water body prediction result.

[0025] Preferably, the screening criterion for the gently sloping area in step 4 is: the slope is less than or equal to 5°.

[0026] Preferably, the step 4, which involves statistically analyzing the elevation range of the high-confidence water bodies as a global geographic feature constraint, is as follows: statistically analyzing the elevation distribution characteristics of the high-confidence water bodies, and using the elevation range from the 5th percentile to the 95th percentile as the global geographic feature constraint.

[0027] Preferably, in step 5, the first threshold is 50% and the second threshold is 20%.

[0028] Preferably, in step 5, the criterion for determining whether the slope meets the requirements of the water body is: the average slope of the superpixel is less than 15°.

[0029] Beneficial effects:

[0030] (1) Data Synergy and Complementary Advantages: Fully leveraging the high sensitivity of CYGNSS data to changes in surface roughness (differences in scattering mechanisms between water and land) and the high spatial resolution of SAR data, a random forest model is used to achieve deep fusion of the two types of data, overcoming the limitations of a single data source. Figure 5 and Figure 6 As shown in neutron diagram (a), the separability between water bodies and non-water bodies is enhanced, thereby improving the accuracy of water body identification.

[0031] (2) High-resolution downscaling: The mapping relationship between SAR high-resolution features and CYGNSS reflectivity is established by using a random forest model. The 1km resolution CYGNSS data is downscaled to 10 meters, which takes into account the water body identification sensitivity of CYGNSS data and the spatial detail expression ability of SAR data. This solves the resolution adaptation problem between CYGNSS and SAR data and meets the needs of refined monitoring.

[0032] (3) Dual constraints of terrain and elevation: The superpixel improvement algorithm based on DEM perception is no longer limited to a single slope threshold for removal. Instead, it constructs an effective water body elevation range by analyzing the elevation distribution of high-confidence water bodies in the entire map. By using this range to constrain the superpixels, it can effectively remove shadows of high-altitude mountains (false water bodies) and recall weak signal water bodies that were missed by using elevation features, which significantly improves the accuracy of water body extraction under complex terrain.

[0033] (4) Automation and efficiency: The entire process, from data preprocessing, reflectivity calculation, downscaling, terrain interference removal to water extraction, is automated without much manual intervention. Furthermore, through superpixel and clustering algorithms, the computational efficiency is improved while ensuring accuracy, making it suitable for the rapid response needs of flood disaster emergency monitoring.

[0034] (5) Verification of feature separability and algorithm robustness: Through M-statistic and probability density function (PDF) analysis, it is shown that the SNR signal-to-noise ratio data of CYGNSS is significantly better than that of traditional SAR polarization features (VH M=2.408, VV M=2.123) than that of water bodies and non-water bodies (M=3.231), which verifies the necessity of introducing CYGNSS data. At the same time, through threshold sensitivity and overlap (OVL) analysis, the feature overlap of the two types of land cover is extremely low (OVL≤0.03), which proves that the method has extremely high robustness in feature fusion and threshold determination and is not easily affected by small parameter fluctuations. Attached Figure Description

[0035] Figure 1 This is a flowchart of the water body extraction method that integrates CYGNSS data and SAR data according to the present invention.

[0036] Figure 2 The graph shows the probability distribution of water / non-water bodies with multi-source characteristics and the analysis of the M-statistic.

[0037] Figure 3 This is a graph showing the feature threshold sensitivity and overlap (OVL) analysis.

[0038] Figure 4 A scatter plot for the accuracy assessment of signal-to-noise ratio reflectivity at 10-meter resolution.

[0039] Figure 5 This is an illustration of the water extraction effect based on a single SAR image using existing technology.

[0040] Figure 6 This is a water body extraction effect diagram based on the fusion of CYGNSS reflectance and SAR data using random forest, as presented in this invention. Detailed Implementation

[0041] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0042] This invention provides a method for water body extraction that integrates CYGNSS and SAR data. The core idea is to establish an effective mapping relationship between high-resolution SAR polarization features and low-resolution CYGNSS reflectance features using a random forest model. The trained random forest model is then used to fuse CYGNSS signal-to-noise ratio (SNR) reflectance and polarization features, simultaneously improving the resolution of the originally low-resolution CYGNSS SNR reflectance to the same level as SAR data, thus obtaining high-resolution CYGNSS SNR reflectance and solving the resolution compatibility problem between CYGNSS and SAR data. Furthermore, the mapping relationship constructed based on the random forest model also includes polarization features as input, meaning the output high-resolution CYGNSS SNR reflectance incorporates the characteristics of both CYGNSS and SAR data. The high-resolution SNR reflectance data combines the spatial details of SAR with the high water body sensitivity of CYGNSS, enhancing the separability between water and non-water bodies and laying the foundation for subsequent high-precision extraction.

[0043] Furthermore, when eliminating false water bodies, this invention combines global geographic feature constraints constructed based on the elevation distribution characteristics of high-confidence water bodies to implement DEM-sensing weak signal recall. By recalling weak signal water bodies that are easily missed due to vegetation obstruction and the influence of wind and waves, the accuracy of water body extraction is further improved.

[0044] Figure 1 A flowchart of the water body extraction method integrating CYGNSS and SAR data according to the present invention is shown in the figure. The method includes the following steps:

[0045] Step 1: Data Acquisition and Preprocessing.

[0046] Acquire CYGNSS, SAR, and elevation data for the area to be extracted from the water body. In practice, CYGNSS Level 1 Science Data Record Version 3.2 data, Sentinel 1 SAR data, and Digital Elevation Model (DEM) data can be acquired.

[0047] Multidimensional constraint quality control was performed on CYGNSS data: parameters such as gain, incident angle, signal-to-noise ratio, and quality flags were read from the data; a physical parameter filter was constructed to remove invalid observations with abnormal antenna gain, large incident angle, and low signal-to-noise ratio; a peak position window for the Delayed Doppler Map (DDM) was set to remove edge-truncated data; and bitwise masking was used to decode and filter the quality flags to strictly remove noisy data containing attitude instability, instrument calibration abnormalities, and invalid specular reflection points, and to extract the physical quantities of valid observation points.

[0048] SAR data preprocessing includes at least radiometric and geometric correction to ensure spatial registration accuracy between the image and CYGNSS data is better than one pixel. Preferably, SAR data preprocessing also includes denoising, preferably using Lee filtering to remove speckle noise from the SAR image, thus improving feature extraction accuracy.

[0049] Digital elevation model (DEM) data downloaded from a geospatial data cloud was used to calculate slope information for the study area, providing a basis for subsequent terrain interference processing. In one specific scheme, based on 30m resolution DEM data, ArcGIS software was used to calculate the slope map of the study area, with the slope calculation window set to 3×3. Then, resampling was performed to spatially register it with the 10m resolution reflectance results.

[0050] Step 2: Calculate the CYGNSS signal-to-noise ratio reflectance (hereinafter referred to as signal-to-noise ratio reflectance) based on CYGNSS data; extract polarization features from SAR data.

[0051] The physical scattering mechanism based on CYGNSS data is used to calculate the signal-to-noise ratio reflectance using the following formula. :

[0052]

[0053] in, The signal-to-noise ratio (SNR) of the DDM (unit: dB) is calculated from the original DDM maximum value and average noise. GPS effective isotropic radiated power (unit: W); Receiver antenna gain (unit: dBi); Distance from the GPS signal transmitter to the mirror reflection point (unit: m); Distance from the CYGNSS receiver to the mirror reflection point (unit: m); The wavelength of the GPS L1 band signal is approximately 19.3 cm. The angle of incidence at the point of reflection of the mirror (unit: degrees).

[0054] Substituting the preprocessed CYGNSS parameters into the above formula, the signal-to-noise ratio reflectivity of a 1km grid is calculated. Wherein, the GPS L1 band wavelength... Take 19.3cm, angle of incidence It is calculated using orbital parameters and DEM data provided by CYGNSS data.

[0055] In this step, polarization features are extracted from SAR data, specifically VV and VH polarization features.

[0056] Step 3: Data fusion based on the random forest model.

[0057] This step uses a random forest model to fuse signal-to-noise ratio reflectance and polarization features to obtain high-resolution signal-to-noise ratio reflectance that integrates SAR data and improves the resolution to the same level as SAR data.

[0058] Before constructing the downscaling model, we first perform feature distribution analysis on the preprocessed SAR polarization features (VH, VV) and the calculated CYGNSS signal-to-noise ratio reflectance (SNR):

[0059] Probability distribution analysis (PDF): Statistically analyze the probability density distribution of water and non-water samples under three characteristics: VH, VV, and SNR. See also... Figure 2 The results showed that the peak separation between water and non-water bodies was good for all three characteristics.

[0060] M-statistic assessment: The M-statistic is introduced to quantify the distinguishing ability of three different characteristics on water bodies (the larger the M value, the higher the separation). See also Figure 2 The analysis results show that the M-statistic of the SNR feature is as high as 3.231, which is superior to VH polarization (M=2.408) and VV polarization (M=2.123). This fully demonstrates that CYGNSS data has a stronger physical sensitivity to water body identification mechanisms than traditional SAR data, and establishes the rationality of the fusion strategy of using high spatial resolution SAR features to predict high physical sensitivity CYGNSS features.

[0061] Stability and overlap analysis: Threshold sensitivity analysis was performed on the three features. See also Figure 3 The calculation results show that the water / non-water body distribution overlap (OVL) of the three remains at an extremely low level (VH OVL=0.02, VV OVL=0.03, SNR OVL=0.03), indicating minimal feature confusion. Furthermore, within the central interval of the normalized threshold, the rate of change of extracted area approaches 0, indicating that the algorithm is insensitive to threshold selection and has high robustness.

[0062] Based on the above analysis, M-statistic and probability density function (PDF) analysis demonstrates that the SNR signal-to-noise ratio data of CYGNSS is significantly better than that of traditional SAR polarization features (VH M=2.408, VV M=2.123) in terms of separability between water and non-water bodies (M=3.231), verifying the necessity of introducing CYGNSS data. Simultaneously, threshold sensitivity and overlap (OVL) analysis shows that the feature overlap between the two types of land cover is extremely low (OVL≤0.03), proving that this method has extremely high robustness in feature fusion and threshold determination, and is not easily affected by small parameter fluctuations. This establishes the technical route of inverting highly separable SNR features using high-resolution SAR features. Using latitude and longitude information obtained by calculating CYGNSS reflectance on the Google Earth Engine platform as sample points, the 1km CYGNSS SNR reflectance (dependent variable) and 10m SAR VV and VH polarization features (predictor variables) corresponding to each sample point are extracted for model training.

[0063] Specifically, during the training phase, the latitude and longitude coordinates (1km grid) inherent in the CYGNSS data are used as sample locations. The signal-to-noise ratio (SNR) reflectance calculated from the CYGNSS data is then further divided into 1km grids to obtain the SNR reflectance at each sample location, which is used as the dependent variable. Based on the sample locations, VV and VH polarization features extracted from SAR data are obtained as prediction variables. The original resolution of the polarization features is relatively high, typically 10 meters. During training, to match the SNR reflectance calculated from the CYGNSS data, low-resolution (1km grid) VV and VH polarization features are extracted here based on the sample locations.

[0064] Samples with the same signal-to-noise ratio reflectance and polarization feature resolution at the same location (both low resolution) are used as training samples to train a random forest model. This allows the random forest model to learn the fusion and mapping relationship between polarization features and signal-to-noise ratio reflectance. By constructing random forest inputs and labels with the same resolution, an effective mapping relationship between high-resolution SAR polarization features and low-resolution CYGNSS reflectance features is established, solving the resolution adaptation problem between CYGNSS and SAR data.

[0065] The downscaling process involves applying a trained random forest model to 10-meter SAR feature data across the entire study area, outputting CYGNSS signal-to-noise ratio reflectance at 10-meter resolution. Specifically, during the inference phase, the high-resolution polarization features extracted from the SAR data in step 1 are input into the random forest model. The random forest model outputs a signal-to-noise ratio reflectance that fuses the SAR data and is of the same high resolution as the SAR data. This downscales the 1km resolution CYGNSS signal-to-noise ratio reflectance to 10-meter resolution, yielding the 10-meter resolution signal-to-noise ratio reflectance result.

[0066] like Figure 4 As shown, after evaluation, the predicted values, after linear fitting, have an R (correlation coefficient) of 0.69, a p (significance level) of <0.05, and an RMSE (root mean square error) of 1.542, meeting the requirements for signal-to-noise ratio reflectance and ensuring the reliability of the downscaling results. Combined with the aforementioned M-statistic analysis, the downscaled high-resolution SNR data retains both the spatial detail of SAR (10 m) and the high water sensitivity of CYGNSS (M=3.231), laying the foundation for subsequent high-precision extraction.

[0067] In one preferred scheme, the number of decision trees in the random forest model is set to 100, the maximum tree depth is 10, and the minimum number of sample splits is 5. The training samples are divided into training set and validation set in a 7:3 ratio. The model parameters are optimized through the training set, and the model performance is evaluated through the validation set.

[0068] Step 4: Based on the improved signal-to-noise ratio reflectance obtained in Step 3, extract water bodies and generate initial water body prediction results.

[0069] The signal-to-noise ratio reflectance data at 10-meter resolution were normalized by using the min-max normalization method to map the data to the [0,1] interval, thus eliminating dimensional differences.

[0070] A multi-algorithm initial extraction scheme is adopted, using at least two classification algorithms for water body extraction. The water body extraction results from the two classification algorithms are then fused through a voting mechanism to generate initial water body prediction results. In an optimal scheme, based on the feature differences between water bodies and non-water bodies obtained in step three, three unsupervised classification algorithms are run respectively:

[0071] K-means clustering: Set up 2 classes (water body, background), and take the class with the smallest mean as the water body;

[0072] Otsu thresholding method: Taking advantage of the long-tail distribution characteristics, the high-value portion is segmented twice to accurately locate the threshold;

[0073] Gaussian Mixture Model (GMM): It is set with 3 Gaussian components and uses a probability model to adapt to complex distributions.

[0074] If a pixel is identified as a water body by at least two algorithms, that pixel is included in the initial water body prediction result. In practice, the extraction results from the three algorithms are overlaid and averaged (Voting) to obtain a probability map of each pixel belonging to a water body. If the probability of a pixel is ≥0.5 (i.e., identified as a water body by at least two algorithms), that pixel is included in the initial water body prediction result, generating an initial binarized water body mask and a CYGNSS signal-to-noise ratio reflectance map after the mask. The initial binarized water body mask is a binary mask, and the signal-to-noise ratio reflectance image after the mask is a CYGNSS signal-to-noise ratio reflectance data image obtained using the binary mask. This step utilizes the advantages of multi-algorithm integration to initially suppress random noise at the pixel level.

[0075] Step 5: Based on the initial water body prediction results, perform DEM elevation feature analysis and construct global geographic feature constraints.

[0076] In this step, DEM data and slope maps are used to assist in the analysis. The initial water body prediction results are used to screen for areas with gentle slopes to obtain high-confidence water body samples. The preferred screening criterion is that a pixel with a slope of 5° or less is considered a high-confidence water body pixel. These high-confidence water body pixels form areas with gentle slopes, which are then used as high-confidence water body samples.

[0077] The elevation distribution characteristics of DEM samples with high confidence are statistically analyzed to calculate the effective water body elevation range. The preferred range is between the 5th and 95th percentiles, with a buffer value (e.g., 5 meters) allowed at both ends to determine the effective water body elevation range for the current scenario. This range also serves as a global geographic feature constraint for subsequent DEM-based weak signal retrieval.

[0078] Step Six: DEM-aware superpixel generation and global optimization.

[0079] This step aims to utilize the initial prediction map and effective water body elevation range obtained in step five to address the salt-and-pepper noise problem through superpixel segmentation technology, and to identify missed weak signal water bodies using elevation context information. The specific operations are as follows:

[0080] Step 61: Generate superpixels for the entire image.

[0081] This step employs the Felzenszwalb graph theory segmentation algorithm to perform full-image segmentation on the signal-to-noise ratio reflectance image normalized in step four. To prevent invalid values ​​(NaN) from image edges or masked regions from interfering with the segmentation algorithm, invalid regions are filled with minimum values ​​as background before segmentation. The superpixel scale parameter (scale) is set to 100, and the minimum size (min_size) is set to 50 to generate superpixel blocks that fit the boundaries of ground features.

[0082] Step 62: Superpixel-level feature statistics and hierarchical determination.

[0083] This step iterates through each generated superpixel, calculating three key features within the block: the initial predicted water body proportion (calculated based on the initial water body prediction results from step four), the average slope, and the average elevation. The final water body determination is then performed using the following dual logic:

[0084] First, water body determination is performed based on majority voting and terrain constraints (Logic A): If the proportion of the initially predicted water body within a superpixel is high and the slope meets the water body requirements, then the superpixel area is considered to have a strong water body signal and conforms to the terrain logic, and is directly confirmed as a water body. The determination of "high proportion" can be that the water body proportion is greater than a first threshold, for example, set to 50%. "Slope meets water body requirements" is determined using the average slope of the superpixel, preferably less than 15°. This step effectively removes false water bodies caused by mountain shadows, even if they have low reflectivity (like water) but are located on steep slopes (average slope > 15°).

[0085] Then, weak signal recall based on DEM perception is performed: For weak signal water areas in SAR imagery that are susceptible to vegetation obstruction or wind and waves, if the superpixel does not meet logic A, but the initially predicted water body proportion shows a weak signal (e.g., the water body proportion is greater than the second threshold of 20%), it is further determined whether the elevation meets the above-mentioned global geographic feature constraints and the slope meets the water body requirements. If the average slope is gentle (≤15°) and the average elevation strictly falls within the effective water body elevation range, the elevation consistency principle is used to forcibly recall and mark it as a water body. DEM perception optimization is achieved through weak signal recall based on DEM perception.

[0086] Step 7: The final output is a high-precision flood inundation map optimized by DEM perception.

[0087] In this step, based on the aforementioned judgment results, a final binarized flood inundation map is generated, and morphological processing is performed to remove fine noise, resulting in a high-precision water body extraction result with clear boundaries and uniform internal structure.

[0088] This concludes the process.

[0089] Example

[0090] This example selects a flood-affected area in a certain region for verification, using ground-measured water body range data as the ground truth to evaluate the accuracy of the extraction results. For example... Figure 5 and Figure 6 As shown, comparing single SAR data ( Figure 5 ) and CYGNSS signal-to-noise ratio reflectance and SAR fusion data based on random forest ( Figure 6The processing effect can be observed to be that... Figure 5 The original SAR image shown in (a) exhibits significant speckle noise and low ground feature contrast; while Figure 6 The fused data shown in (a) significantly suppresses background noise while preserving high-resolution spatial details, and enhances the separability between water and non-water bodies. (Observation) Figure 5 and Figure 6 The extraction results of subgraphs (e) K-means, (f) Otsu and (g) GMM in the figure show that the traditional algorithm is greatly affected by terrain and noise. The extracted water body boundaries are broken and there is obvious salt and pepper noise. In addition, there are still some misjudgments in the high slope area shown by the slope mask (red area in subgraph c).

[0091] Advantages of the method of this invention: Sub-image (h) shows the extraction results of the superpixel improved algorithm for DEM perception of this invention. Compared with traditional methods, the method of this invention utilizes DEM elevation and slope information (sub-images b and c) for effective global constraints, completely eliminating the interference of pseudo-water bodies caused by mountain shadows under complex terrain; at the same time, thanks to the local consistency constraints of superpixels, the extracted water body patches have good internal uniformity, and the connectivity and boundary smoothness of the river network system are significantly better than those of sub-images (e)-(g).

[0092] Quantitative evaluation results show that the overall accuracy of water body extraction by the method of the present invention reaches 95.8%, and the Kappa coefficient is 0.91. Compared with the single SAR data extraction method (overall accuracy 87.3%), the accuracy is significantly improved, which verifies the effectiveness and robustness of the method in flood monitoring in complex terrain. The extracted water body boundary has a high degree of agreement with the actual range.

[0093] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for water body extraction that integrates CYGNSS and SAR data, characterized in that, include: Step 1: Acquire CYGNSS and SAR data of the area to be water extracted; Calculate CYGNSS signal-to-noise ratio reflectance based on CYGNSS data; Extracting polarization features from SAR data; Obtain slope maps using digital elevation models (DEMs); Step 2: Perform data fusion of CYGNSS signal-to-noise ratio reflectance and polarization features based on the random forest model to obtain high-resolution CYGNSS signal-to-noise ratio reflectance that incorporates SAR data and has a resolution improved to the same level as SAR data; The specific data fusion method based on the random forest model is as follows: During the training phase, the latitude and longitude coordinates inherent in the CYGNSS data are used as sample locations. The signal-to-noise ratio reflectance is calculated from the CYGNSS data and then gridded to obtain the CYGNSS signal-to-noise ratio reflectance at the sample location. Polarization features with the same resolution as the CYGNSS data are extracted based on the sample location; these are called low-resolution polarization features. The CYGNSS signal-to-noise ratio reflectance at the sample location and the low-resolution polarization features constitute training samples to train a random forest model, enabling the random forest model to learn the mapping relationship between polarization features and CYGNSS signal-to-noise ratio reflectance. During the inference phase, the high-resolution polarization features extracted from the SAR data in step 1 are input into the random forest model to obtain the CYGNSS signal-to-noise ratio reflectance that fuses the SAR data and is of the same high resolution as the SAR data; this is the high-resolution CYGNSS signal-to-noise ratio reflectance. Step 3: Based on the high-resolution CYGNSS signal-to-noise ratio reflectance, water body extraction is performed to generate initial water body prediction results; Step 4: Based on the slope map, filter the areas with gentle slopes from the initial water body prediction results to obtain high-confidence water bodies; statistically analyze the elevation range of the high-confidence water bodies as a global geographic feature constraint; Step 5: Perform superpixel segmentation on the high-resolution CYGNSS signal-to-noise ratio reflectance, and determine the water proportion of each superpixel based on the initial water body prediction results; if the water proportion of a superpixel is greater than the first threshold and the slope meets the water body requirements, it is determined to be a water body area; then perform DEM-sensing weak signal recall: for superpixels that do not meet the first threshold, if the water proportion is greater than the second threshold, the second threshold is less than the first threshold, and the elevation meets the global geographic feature constraints, and the slope meets the water body requirements, then the superpixel is recalled and marked as a water body using the elevation consistency principle; finally, output a high-precision flood inundation map optimized by DEM-sensing weak signal recall.

2. The water body extraction method by fusing CYGNSS data and SAR data as described in claim 1, characterized in that, Step 1 further includes a preprocessing operation, including: Multidimensional constraint quality control was performed on the acquired CYGNSS data: antenna gain, incident angle, signal-to-noise ratio, and quality flag bits were read from the CYGNSS data; a physical parameter filter was constructed to remove invalid observation data with abnormal antenna gain, large incident angle, and low signal-to-noise ratio; a window for the peak position of the delay Doppler image was set to remove edge-truncated data; and bit masking technology was used to decode and filter the quality flag bits to remove noise data containing attitude instability, instrument calibration anomalies, and invalid specular reflection points. The acquired SAR data is subjected to radiometric correction, geometric correction, and denoising.

3. The water body extraction method by fusing CYGNSS data and SAR data as described in claim 1, characterized in that, In the random forest model, the number of decision trees is set to 100, the maximum tree depth is 10, and the minimum number of sample splits is 5. The training samples are divided into training set and validation set in a 7:3 ratio. The model parameters are optimized through the training set, and the model performance is evaluated through the validation set.

4. The water body extraction method by fusing CYGNSS data and SAR data as described in claim 1, characterized in that, Step 3 specifically includes: Based on the high-resolution CYGNSS signal-to-noise ratio reflectance obtained in step 2, water bodies are extracted using at least two classification algorithms. The water body extraction results of the two classification algorithms are then fused through a voting mechanism to generate initial water body prediction results.

5. The water body extraction method by fusing CYGNSS data and SAR data as described in claim 4, characterized in that, The step of using at least two classification algorithms to extract water bodies and fusing the water body extraction results of the two classification algorithms through a voting mechanism to generate an initial water body prediction result specifically includes: Three algorithms—K-means clustering, Otsu thresholding, and Gaussian mixture model (GMM)—are used for water body extraction. If a pixel is identified as a water body by at least two of the algorithms, then that pixel is included in the initial water body prediction result.

6. The water body extraction method by fusing CYGNSS data and SAR data as described in claim 1, characterized in that, The screening criteria for gently sloping areas in step 4 is: slope less than or equal to 5°.

7. The water body extraction method by fusing CYGNSS data and SAR data as described in claim 1, characterized in that, The step 4, which involves statistically analyzing the elevation range of the high-confidence water bodies and using it as a global geographic feature constraint, is as follows: statistically analyze the elevation distribution characteristics of the high-confidence water bodies and use the elevation range from the 5th percentile to the 95th percentile as the global geographic feature constraint.

8. The water body extraction method by fusing CYGNSS data and SAR data as described in claim 1, characterized in that, In step 5, the first threshold is 50% and the second threshold is 20%.

9. The water body extraction method by fusing CYGNSS data and SAR data as described in claim 1, characterized in that, In step 5, the criterion for determining whether the slope meets the requirements of the water body is: the average slope of the superpixel is less than 15°.

Images