Water depth inversion method based on spectrum and geographic space information

Abstract

The invention provides a water depth inversion method based on spectrum and geospatial information, which relates to the technical field of water depth inversion and comprises the following steps: acquiring a satellite high-resolution spectral image of a target position, and preprocessing the satellite high-resolution spectral image; extracting remote sensing reflectivity data from the preprocessed satellite high-resolution spectral image; constructing a prediction data set based on the remote sensing reflectivity data and the geographic space information of the target position; and inputting the prediction data set into the random forest regression model for processing to obtain water depth data of the target position. According to the method, the geospatial information and the spectral data are combined, so that the model can capture the spatial mode of the water depth instead of only depending on the color or transparency of water, and therefore, the method provided by the invention can obtain better water depth surveying and mapping accuracy in a turbid and complex watershed environment compared with the prior art; and reliable data support is provided for hydrology, landform and river management research.

Description

Technical Field

[0001] The present invention relates to the technical field of water depth inversion, and in particular to a water depth inversion method based on spectrum and geographic space information. Background Art

[0002] Accurate bathymetric data is crucial for a wide range of applications, including navigation safety, coastal management, marine habitat mapping, and hydrodynamic modeling. Traditional field measurements, while accurate, are time-consuming, costly, and have limited spatial coverage, particularly in dynamic, remote environments. Bathymetric remote sensing technology offers a more convenient alternative, offering wide spatial coverage, frequent updates, and a cost-effective solution for monitoring shallow water depths. However, for turbid and complex watersheds, particularly at river confluences, existing bathymetric remote sensing technologies estimate depth based solely on spectral analysis, failing to capture the true underwater topography at river confluences and thus compromising accuracy. Summary of the Invention

[0003] The purpose of the present invention is to provide a water depth inversion method based on spectral and geospatial information to improve the accuracy of water depth mapping in turbid and complex watershed environments.

[0004] To achieve the above-mentioned purpose, the present invention provides the following technical solutions:

[0005] A water depth inversion method based on spectral and geospatial information, the method comprising:

[0006] Acquire a satellite high-resolution spectral image of the target location and preprocess the satellite high-resolution spectral image;

[0007] Extract remote sensing reflectance data from pre-processed satellite high-resolution spectral images;

[0008] Construct a prediction dataset based on remote sensing reflectance data and geospatial information of target locations;

[0009] The predicted data set is input into the random forest regression model for processing to obtain the water depth data of the target location.

[0010] Furthermore, the satellite high-resolution spectral images are preprocessed, including atmospheric correction, glare correction, and geometric correction.

[0011] Furthermore, before the prediction dataset is input into the random forest regression model for processing, the random forest regression model is trained, specifically including the following operations:

[0012] Extract remote sensing reflectance data for multiple locations from pre-processed satellite high-resolution spectral images;

[0013] Collect water depth data at corresponding locations, and associate remote sensing reflectivity data with water depth data at corresponding locations to form a comprehensive data set;

[0014] A random forest regression model was trained using the synthetic dataset.

[0015] Furthermore, the comprehensive dataset is randomly divided into 80% as a training set and the other 20% as a test set, and the data in the training set is standardized.

[0016] Furthermore, a 4×4 hyperparameter grid was defined using the random search class in the scikit-learn library, and 20 random search iterations were performed with 5-fold cross validation to optimize the hyperparameters of the random forest regression model.

[0017] Furthermore, a prediction dataset is constructed based on remote sensing reflectivity data and geospatial information of the target location, specifically including:

[0018] Extracting longitude and latitude information from the geospatial information of the target location and binding the longitude and latitude information with the remote sensing reflectance data of the corresponding location;

[0019] Perform feature engineering process based on remote sensing reflectance and geospatial information.

[0020] Furthermore, the feature engineering process based on remote sensing reflectivity and geospatial information specifically includes the following operations:

[0021] The normalized difference water index was calculated using the formula: (Band_3-Band_4) / (Band_3+Band_4);

[0022] Calculate the first band ratio Band_1 / Band_2 and the second band ratio Band_3 / Band_4;

[0023] Calculation of geographic features, including longitude-latitude interaction calculations and origin distance calculations;

[0024] Spectral combination calculation, including band sum, band average, and band standard deviation;

[0025] Nonlinear transformation, creating two additional features for each band through logarithmic and square transformations;

[0026] Among them, Band_1, Band_2, Band_3, and Band_4 are the spectral information of the four bands of the satellite high-resolution spectral image.

[0027] Furthermore, after obtaining the water depth data of the target location, the water depth data is tidal corrected, and the corrected water depth data is converted into an image format to draw a comprehensive water depth map of the target location.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention provides a water depth inversion method based on spectral and geospatial information. By combining geospatial information with spectral data from high-resolution satellite spectral images and processing them through a random forest regression model, the model can capture the spatial pattern of water depth, rather than relying solely on the color or transparency of the water. This enables the method provided by the present invention to achieve better water depth mapping accuracy than existing technologies in turbid and complex watershed environments, providing reliable data support for hydrological, geomorphological and river management research. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0031] Figure 1 It is a schematic diagram of the overall structure of a water depth inversion method based on spectral and geospatial information provided by an embodiment of the present invention.

[0032] Figure 2 This is a scatter diagram of the results of water depth measurement using the Stumpf model.

[0033] Figure 3 This is a scatter diagram of the results of water depth measurement using a log-linear model.

[0034] Figure 4 This is a scatter diagram of the results of water depth measurement using the random forest model.

[0035] Figure 5 It is a scatter diagram of the results of water depth measurement using the method provided by the present invention.

[0036] Figure 6 It is a water depth map of the target study area drawn using the Stumpf model.

[0037] Figure 7 It is a bathymetric map of the target study area drawn using a log-linear model.

[0038] Figure 8 It is a water depth map of the target study area drawn using the random forest model.

[0039] Figure 9 It is a water depth map of the target research area drawn using the method provided by the present invention. DETAILED DESCRIPTION

[0040] The principles and features of the present invention are described below with reference to the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0041] Reference Figure 1 This embodiment provides a water depth inversion method based on spectral and geospatial information, the method comprising the following steps:

[0042] S101: Acquire a satellite high-resolution spectral image of a target location and preprocess the satellite high-resolution spectral image.

[0043] For example, the GF-1WFV1 satellite sensor can be used to obtain high-resolution satellite spectral imagery. This high-resolution satellite spectral imagery provides detailed spectral information, essential for accurate bathymetric mapping. The GF-1 satellite's wide-field-of-view (WFV) sensor offers a balance of spatial and spectral resolution, making it particularly well-suited for monitoring large river systems and their complex underwater topography.

[0044] To improve measurement accuracy, this embodiment preprocesses the satellite's high-resolution spectral imagery, including: atmospheric correction using the FLAASH (Fast Line-of-Sight Atmospheric Analysis of Hypercube) module to mitigate atmospheric effects and obtain more accurate water surface reflectance values; glare correction using a flare removal algorithm to address the effect of solar glare on the water surface, thereby further improving the quality of the bathymetric data; and geometric correction to ensure accurate spatial alignment of the images.

[0045] S102. Extract remote sensing reflectance data from the preprocessed satellite high-resolution spectral image.

[0046] For example, ENVI software can be used to analyze satellite high-resolution spectral images to obtain remote sensing reflectance data.

[0047] S103: Construct a prediction data set based on the remote sensing reflectance data and the geospatial information of the target location.

[0048] S104: Input the predicted data set into a random forest regression model for processing to obtain water depth data of the target location.

[0049] Before inputting the prediction set data into the random forest regression model for processing, the random forest regression model needs to be trained, which includes the following operations:

[0050] S201. Extract remote sensing reflectance data of multiple locations from preprocessed satellite high-resolution spectral images.

[0051] S202: Collect water depth data at corresponding locations, and associate the remote sensing reflectivity data with the water depth data at corresponding locations to form a comprehensive data set.

[0052] S203. Use the comprehensive dataset to train the random forest regression model.

[0053] To ensure reproducibility, the synthetic dataset is randomly partitioned using a fixed random seed, with 80% of the data set used as a training set and the remaining 20% ​​used as a test set. The data in the training set is then normalized. For example, the normalization process can be implemented using the Standard Scaler in the scikit-learn library.

[0054] As a possible implementation, to optimize model performance, we used the Random Search class in the scikit-learn library to define a 4×4 hyperparameter grid. We then performed 20 random search iterations with 5-fold cross-validation to optimize the hyperparameters of the random forest regression model. The hyperparameter grid is shown in Table 1.

[0055] Table 1 Hyperparameter grid

[0056] Hyperparameters value n_estimators [100,200,300,400] Maximum depth [10,20,30,none] Minimum Sample Split [2,5,10,15] Minimum sample leaf [1,2,4,8]

[0057] Construct a prediction dataset based on remote sensing reflectance data and geospatial information of the target location, including:

[0058] S301 , extracting longitude and latitude information from the geographic space information of the target location, and binding the longitude and latitude information with remote sensing reflectivity data of the corresponding location.

[0059] S302: Execute a feature engineering process based on remote sensing reflectivity and geospatial information.

[0060] Step S302 specifically includes the following operations:

[0061] (1) Calculate the normalized difference water index (NDWI) using the formula: (Band_3-Band_4) / (Band_3+Band_4). NDWI is sensitive to water content and is related to water depth.

[0062] (2) Calculate the first band ratio Band_1 / Band_2 and the second band ratio Band_3 / Band_4. The first band ratio and the second band ratio help normalize the atmospheric effects and highlight specific spectral characteristics related to water depth.

[0063] (3) Calculation of geographical features, including longitude-latitude interaction calculation and origin distance calculation.

[0064] Among them, the longitude interaction calculation is calculated as longitude × latitude, which can capture the potential depth changes related to geographical location.

[0065] The calculation formula for the origin distance is: sqrt(longitude 2 +Latitude 2 ), which represents the distance from an arbitrary origin and can capture trends in water depth related to distance from the coast or other geographic features.

[0066] (4) Spectral combination calculation, including band sum, band average, and band standard deviation.

[0067] Among them, the band synthesis is the sum of all bands, that is, (Band_1+Band_2+Band_3+Band_4);

[0068] The band average is the average of all bands;

[0069] Band standard deviation is the standard deviation of all bands and is used to capture spectral variations.

[0070] (5) Nonlinear transformation, creating two additional features for each band through logarithmic transformation and square transformation.

[0071] The calculation formula for logarithmic transformation is: log(1+Band_n); the calculation formula for square transformation is: Band_n 2 The nonlinear transformation can capture the potential nonlinear relationship between spectral reflectance and water depth.

[0072] Band_1, Band_2, Band_3, and Band_4 are spectral information of four bands of the satellite high-resolution spectral image. This embodiment captures the complex relationship between spectral characteristics, geographical location, and water depth by performing the above feature engineering process.

[0073] After obtaining the water depth data at the target location, the water depth data is tidally corrected to account for water level changes during satellite image acquisition and field measurements. The corrected water depth data is converted into an image format and a comprehensive water depth map of the target location is drawn.

[0074] We collected 2,000 in-situ depth points using an RDI 600kHz ADCP River Ray, providing a reliable ground-truth dataset for model training and validation. The water depth estimates from the random forest regression model were compared with the ground-truth dataset to verify the accuracy of the model estimates.

[0075] In a specific implementation, the method provided in this embodiment was applied to the complex aquatic environment at the confluence of the Yellow River and the Yiluo River to estimate water depth along with three other traditional bathymetric estimation models, including the Stumpf model, the log-linear method, and the random forest model.

[0076] The Stumpf model and log-linear model, which represent traditional bathymetric mapping methods, performed poorly in the study area, e.g. Figure 2 and Figure 3 As shown, the coefficient of determination of these two models is low - R 2 The predicted depths are 0.01 and 0.00, respectively, with high error rates—MAE of 1.20 meters and 1.19 meters, respectively; and RMSE of 1.51 meters for both. The scatter plots for these models show a clear deviation between the predicted depths and the actual depths. Regardless of the actual depth, the predicted depths are concentrated in a narrow range. Figure 6 and Figure 7 This poor performance is further demonstrated by the fact that the bathymetric maps produced by these two models lack details and fail to capture the true underwater topography at the confluence of the rivers.

[0077] Compared with traditional bathymetric estimation models, random forest models have made significant improvements, such as Figure 4 As shown, its determination coefficient (R 2 =0.75) and a lower error rate (MAE = 0.50 m, RMSE = 0.76 m). The scatter plot shows that the correlation between the predicted depth and the actual depth is stronger, but there is still some obvious dispersion, especially at larger depths. Figure 8 The bathymetric maps reflect this improvement, showing a more detailed and diverse underwater terrain.

[0078] By comparison, it can be seen that the method provided in this embodiment has achieved the most significant results. Figure 5 The results show the excellent performance of the method. In all models, its coefficient of determination (R 2 =0.92) is the highest, and the error rate (MAE = 0.28m, RMSE = 0.45m) is the lowest. Figure 9 shows a tight clustering of points along the 1:1 line, indicating high accuracy across the entire depth range. Figure 9 As shown, this exceptional performance translates into the most detailed and nuanced bathymetric maps. Clearly, the method provided by this embodiment can capture subtle changes in riverbed topography, clearly outlining river channels, shoals, and other underwater features that are crucial for understanding river dynamics and navigation.

[0079] The superior performance of the method provided in this embodiment is attributed to its combination of geospatial information with spectral data. This approach enables the model to capture spatial patterns of water depth that are not solely dependent on water color or transparency, which is particularly beneficial in turbid and complex river systems like the Yellow River.

[0080] Table 2 Comparison of RMSE errors for different water depths and different bathymetric methods

[0081]

[0082]

[0083] In our implementation, we conducted a comprehensive evaluation of four bathymetric estimation models at different depth ranges at the confluence of the Yellow and Yiluo rivers. Table 2 provides a detailed comparison of the RMSE of each model at different depth ranges.

[0084] In shallow waters (0-2 meters), accounting for 567 measurement points, the method provided by this example demonstrates superior performance, followed by the random forest model. The traditional Stumpf model and log-linear model exhibited significant errors. This demonstrates that machine learning methods, particularly those incorporating geospatial information, can more effectively capture the complexities of shallow river bathymetry.

[0085] For the mid-depth range (2-4 meters), which includes 970 measurement points, the performance of all models improved. Specifically, the method provided in this embodiment achieved the lowest RMSE, maintaining its leading position and significantly outperforming the other methods. The performance of the random forest model followed closely behind. The performance of the Stumpf and log-linear models was comparable. This trend highlights the consistent advantage of machine learning methods in this critical depth range.

[0086] Similar results were observed in the 4-6 meter depth range. The method provided in this example again demonstrated the best performance, followed by the random forest model. However, the accuracy of traditional bathymetric estimation models decreased significantly in this depth range, with the Stumpf model achieving a slightly better RMSE than the log-linear model.

[0087] For depths greater than 6 meters, the errors of all models increased, likely due to the challenges of remote sensing deeper waters. The performance gap between models also widened significantly. The method provided in this example maintained relatively high accuracy, while the errors of the other models increased significantly. The random forest model still outperformed the traditional bathymetric estimation model, while the Stumpf and log-linear models had the largest errors.

[0088] Overall, among all 2000 measurement points, the RMSE value of the method provided in this embodiment is 0.45 meters, which always maintains excellent performance; followed by the random forest model; and the overall performance of the Stumpf and log-linear models is the same. It can be seen that in all depth ranges, the performance of the method provided in this embodiment is always better than other methods, highlighting the effectiveness of combining geospatial information with spectral data in bathymetric mapping. Compared with traditional models, machine learning methods are more adaptable to different depth conditions. The performance gap between machine learning and traditional models widens with increasing depth, especially in areas with a water depth of more than 4 meters. The performance gap of traditional models is smaller, but it is difficult to work in extremely shallow (<2 meters) and deeper (>4 meters) waters.

[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A water depth inversion method based on spectral and geospatial information, characterized in that: The method comprises: Acquire a satellite high-resolution spectral image of the target location and preprocess the satellite high-resolution spectral image; Extract remote sensing reflectance data from pre-processed satellite high-resolution spectral images; Construct a prediction dataset based on remote sensing reflectance data and geospatial information of target locations; The predicted data set is input into the random forest regression model for processing to obtain the water depth data of the target location.

2. The method for water depth inversion based on spectral and geographic spatial information according to claim 1, characterized in that: Preprocessing of satellite high-resolution spectral images, including atmospheric correction, glare correction, and geometric correction.

3. The method for water depth inversion based on spectral and geographic spatial information according to claim 1, characterized in that: Before inputting the prediction dataset into the random forest regression model for processing, the random forest regression model is trained, which includes the following operations: Extract remote sensing reflectance data for multiple locations from pre-processed satellite high-resolution spectral images; Collect water depth data at corresponding locations, and associate remote sensing reflectivity data with water depth data at corresponding locations to form a comprehensive data set; A random forest regression model was trained using the synthetic dataset.

4. The method for water depth inversion based on spectral and geographic spatial information according to claim 3, characterized in that: The comprehensive dataset is randomly divided into 80% as the training set and the other 20% as the test set, and the data in the training set is standardized.

5. The method for water depth inversion based on spectral and geographic spatial information according to claim 1, characterized in that: The random search class in the scikit-learn library defines a 4×4 hyperparameter grid and performs 20 random search iterations with 5-fold cross validation to optimize the hyperparameters of the random forest regression model.

6. The method for water depth inversion based on spectral and geographic spatial information according to claim 1, characterized in that: Construct a prediction dataset based on remote sensing reflectance data and geospatial information of the target location, including: Extracting longitude and latitude information from the geospatial information of the target location and binding the longitude and latitude information with the remote sensing reflectance data of the corresponding location; Perform feature engineering process based on remote sensing reflectance and geospatial information.

7. The method for water depth inversion based on spectral and geographic spatial information according to claim 6, characterized in that: The feature engineering process based on remote sensing reflectivity and geospatial information specifically includes the following operations: The normalized difference water index was calculated using the formula: (Band_3-Band_4) / (Band_3+Band_4); Calculate the first band ratio Band_1 / Band_2 and the second band ratio Band_3 / Band_4; Calculation of geographic features, including longitude-latitude interaction calculations and origin distance calculations; Spectral combination calculation, including band sum, band average, and band standard deviation; Nonlinear transformation, creating two additional features for each band through logarithmic and square transformations; Among them, Band_1, Band_2, Band_3, and Band_4 are the spectral information of the four bands of the satellite high-resolution spectral image.

8. The method for water depth inversion based on spectral and geographic spatial information according to claim 1, characterized in that: After obtaining the water depth data of the target location, the water depth data is tidally corrected, and the corrected water depth data is converted into an image format to draw a comprehensive water depth map of the target location.

Images