Improving Satellite-Derived Bathymetry Estimation with a Joint Classification-Regression Model.

Emerging deep learning methods for satellite-derived bathymetry (SDB), in which water depth is estimated using satellite band reflectance values, typically treat the problem as either classification or regression tasks, which can underperform, particularly when the depth data exhibit a skewed distribution. In this work, we propose a novel jointly trained classification-regression (JTCR) model for SDB that first classifies the input band reflectance values to correspond to a depth range and then performs regression within each range. Using Shetrunji reservoir, an inland reservoir in India, as a case study, with Sentinel-2 band reflectance values, we demonstrate that our proposed model outperforms other competitive deep learning models, including the model derived from the separate training of classification and regression tasks in the proposed classification-regression architecture. Concretely, we observe root-mean-square error (RMSE), mean absolute error (MAE), and R-squared ( R-2) values of 0.17, 0.05, and 0.99, respectively, in the proposed JTCR model, compared with 0.99, 0.71, and 0.85 in the feedforward neural network (FFNN) model.