A Deep Learning Approach to Predict Surface Soil Wetness and Its Uncertainty Analysis Over the Tel River Basin, India

Abstract

Surface soil moisture (SSM) refers to the capacity of the top layer of soil to hold moisture. It is an essential part of the budget for surface water. Soil moisture monitoring is crucial to reduce the effects of precipitation deficits and determine the best ways to manage natural ecosystems in the face of climate change. The current study collected daily SSW data from MERRA-2 for the Tel River Basin in Odisha, India, from 2001 to 2020 with a spatial resolution of 0.5° × 0.625°. To forecast SSW time series (SSWTS) one step ahead, this study examines the reliability of three deep learning (DL) models: gated recurrent unit (GRU), long short-term memory (LSTM), and simple recurrent neural network (simpleRNN). This study aims to address the following research questions: (1) How accurately can DL models predict SSWTS? (2) Which DL model—GRU, LSTM, or simpleRNN—is the most reliable for SSW forecasting? (3) How can the uncertainty in the predicted SSW be quantified and analyzed? Further, in an uncertainty investigation on SSW projected values, a Wilson score technique was employed to evaluate the uncertainty of the DL methods. GRU has outdone the other two models in forecasting monthly SSW with a 12-lookback timestep with a lower error for all the stations. The model appeared more accurate as it declined in gradient on larger sequencing samples. GRU's ability to remember significant prior knowledge, whereas discarding irrelevant data may assist in finding a novel, dependable solution for SSWTS forecasting.