Abstract

High-accuracy lake bathymetry and mapping are crucial for estimating lake water storage on the Tibetan Plateau (TP). In this study, we constructed traditional empirical (TE) models and machine learning (ML) models to compare the prediction accuracy and remote sensing bathymetric mapping performance by using Sentinel-2 satellite imagery and in situ measured water depth from Caiduochaka (CK) and QiXiang Co in the central TP. We analyzed the relationship between the band reflectance and depth and explored the universality of the model in different lakes. The results indicated that when using the TE model, the mean absolute percentage error (MAPE) varied between 14.5% and 26.5% for the test dataset at different study sites. When using the ML models, the MAPE varied between 7.6% and 18.9%, and it was the better choice overall. For the test dataset of the random forest (RF) model with the highest accuracy, in the CK with the maximum depth of approximately 16 m, the mean absolute error (MAE) and root mean square error (RMSE) were 0.54 m and 0.89 m, and the precision was the highest with an MAE of 1.13 m and RMSE of 1.67 m in QiXiang Co with a maximum depth of approximately 28 m, whereas the portability of the model was not satisfactory. Overall, the results indicated that the ML model can obtain bathymetric maps with high accuracy, good visual performance, and reliability, outperforming the TE model. It can be used effectively for deriving accurate and updated high-resolution bathymetric maps for shallow lakes.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call