Redai dili (Sep 2023)
Assessment of Multi-Spectral Imagery and Machine Learning Algorithms for Shallow Water Bathymetry Inversion
Abstract
The eastern coastal zone of Hainan Island is an important scenic belt and concentrated area of tourism resources in Hainan Province. Due to natural factors such as sea level rise and human factors such as coastal reclamation, the ecosystem in this area is highly sensitive. Water depth data are crucial for the protection and management of the coastal ecosystem. Satellite remote sensing data combined with machine learning algorithms have become an important means of shallow water depth inversion. However, few studies evaluate shallow water depth inversion for different remote sensing data, water environmental conditions, and algorithms. Taking the Wanning Sea area as an example, three scenes of Landsat-8 and Sentinel-2 data with different water environmental conditions were selected to apply water depth inversion. The Sentinel-2 data were collected on February 11th, 2022, with some suspended matter and poor water transparency in the nearshore water, and the image exhibited band-like reflectance anomalies caused by waves. The Landsat-8 data were collected on July 16, 2019, and June 28th, 2021. Both scenes had better water transparency than the Sentinel-2 data, and relatively less suspended matter in the nearshore water. Compared to the Landsat-8 data collected on June 28th, 2021, the image of Landsat-8 data collected on July 16th, 2019, showed stronger wave features in the nearshore water. A 1:25,000 maritime chart of the Potou Port and Dazhou Island (C1516171) area released by the China Maritime Safety Administration was collected to obtain 588 measured water depth data points in the study area. Among these, 295 randomly selected data points were used as training data for the remote sensing water depth inversion, and the remaining 293 data points were used as testing data to evaluate the accuracy of the inversion models. A total of three machine learning methods, including Random Forest regression, Support Vector Machine, and Partial Least Squares Regression, were used for water depth inversion experiments, and their accuracy was evaluated. The results indicated that the Landsat-8 data (20190716) with the best water transparency and weakest wave effect achieved the highest accuracy in water depth inversion. In the water depth range of 0-40 m, the R2 was 0.814, and the MAE, RMSE, and MAPE were 3.39 m, 4.31 m, and 0.366, respectively. In the water depth range of 0-20 m, the R2 was 0.874, and the MAE, RMSE, and MAPE were 2.24 m, 3.24 m, and 0.449, respectively. The RF algorithm obtained relatively high accuracy in the entire water depth range, while the SVM and PLSR algorithms displayed advantages in some shallow water depth inversions. The spatial resolution of optical remote sensing images is not an absolute positive correlation with the accuracy of water depth inversion. The hydrological characteristics of the water bodies in the remote sensing images have a significant impact on water depth inversion accuracy. Factors such as water transparency, suspended matter concentration, and seawater waves will affect the inversion accuracy. In the process of using optical remote sensing data for shallow water depth inversion, data with high water transparency and calm water conditions should be selected for modeling and inversion. The results have certain reference value for data source and algorithm selection in shallow water depth inversion based on multispectral remote sensing data.
Keywords