Water depth retrieval method based on residual correction model for active-passive remote sensing fusion

Abstract

The present application relates to the technical field of remote sensing information extraction, and particularly relates to a water depth inversion method based on active and passive remote sensing fusion of residual correction model. The method comprises the following steps: S1, using normalized water index and remote sensing pixel threshold to perform land mask and deep water area mask on multi-spectral remote sensing image; S2, using band logarithmic ratio value model to perform initial water depth inversion on the image processed by step S1, to obtain initial water depth inversion result; S3, based on random forest residual correction model, fitting out the residual between the initial water depth result and the spaceborne laser depth sounding data, and using the residual to correct the initial result; S4, using root mean square error, mean absolute error, mean relative error and determination coefficient to perform precision evaluation on the water depth results before and after correction. The present application can make up for the problem that the traditional band logarithmic ratio value model is difficult to accurately estimate the water depth of shallow water and deep water, and improve the precision of water depth inversion.

Description

Technical Field

[0001] This invention relates to the field of remote sensing information extraction technology, specifically to a water depth inversion method based on active and passive remote sensing fusion using a residual correction model. Background Technology

[0002] Accurate water depth data provides crucial support for activities such as maritime navigation, underwater topography monitoring, and coastal ecosystem monitoring, playing a vital role in marine research. Traditional water depth measurement methods primarily rely on shipborne single-beam or multi-beam sonar and airborne lidar. While traditional sonar offers high accuracy, it is relatively time-consuming, labor-intensive, and expensive. Furthermore, shipborne sonar requires on-site operations, making it difficult for survey vessels to reach remote, unknown, or dangerous islands and reefs, or increasing the risk of grounding or running aground. Airborne lidar can measure inaccessible and dangerous areas, but this method is also costly, inefficient, and susceptible to weather and environmental conditions. Therefore, traditional methods yield limited water depth data, restricting the acquisition of large-scale, multi-island and reef water depth data. With the development of remote sensing technology, water depth inversion based on multispectral remote sensing imagery has become an effective method for rapidly and efficiently acquiring large-scale, high-resolution water depth data. Multispectral remote sensing water depth inversion primarily uses measured water depth as prior knowledge, combined with sea surface spectral information, to construct and train an inversion model. Currently, the acquisition of measured water depth mainly relies on traditional methods such as shipborne sonar and airborne lidar. However, the measured water depth data obtained by these traditional methods is relatively limited. Therefore, satellite bathymetry using multispectral remote sensing technology still faces many challenges, and its accuracy and reliability require further research and improvement.

[0003] The ICESat-2 satellite, launched successfully in September 2018, carries the advanced high-resolution topographic laser altimeter system ATLAS. Because ATLAS emits 532nm green laser pulses that can penetrate the water surface, ICESat-2 can provide precise depth sounding points wherever it flies. Combined with multispectral remote sensing satellite imagery, it can obtain depth maps of coastal areas, island and reef areas, and inland waters globally. Satellite depth sounding, combining active and passive remote sensing from the ICESat-2 satellite and the Sentinel-2 multispectral satellite, has become an important technique for water depth measurement. Compared to traditional shallow water depth measurement methods, it has significant advantages. Its characteristics of periodic acquisition of water depth data, low cost, and high-precision measurement are particularly suitable for exploring remote and unknown island and reef areas. Extensive research has been conducted by scholars both domestically and internationally on active and passive remote sensing water depth inversion. Among them, the band logarithmic ratio model is currently the most widely used model in active and passive remote sensing water depth inversion. The band logarithmic ratio model is based on the logarithmic ratio of the blue-green bands. It uses measured water depth data to construct the relationship between water depth and remote sensing reflectance, and then uses this relationship to invert water depth. For example, Chinese Patent Publication No. CN116295285A discloses a regionally adaptive shallow water depth remote sensing inversion method. It summarizes the water depth remote sensing inversion model into single-band model, dual-band model, multi-band model, band ratio model, logarithmic transformation ratio model, neural network model, and machine learning model to optimize the regional model and perform water depth inversion. However, due to the unevenness of seawater quality and the differences in seabed type, the band logarithmic ratio often overestimates the water depth in shallow water (<2m) and underestimates the water depth in deep water (>12m). Summary of the Invention

[0004] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide a water depth inversion method based on the fusion of active and passive remote sensing using a residual correction model. This method can make up for the problem that the traditional band logarithmic ratio model is difficult to accurately estimate the water depth of shallow and deep water areas, thereby improving the accuracy of water depth inversion.

[0005] The technical solution of this invention is as follows:

[0006] A water depth inversion method based on active and passive remote sensing fusion using a residual correction model includes the following steps:

[0007] S1. Use normalized water index and remote sensing pixel threshold to perform land masking and deep water masking on multispectral remote sensing images to eliminate the influence of irrelevant pixels on subsequent water depth inversion.

[0008] S2. Use the band logarithmic ratio model to perform initial water depth inversion on the masked image to obtain the initial water depth inversion results.

[0009] S3. Based on the random forest residual correction model, the residual between the initial water depth result and the spaceborne laser bathymetry data is fitted, and the residual is used to correct the initial result to achieve high-precision active and passive remote sensing fusion water depth inversion.

[0010] S4. Use root mean square error (RMSE), mean absolute error (MAE), mean relative error (MRE), and coefficient of determination (R²). 2 The accuracy of the water depth results before and after correction is evaluated.

[0011] Preferably, the formula for calculating the normalized water index in step S1 is as follows:

[0012]

[0013] Where R(λ) green ) and R(λ NIR () represent the surface reflectance in the green band and near-infrared band, respectively.

[0014] Preferably, the initial water depth inversion is performed using a band logarithmic ratio model. This model links the remote sensing reflectance of the blue and green bands with the prior water depth, derives the logarithmic transformation relationship of the ratio of the green band to the blue band, and establishes a linear model between the logarithmic ratio of the green band to the blue band and the prior water depth. The linear model is calculated using the following formula:

[0015] SDB=m1pSDB+m0 (2)

[0016]

[0017] Where m1 and m0 are model coefficients, SDB is the prior water depth of the model, pSDB is the logarithmic ratio model, n is a fixed constant set to 1500 to avoid negative values ​​when applying logarithms, and R0... λ,ii R represents the remote sensing reflectance in the blue band. λ,j Represents the remote sensing reflectance of the green band.

[0018] Preferably, the fitted log-ratio model is calculated using the following formula:

[0019] SDB = a * pSDB 2 +b*pSDB+c (4)

[0020] Where: a, b, and c are the model fitting coefficients calculated by the least squares method; the initial remote sensing inversion water depth is calculated using the above method.

[0021] Preferably, the initial inverted water depth is extracted using spaceborne laser water depth data, and the spaceborne laser water depth data is compared with the inverted water depth. The residual between the spaceborne laser water depth data and the inverted water depth is calculated using the following formula:

[0022] ε(x i ,y i )=D(x i ,y i )-SDB(x i ,y i (5)

[0023] Where: ε(x) i ,y i ) is in latitude and longitude (x i ,y i The residual between the spaceborne laser water depth and the remotely sensed inverted water depth at the location, D(x) i ,y i ) is in latitude and longitude (x i ,y i The spaceborne laser water depth at position SDB(x) i ,y i ) is in latitude and longitude (x i ,y i Remote sensing inversion of water depth at location.

[0024] Preferably, the step of using a random forest regression model to predict residuals at other locations includes the following steps:

[0025] S31. Use the residuals at known locations to extract the corresponding visible light bands and create a training dataset for model training.

[0026] S32. Apply the trained model to the entire remote sensing image to predict and obtain the residual field of the entire image.

[0027] S33. Use the predicted residual field to correct the initial water depth to obtain the corrected water depth.

[0028] Preferably, the residual correction model of the random forest is:

[0029] ε RF (r i ,c j )=f RF (R green (r i ,c j ),R red (r i ,c j ),R blue (r i ,c j (6)

[0030] Where f RF R represents a random forest regression model. green (ri ,c j ),R red (r i ,c j ),R blue (r i ,c j ) represent the reflectance of the green, red, and blue bands at row and column number (i,j), respectively.

[0031] Preferably, the correction formula in step S33 is:

[0032] Depth = SDB(r) i ,c j )+ε(r i ,c j (7)

[0033] Where Depth represents the water depth after residual correction, SDB(r i ,c j ) represents the water depth inverted by the empirical model at row and column number (i,j).

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] This invention proposes a residual correction model based on random forest, which fits the residual between the initial water depth result and the spaceborne laser bathymetry data, and uses the residual to correct the initial result, thereby achieving high-precision active and passive remote sensing fusion water depth inversion. This can make up for the problem that the traditional band logarithmic ratio model is difficult to accurately estimate the water depth of shallow and deep water areas, and improve the accuracy of water depth inversion. Moreover, this method is simple, reliable, accurate and easy to implement. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart illustrating the method of the present invention.

[0038] Figure 2 This is a schematic diagram of the result of masking remote sensing images in an embodiment of the present invention.

[0039] Figure 3 This is a schematic diagram of the result after residual correction in an embodiment of the present invention.

[0040] Figure 4 This is the accuracy evaluation result before and after residual correction in an embodiment of the present invention.

[0041] Figure 5 This is the result of the consistency analysis of residuals before and after correction in an embodiment of the present invention. Detailed Implementation

[0042] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0043] Example

[0044] Figure 1 This diagram illustrates the flow chart of the water depth inversion method based on the residual correction model. The following describes the method of the present invention in further detail for each step in the implementation process.

[0045] (1) Land / deep water masking: The normalized water index and pixel threshold are used to mask the multispectral remote sensing images to eliminate the influence of irrelevant pixels on the subsequent water depth inversion, so as to improve the efficiency of the work.

[0046] The Normalized Difference Water Index (NDWI) can highlight water information in images and can be used to separate water bodies from land. Formula (1) is the calculation formula for NDWI.

[0047]

[0048] Where R(λ) green ) and R(λ NIR The NDWI values ​​represent the surface reflectance in the green band and near-infrared band, respectively. Water bodies have an NDWI greater than 0, while land bodies have an NDWI less than 0; therefore, 0 is used as the threshold for separating land and water bodies. Areas with NDWI values ​​greater than the set threshold are classified as water bodies, and areas with NDWI values ​​less than the set threshold are classified as land bodies.

[0049] When the seabed is too deep, spaceborne laser satellites cannot receive signals returned from the seabed, thus each region has a maximum detectable depth. Therefore, after segmenting the land area, further masking processing is required for the deep water area. First, by statistically analyzing the pixel brightness values ​​of each band in the water area, the tenths of the brightness value is used as the initial threshold for deep water pixels. Next, a 3×3 moving window is transmitted to the image. When more than 50% of the pixels in the window are at or below the initial brightness threshold, the center pixel is designated as a deep water pixel and used to calculate the average deep water reflectance and standard deviation for each band. Finally, thresholding is applied to the blue and green bands. The threshold is the average deep water reflectance plus three times the standard deviation. If the reflectance of both the green and blue bands exceeds their respective thresholds, the pixel is identified as a shallow water pixel; otherwise, the pixel is identified as a deep water pixel. Figure 1 It is a remote sensing image that has been masked.

[0050] (2) Initial water depth inversion: Multispectral remote sensing water depth inversion is performed using the band logarithmic ratio model to obtain the initial water depth inversion results.

[0051] The band logarithmic ratio model is an empirical water depth model with exponentially decaying spectral characteristics. This model links the remote sensing reflectance of the blue and green bands with prior water depth, derives the logarithmic transformation relationship of the ratio of the green band to the blue band, and establishes a linear model of the logarithmic ratio of the green and blue bands with prior water depth. This model is calculated using formulas (2) and (3):

[0052] SDB=m1pSDB+m0 (2)

[0053]

[0054] Where m1 and m0 are model coefficients, SDB is the prior water depth of the model, pSDB is the band logarithmic ratio model, n is a fixed constant set to 1500 to avoid negative values ​​when applying logarithms, and R... λ,i R represents the remote sensing reflectance in the blue band. λ,j Represents the remote sensing reflectance of the green band.

[0055] By applying formula (3) pixel by pixel, a remote sensing image with a band logarithmic ratio is generated. Then, the satellite-borne laser satellite bathymetry control points are used as the seed depth for retrieving water depth for regression. The linear relationship between pSDB and water depth is maintained only within a certain depth range of water bodies, and then becomes nonlinear. Especially in deep water bodies, the model does not show good results. This invention uses a quadratic term (formula (4)) to fit the logarithmic ratio model:

[0056] SDB = a * pSDB 2 +b*pSDB+c (4)

[0057] Where a, b, and c are the model fitting coefficients calculated using the least squares method. The initial remote sensing inversion water depth can be calculated using the above method.

[0058] (3) Initial water depth result correction: The initial remote sensing inversion water depth result is corrected using a residual correction model to obtain more accurate and reliable water depth results.

[0059] To make the water depth retrieved by the model closer to the distribution of the real data, residual correction is performed based on the initial retrieved water depth. The initial retrieved water depth is extracted using spaceborne laser water depth data, and the spaceborne laser water depth data is compared with the retrieved water depth. The residual between the two is calculated using formula (5):

[0060] ε(x i ,y i )=D(x i ,y i )-SDB(x i ,y i (5)

[0061] Where: ε(x) i ,y i ) is in latitude and longitude (x i ,y i The residual between the spaceborne laser water depth and the remotely sensed inverted water depth at the location, D(x) i ,y i ) is in latitude and longitude (x i ,y i The spaceborne laser water depth at position SDB(x) i ,y i ) is in latitude and longitude (x i ,y i Remote sensing inversion of water depth at location.

[0062] Random forest regression models do not need to consider the actual physical characteristics of light propagation in water, and can quickly generate large-scale, high-resolution seabed topography using sufficient measured and spectral data. They have high accuracy in the field of multispectral bathymetry. This invention uses a random forest regression model to predict residuals at other locations. The visible light bands of the multispectral remote sensing image are used as features for model training, and formula (6) is the residual correction model of the constructed random forest. First, the corresponding visible light bands are extracted using the residuals at known locations to create a training dataset for model training; then, the trained model is applied to the entire remote sensing image to predict and obtain the residual field of the entire image; finally, the predicted residual field is used to correct the initial water depth (formula (7)) to obtain the corrected water depth.

[0063] ε RF (ri ,c j )=f RF (R green (r i ,c j ),R red (r i ,c j ),R blue (r i ,c j (6)

[0064] Depth = SDB(r) i ,c j )+ε(r i ,c j (7)

[0065] Where: ε(r) i ,c j ) represents the residual at row and column number (i,j), f RF R represents a random forest regression model. green (r i ,c j ),R red (r i ,c j ),R blue (r i ,c j ) represent the reflectance of the green, red, and blue bands at row and column number (i,j), respectively. Depth represents the water depth after residual correction, SDB(r i ,c j ) represents the water depth inverted by the empirical model at row and column number (i,j).

[0066] (4) Water depth inversion accuracy assessment: Combining traditional measured water depth data, the root mean square error, mean absolute error, mean relative error and coefficient of determination are used to assess the accuracy of the water depth before and after correction.

[0067] The traditional measured water depth data used in this embodiment is acquired by airborne lidar. The latitude and longitude information in the measured water depth data is used to extract the pixel values ​​in the remote sensing inversion water depth raster map for accuracy evaluation. Figure 4 , Figure 5 The results of the accuracy assessment and consistency analysis of the embodiments are presented respectively. Figure 4 In this embodiment, the RMSE is 1.33m, the MAE is 1.02m, and the MRE reaches 18.09%, all of which are higher than the accuracy without residual correction. Figure 5 Based on the consistency analysis results of this embodiment, it can be seen that the measured water depth data retrieved by the residual correction model are in good agreement, R 2A value greater than 0.9 can be used to determine the accurate water depth.

[0068] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the invention should also be covered within the protection scope of the invention. Therefore, the protection scope of the invention should be determined by the scope of the claims.

Claims

1. A water depth inversion method based on active and passive remote sensing fusion using a residual correction model, characterized in that, Includes the following steps: S1. Use normalized water index and remote sensing pixel threshold to perform land masking and deep water masking on multispectral remote sensing images to eliminate the influence of irrelevant pixels on subsequent water depth inversion. S2. Use the band logarithmic ratio model to perform initial water depth inversion on the image after masking in step S1 to obtain the initial water depth inversion result. S3. Based on the random forest residual correction model, the residual between the initial water depth result and the spaceborne laser bathymetry data is fitted, and the residual is used to correct the initial result to achieve high-precision active and passive remote sensing fusion water depth inversion. The initial water depth inversion result is extracted from spaceborne laser water depth data, and the spaceborne laser water depth data is compared with the inverted water depth. The residual between the spaceborne laser water depth data and the inverted water depth is calculated using the following formula: （5） in: In latitude and longitude ( The residual between the spaceborne laser water depth and the remotely sensed inverted water depth at the location. In latitude and longitude ( Spaceborne laser water depth at location ) In latitude and longitude ( Remote sensing inversion of water depth at location; The residual correction model for the random forest is as follows: （6） in: This represents a random forest regression model. They represent the row and column numbers respectively. i,j Reflectance of the green, red, and blue bands at point ( ); The residual correction model of the random forest for predicting residuals at other locations includes the following steps: S31. Use the residuals at known locations to extract the corresponding visible light bands and create a training dataset for model training. S32. Apply the trained model to the entire remote sensing image to predict and obtain the residual field of the entire image. S33. The initial water depth is corrected using the predicted residual field to obtain the corrected water depth. The correction formula is as follows: （7） in, Depth This indicates the water depth after residual correction. For row and column numbers ( i , j Empirical model inversion of water depth at location ) S4. Use root mean square error, mean absolute error, mean relative error, and coefficient of determination to evaluate the accuracy of the water depth results before and after correction.

2. The water depth inversion method based on active and passive remote sensing fusion using a residual correction model as described in claim 1, characterized in that, The formula for calculating the normalized water index in step S1 is as follows: （1） in, and These represent the surface reflectance in the green band and near-infrared band, respectively.

3. The water depth inversion method based on active and passive remote sensing fusion using a residual correction model as described in claim 1, characterized in that, The initial water depth inversion is performed using a band logarithmic ratio model, which links the remote sensing reflectance of the blue and green bands with the prior water depth. The model derives the logarithmic transformation relationship between the green and blue band ratios and establishes a linear model between the green and blue band logarithmic ratios and the prior water depth. This linear model is calculated using the following formula: （2） （3） in, , These are the model coefficients. The prior water depth for the model, For log-ratio models, n This is a fixed constant, set to 1500, to avoid negative numbers when applying logarithms. Represents remote sensing reflectance in the blue band. Represents the remote sensing reflectance of the green band.

4. The water depth inversion method based on active and passive remote sensing fusion using a residual correction model as described in claim 3, characterized in that, Use the following formula to fit the log-ratio model: （4） in: a , b , c The model fitting coefficients are calculated using the least squares method. The initial remote sensing inversion water depth is calculated using formula (4).

Images