A novel hybrid deep learning framework integrating atmospheric teleconnection indices for daily soil moisture prediction in Kansas, USA

Hydrological cycles are affected by changes in soil moisture, and this parameter affects agricultural sciences and water resources management. The nonlinear interactions among climate parameters necessitate the use of advanced and intelligent modeling approaches to enhance accuracy and facilitate the identification of intricate relationships. In this study, seven hybrid deep learning models that include LSTM, Bi-LSTM, SSA-LSTM, EMD-LSTM, VMD-LSTM, CEEMD-LSTM, Autoencoder-LSTM, and ADIPLS-LSTM were utilized to predict daily soil moisture in Kansas, USA. For the accuracy evaluation of models uses daily data on precipitation, temperature, air humidity, solar radiation, and wind speed, from 2009 to 2020. Additionally, in another part of this study, the influence of Teleconnection Indices, including ENSO, PDO, PNA, NAO, AO, EPO, and SOI, on climatology parameters was evaluated. According to findings, it can be stated that combining data preprocessing techniques with LSTM has enhanced prediction accuracy. The findings indicated that integrating the CEEMD algorithm with LSTM enhanced the R2 value by roughly 20%. Among the models assessed, the ADIPLS-LSTM model revealed the most favorable predictive performance, with average values of CC, R², NSE, MAE, MSE, and RMSE recorded at 0.99, 0.98, 0.98, 0.33, 0.22, and 0.12, respectively, all of which are closely associated with ideal values. Teleconnection Indices, especially ENSO and PDO, have a notable impact on meteorological parameters and soil moisture in the state of Kansas, and the impact of the two mentioned indices on diverse parameters is maintained between 2 and 3 months.