Accurate sub-seasonal root-zone soil moisture prediction using attention-based autoregressive transfer learning and SMAP data

Abstract

Root zone soil moisture (RZSM) is an important hydrological variable for agricultural planning and water resources management. The Soil Moisture Active Passive Level 4 (SMAP L4) data demonstrates great value in RZSM estimation. Accurate sub-seasonal RZSM prediction based on SMAP L4 holds great significance for agricultural management and drought assessment. Current deep learning-based RZSM prediction models tend to accumulate error in long-term forecasting and the limited SMAP RZSM samples may result in insufficient model generalization. To address these issues, this study proposes a multi-head self-attention-based autoregressive transfer learning model based on long short-term memory (MAATL) model for sub-seasonal RZSM prediction. The proposed MAATL model is evaluated over the Continental United States (CONUS) for 1- to 60-day RZSM prediction and compared with some ablation and long short-term memory (LSTM) models. The results showed that compared with LSTM, the skills of the MAATL model were significantly improved, with an average correlation coefficient increase of 18.26% and a root mean square error (RMSE) reduction of 42.55%. Furthermore, 118 in-situ soil moisture stations are used for predictive validation and the proposed MAATL model demonstrates higher accuracy compared to the Global Forecast System (GFS) and the LSTM model, with an average correlation skill improvement of 16.02% and 15.08% for MAATL over GFS and LSTM, respectively. These findings indicate superior performance for the proposed MAATL model in sub-seasonal RZSM prediction, which has great potential for agricultural drought preparations.

Keywords

• RZSM
• SMAP
• Multi-head self-attention
• Autoregressive
• Transfer learning