Ultrahigh spatial resolution water depth inversion method fusing panchromatic spectral data

Abstract

The fusion full-color spectral data super-high spatial resolution water depth retrieval method comprises the following steps: collecting ICESat-2 satellite water depth data and satellite multispectral image and full-color image data; water depth retrieval is carried out on the original spatial resolution multispectral image to obtain the original spatial resolution water depth data; the multispectral image and the full-color image data are subjected to full-color spectral fusion processing to obtain super-high spatial resolution multispectral image; a machine learning algorithm is used for model training to construct a complex nonlinear mapping relationship between water depth and multispectral reflectivity; and the trained model is used for pixel-by-pixel prediction of the super-high spatial resolution multispectral image to obtain super-high spatial resolution water depth data. The nonlinear mapping relationship between the multispectral image spectral reflectivity and the water depth data of different spatial resolutions is constructed through the machine learning algorithm, and the problem of the decline of the retrieval precision caused by the loss of spectral information in the full-color spectral fusion process is effectively solved.

Description

TECHNICAL FIELD

[0001] The application relates to a fusion panchromatic spectral data super-high spatial resolution water depth inversion method and belongs to the field of marine surveying and mapping. BACKGROUND

[0002] As an important parameter of the marine environment, water depth is one of the basic elements of marine geographic space information, and is of great significance to maritime traffic, coastal engineering, marine resource development, and benthic environment mapping. Although the traditional shipborne sonar sounding technology has high precision, it is expensive and has limited coverage. With the rapid development of remote sensing technology, satellite optical remote water depth inversion technology has gradually become an important supplement to traditional sounding technology due to its advantages of large-scale synchronous observation and multi-temporal continuous observation. In particular, in recent years, the successful launch of satellites carrying active laser altimetry systems has promoted the development of active-passive fusion water depth inversion methods, providing a new way for large-scale and low-cost acquisition of shallow underwater topography. At present, water depth inversion models based on multispectral images mainly include three types: theoretical analytical models, semi-theoretical and semi-empirical models, and statistical models.

[0003] With the deepening of human development of the ocean, the demand for accurate measurement of seabed topography and geomorphology is increasing. Super-high spatial resolution sounding maps can accurately depict complex seabed topography, especially in areas with dramatic changes in water depth or complex seabed geomorphology, and can accurately capture rich seabed topographic details to achieve accurate expression of seabed micro-topography and geomorphology. However, existing water depth inversion mainly relies on multispectral images, which are limited by spatial resolution and cannot fully reveal the subtle undulations of seabed geomorphology. In contrast, panchromatic images have higher spatial resolution and can provide more detailed seabed topographic information. Through panchromatic spectral fusion technology, the spatial resolution of the fused image can be significantly improved while retaining the rich spectral information of the multispectral image, making the fused remote sensing image information more accurate and reliable for obtaining high-precision water depth data.

[0004] Although panchromatic spectral fusion technology has been widely used in land topographic mapping, the inevitable loss of spectral information during the fusion process limits its application in water depth inversion. Existing methods usually directly use the fused image for water depth inversion without fully considering the loss of spectral information during image fusion, especially for the fusion of different satellite images, where the loss of spectral information is more significant. Although the existing method directly uses the fused image for water depth inversion, it improves the spatial resolution but loses the inversion accuracy to some extent. Therefore, there is an urgent need to develop a super-high spatial resolution water depth inversion method that can effectively reduce the impact of spectral information loss. SUMMARY

[0005] The application aims to provide a fusion panchromatic spectral data super-high spatial resolution water depth inversion method, which can improve spatial resolution while effectively ensuring water depth inversion accuracy, and solve the problems of low efficiency and accuracy and high cost of super-high spatial resolution water depth data acquisition in current water depth inversion.

[0006] A fusion panchromatic spectral data super-high spatial resolution water depth inversion method, comprising the following steps:

[0007] Step S1: Collect ICESat-2 satellite water depth data, satellite multispectral image and panchromatic image data, and preprocess the obtained original data to ensure data consistency and availability.

[0008] Step S2: Spatially match the spectral reflectance of ICESat-2 water depth data and original spatial resolution multispectral image, construct a water depth inversion dataset, select a water depth inversion model, and perform water depth inversion on the original spatial resolution multispectral image to obtain original spatial resolution water depth data, providing reliable basic data for super-high spatial resolution water depth inversion.

[0009] Step S3: Perform panchromatic spectral fusion processing on the preprocessed multispectral image and panchromatic image data to obtain super-high spatial resolution multispectral image.

[0010] Step S4: Perform pixel matching between original spatial resolution water depth data and super-high spatial resolution multispectral image, extract the spectral reflectance of each band in the super-high spatial resolution multispectral image, and combine the water depth data at the corresponding position to construct a training sample set; based on the sample set, a machine learning algorithm is used for model training to construct a complex nonlinear mapping relationship between water depth and multispectral reflectance.

[0011] In the training process, cross-validation and hyperparameter optimization are used to ensure the generalization ability of the model.

[0012] Step S5: Use the trained model to perform pixel-by-pixel prediction on the super-high spatial resolution multispectral image to finally obtain super-high spatial resolution water depth data.

[0013] The fusion full-color spectral data super-high spatial resolution water depth inversion method provided by the application solves the problem of decreased inversion accuracy caused by loss of spectral information in the full-color spectral fusion process through a machine learning algorithm to construct a nonlinear mapping relationship between the multispectral image spectral reflectance and the water depth data of different spatial resolutions, thereby improving the spatial resolution while ensuring the inversion accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is a flowchart of the fusion full-color data super-high spatial resolution water depth inversion method described in the application.

[0015] Figure 2 is a schematic diagram of the experimental area of the embodiment of the application, wherein (a) is an image of the satellite of the experimental area, and the red line is the distribution of the ICESat-2 measuring line, and (b) is the distribution of the verification points in the local area in (a).

[0016] Figure 3 is a result example diagram (water depth raster diagram) obtained by inversion in the Ganquan Island area in the embodiment of the application, wherein (a) is the inversion of the original multispectral image, (b) is the direct inversion of the full-color spectral fusion super-resolution image, and (c) is the inversion of the full-color spectral fusion super-resolution image combined with the random forest.

[0017] Figure 4 is an enlarged diagram of the local details of the water depth results obtained by inversion in the embodiment of the application, wherein the first line is the northwest edge area of the reef disc, the second line is the northeast edge area of the reef disc, and the third line is the reef rock protruding area in the southern research area. DETAILED DESCRIPTION

[0018] The technical solutions of the application will be described clearly and completely in combination with the drawings.

[0019] The application provides a fusion full-color spectral data super-high spatial resolution water depth inversion method to obtain super-high spatial resolution shallow sea water depth data.

[0020] In order to more accurately evaluate the effectiveness of the fusion full-color spectral data super-high spatial resolution water depth inversion method, the application selects the sea area around the Ganquan Island as the research area for experiment, and the range of the research area is shown in Figure 2Airborne LiDAR bathymetric data was collected simultaneously for accuracy assessment. The accuracy of the bathymetric data obtained by the present invention was compared with that obtained by direct inversion using panchromatic image fusion, further demonstrating the effectiveness of the present method.

[0021] Reference Figure 1 , this method comprises the following steps:

[0022] Step S1: Collect and process ICESat-2 satellite water depth data and GF-2 satellite multispectral and panchromatic image data. For validation purposes, also collect on-site Optech Aquarius airborne LiDAR water depth measurement data. Specifically:

[0023] The acquired GF-2 satellite 4m spatial resolution multispectral images were subjected to data preprocessing such as radiometric calibration, atmospheric correction, orthorectification, and flare correction to reduce the impact of various environmental factors including the atmosphere and seawater on the image brightness value, highlight the water body information, and ensure the accuracy of the inversion results.

[0024] Radiometric calibration, orthorectification, and image registration are performed on the GF-2 satellite 1m spatial resolution panchromatic image to facilitate subsequent panchromatic spectral fusion processing.

[0025] ICESat-2 data was preprocessed by point cloud denoising, refraction correction, and tide correction to obtain high-precision water depth data for calibrating the inversion model. ICESat-2 has too few survey lines that pass directly through the study area. This paper selected a total of 6 survey lines that pass through the Ganquan Island area and nearby islands and reefs. The survey line locations are shown in the figure. Figure 2 (a) Because the northeastern part of the GF-2 satellite imagery used for inversion is partially covered by clouds, the cloud-covered areas of the ICESat-2 data were removed to reduce the impact of cloud cover on the accuracy of the water depth inversion results.

[0026] The collected measured data were tidal corrected and unified with the water depth data extracted by ICESat-2 for subsequent accuracy evaluation. In order to improve the efficiency of accuracy evaluation, the measured data were thinned before use. 1334 points evenly distributed in the study area were randomly selected. The distribution positions are shown in Figure 2 (b).

[0027] Step S2: Use ICESat-2 water depth data and the spectral reflectance of 4m spatial resolution multispectral images to perform spatial matching, build a water depth inversion dataset, select the logarithmic ratio inversion model, perform water depth inversion on the 4m spatial resolution multispectral images, and obtain 4m spatial resolution water depth data. Specifically:

[0028] In this step, first, the pre-processed ICESat-2 photon point water depth data is spatially matched with the spectral reflectance of the 4m spatial resolution multispectral image to construct the water depth inversion dataset; then, the spectral reflectance of the blue and green bands with the highest water depth correlation in the image is selected as the input variable, the processed ICESat-2 photon point water depth data is selected as the dependent variable, and the logarithmic ratio water depth inversion model is used for water depth inversion to obtain 4m spatial resolution water depth data, which provides reliable basic data for subsequent super-high spatial resolution water depth inversion. The logarithmic ratio model can to some extent weaken the influence of seabed differences in the water depth inversion process, and improve the accuracy and reliability of the water depth inversion results. The model contains only two key parameters, avoiding the problem of numerous and complex parameters in the traditional water depth inversion process.

[0029] Step S3: Selecting a suitable algorithm to perform panchromatic spectral fusion processing on the 1m spatial resolution panchromatic data and the 4m spatial resolution multispectral data to obtain 1m spatial resolution multispectral image. Specifically:

[0030] In this step, the Gram-Schmidt fusion algorithm is used to fuse the 4m spatial resolution multispectral image of GF-2 satellite and the 1m spatial resolution panchromatic image into a 1m spatial resolution multispectral image. The quality and visual clarity of the remote sensing image fused by the Gram-Schmidt algorithm are obviously improved, the texture details are more abundant, the spectral characteristics of the original multispectral image are preserved to the greatest extent, and the loss of spectral information is reduced.

[0031] Step S4: Establishing a nonlinear mapping relationship between the 4m spatial resolution water depth data and the spectral reflectance of the 1m spatial resolution multispectral image, and then inversely obtaining 1m spatial resolution water depth data. Specifically:

[0032] In this step, due to the difference in spatial resolution, the 4m spatial resolution water depth data corresponds to multiple pixels of the 1m spatial resolution multispectral image. In the 4m range, the middle position of the 1m spatial resolution multispectral image is selected to match the water depth data, ensuring that each 4m spatial resolution water depth data corresponds to a 1m spatial resolution multispectral image pixel. The spectral reflectance of each band in the 1m spatial resolution multispectral image is extracted, and combined with the corresponding water depth data to construct a training sample set.

[0033] The random forest algorithm is used to establish a nonlinear mapping relationship between the spectral reflectance of 1m spatial resolution multispectral images and the water depth data of 4m spatial resolution. Based on the constructed training sample set, the feature data set composed of the spectral reflectance of the red light, green light, blue light and near-infrared four bands of the 1m spatial resolution multispectral image is taken as the input variable, and the water depth data is taken as the target variable, and the random forest algorithm is used to train the model and predict the water depth value, and finally the 1m ultra-high spatial resolution water depth data is obtained. The construction process of the random forest model is simple and efficient, and is suitable for solving nonlinear complex problems.

[0034] In the training process, the optimization of the model performance is realized by changing the number of decision trees and the number of random features of a single decision tree.

[0035] Step S5: using the trained model to perform per-pixel prediction on the 1m spatial resolution multispectral image, and finally obtaining the 1m spatial resolution water depth data through inversion.

[0036] Embodiment

[0037] The inversion effect of the present application is compared with the inversion effect of the existing method of directly using the fused image for water depth inversion, which can further intuitively illustrate the effectiveness of the method.

[0038] The water depth inversion results of the Ganquan Island region are shown in Figure 3 , including the 4m spatial resolution water depth grid map of the Ganquan Island region ( Figure 3 a), the 1m spatial resolution water depth grid map of the Ganquan Island region obtained by the traditional method ( Figure 3 b) and the 1m spatial resolution water depth grid map of the Ganquan Island region obtained by the method proposed in the present application ( Figure 3 c). Overall, the overall change trend of the 1m resolution water depth data obtained by different methods is basically consistent with that of the 4m resolution water depth data. In order to more accurately compare the description ability of the water depth data after the spatial resolution is improved, three local areas with obvious topographic features are selected in the study area for enlarged display, and the water depth inversion results are compared (see Figure 4 ). In the reef protruding area in the south of the study area, the full-color spectral fusion super-resolution method combined with the random forest proposed in the present application can more clearly show the distribution details of the reefs, and the inversion obtained reef boundary is clear, and the description of the seabed topographic details is more accurate.

[0039] In order to quantitatively compare the inversion effects of the image data after the spatial resolution is improved by different methods, the present application selects three indexes of determination coefficient R2, mean absolute error MAE and root mean square error RMSE to evaluate the inversion accuracy of several different methods, and the calculation formulas are as follows:

[0040] (1)

[0041] (2)

[0042] (3)

[0043] in, For the The measured water depth value at the test point, For the Inverted water depth value of the checkpoint; is the average water depth of the test points; n is the number of test points. 2 The larger the value, the larger the MAE and RMSE values, indicating that the method is more accurate and the fitting effect is better.

[0044] Table 1 Comparison of inversion accuracy after improving spatial resolution using different methods

[0045] Method Spatial resolution [R 2 ]] MAE / m RMSE / m Original multispectral image inversion 4m 0.93 1.26 1.52 Panchromatic spectral fusion super-resolution image direct inversion 1m 0.87 1.66 2.00 Panchromatic spectral fusion super-resolution image inversion combined with random forest 1m 0.96 1.18 1.40

[0046] Comparison results show that the most accurate method is the panchromatic spectral fusion super-resolution method combined with the random forest algorithm proposed in this paper. The water depth data with a spatial resolution of 1m obtained by this method has an R2 of 0.96, a MAE of 1.18m, and an RMSE of 1.40. By constructing a nonlinear mapping relationship between the spectral reflectance of multispectral images of different spatial resolutions before and after image fusion and the water depth data, this method not only fully utilizes the rich spectral information of the original multispectral image before fusion, effectively solving the problem of spectral information loss in image fusion, but also integrates the effective spectral data of the four bands of blue, green, red, and near-red based on the random forest algorithm. This improves the spatial resolution while ensuring the accuracy of the water depth data inversion. Compared with the direct use of the panchromatic spectral fusion super-resolution method, the water depth data inverted by the proposed method has a MAE of 0.48m and an RMSE of 0.60m.

Claims

1. An ultra-high spatial resolution water depth inversion method integrating panchromatic spectral data, characterized by The following steps are involved: Step S1: Collect ICESat-2 satellite water depth data and satellite multispectral image and panchromatic image data, and pre-process the acquired raw data to ensure data consistency and availability; Step S2: Spatially match the spectral reflectance of ICESat-2 water depth data with the original spatial resolution multispectral image, construct a water depth inversion dataset, select a water depth inversion model, perform water depth inversion on the multispectral image at the original spatial resolution, and obtain water depth data at the original spatial resolution, providing reliable basic data for ultra-high spatial resolution water depth inversion; Step S3: performing panchromatic spectral fusion processing on the pre-processed multispectral image and panchromatic image data to obtain an ultra-high spatial resolution multispectral image; Step S4: Pixel matching is performed between the water depth data of the original spatial resolution and the ultra-high spatial resolution multispectral image, and the spectral reflectance of each band in the ultra-high spatial resolution multispectral image is extracted. The training sample set is then constructed by combining the water depth data of the corresponding locations. Based on this sample set, a machine learning algorithm is used to train the model and construct a complex nonlinear mapping relationship between water depth and multispectral reflectance. Step S5: Use the trained model to perform pixel-by-pixel prediction on the ultra-high spatial resolution multispectral image, and finally invert to obtain ultra-high spatial resolution water depth data.

2. The method for ultra-high spatial resolution water depth inversion by fusing panchromatic spectral data according to claim 1, characterized in that The satellite multispectral image and panchromatic image data adopt GF-2 satellite multispectral image and panchromatic image data.

3. The ultra-high spatial resolution water depth inversion method for fusing panchromatic spectral data according to claim 1 is characterized in that The ICESat-2 data are preprocessed, including point cloud denoising, refraction correction, and tide correction, to obtain high-precision water depth data for calibrating the inversion model.

4. The ultra-high spatial resolution water depth inversion method integrating panchromatic spectral data according to claim 2, The method is characterized in that the satellite multispectral image and panchromatic image data are preprocessed as described in step 1, including performing radiometric calibration, atmospheric correction, orthorectification and flare correction on the acquired GF-2 satellite 4m spatial resolution multispectral image; Radiometric calibration, orthorectification and image registration were performed on the GF-2 satellite 1m spatial resolution panchromatic image.

5. The method for ultra-high spatial resolution water depth inversion by fusing panchromatic spectral data according to claim 2, characterized in that The step 2 is specifically as follows: First, the pre-processed ICESat-2 photon point water depth data were spatially matched with the spectral reflectance of the 4-meter spatial resolution multispectral image to construct a water depth inversion dataset. Subsequently, the spectral reflectance of the blue and green bands in the image, which have the highest correlation with water depth, was selected as the input variables, and the processed ICESat-2 photon point water depth data was used as the dependent variable. The logarithmic ratio water depth inversion model was used to perform water depth inversion, and the water depth data with 4-meter spatial resolution was obtained, thus providing reliable basic data for subsequent ultra-high spatial resolution water depth inversion.

6. The method for ultra-high spatial resolution water depth inversion by fusing panchromatic spectral data according to claim 2, characterized in that In step 3, the Gram-Schmidt fusion algorithm is used to fuse the GF-2 satellite multispectral image with a spatial resolution of 4 meters and the panchromatic image with a spatial resolution of 1 meter into a multispectral image with a spatial resolution of 1 meter.

7. The method for ultra-high spatial resolution water depth inversion by fusing panchromatic spectral data according to claim 2, characterized in that In step S4, the training sample set is constructed as follows: To address the problem that the difference in spatial resolution will cause 4m spatial resolution water depth data to correspond to multiple pixels of 1m spatial resolution multispectral image, the pixel of 1m spatial resolution multispectral image in the middle position within the 4m range is selected to match the water depth data, ensuring that each 4m spatial resolution water depth data corresponds to a pixel of 1m spatial resolution multispectral image; the spectral reflectance of each band in the 1m spatial resolution multispectral image is extracted, and combined with the corresponding water depth data, a training sample set is constructed.

8. The method for ultra-high spatial resolution water depth inversion by fusing panchromatic spectral data according to claim 7, characterized in that In step S4, a nonlinear mapping relationship is established between the water depth data with a spatial resolution of 4 m and the spectral reflectance of the multispectral image with a spatial resolution of 1 m, and then the water depth data with a spatial resolution of 1 m is obtained by inversion, as follows: Based on the constructed training sample set, the feature data set composed of the spectral reflectance of the four bands of red, green, blue and near-infrared of the 1m spatial resolution multispectral image is used as the input variable, and the water depth data is used as the target variable. A nonlinear mapping relationship is established between the spectral reflectance of the 1m spatial resolution multispectral image and the 4m spatial resolution water depth data. The random forest algorithm is used to train the model and predict the water depth value, and finally the water depth data with 1m ultra-high spatial resolution is obtained.

9. The method for ultra-high spatial resolution water depth inversion by fusing panchromatic spectral data according to claim 2, characterized in that In step S5, the trained model is used to perform pixel-by-pixel prediction on the multispectral image with a spatial resolution of 1 m, and finally the water depth data with a spatial resolution of 1 m is inverted.

10. The method for ultra-high spatial resolution water depth inversion by fusing panchromatic spectral data as claimed in claim 1, wherein in step S4, a machine learning algorithm is used to perform model training. During the model training process, cross-validation and hyperparameter optimization are used to ensure the generalization ability of the model.

Images