Evaporation and soil moisture prediction with artificial intelligence and deep learning methods

Evaporation and soil moisture prediction with artificial intelligence and deep learning methods

Understanding future changes and predicting hydrological variables well in advance is practically useful in water resources and drought management measures. This doctoral thesis presents the new methodologies and the findings based on three primary objectives that aim to build artificial intelligence and deep learning hybrid models to forecast drought-related hydrological variables comprised of evaporation, evapotranspiration, and soil moisture within the key drought-prone regions in Queensland, Australia. Data preprocessing techniques that involve feature selection and data decomposition to reveal the patterns or trends in modeling data are used in the model hybridization stage where standalone models are integrated with these techniques and the significance of their influence in enhancing the model performances are tested. In the first objective, the Long Short-Term Memory (LSTM) predictive model is hybridized with the Neighborhood Component Analysis (NCA) feature selection technique to enhance the model’s predictive efficacy that aims to accurately predict pan evaporation (Ep). The second objective aims to develop novel methods to forecast reference evapotranspiration (ET) and is achieved by hybridizing the LSTM model with Boruta-Random Forest (Boruta) feature selection technique and the Multivariate Empirical Mode Decomposition (MEMD) technique to further improve the efficacy. In the third objective, the 1-, 14-, and 30-days ahead soil moisture (SM) within the topsoil layer (0–10 cm depth) is forecasted by employing a hybrid deep learning forecasting model built using LSTM network coupled with Maximum Overlap Discrete Wavelet Transform (moDWT) data decomposition method and Least Absolute Shrinkage and Selection Operator (Lasso) feature selection method. When compared with benchmark models, all the hybrid models developed in this study registered a comparatively high performance with low error performance metrics to demonstrate their usefulness in forecasting Ep, ET, and SM values in the present study region. In the practical sense, as the models developed in this study provide accurate estimations, their capabilities can undoubtedly be employed to successfully manage water resources and drought events. Further, this doctoral study shows that artificial intelligence and deep learning models developed in this study could be a significant forward step in contributing to the advancement of data-driven hydrological forecasting methods that may be useful for understanding the future trend of hydrological variables. The outcomes and implications thus contributed to the advancement of science while creating socio-economic benefits due to their usefulness in water resources and drought event management.

File reproduced in accordance with the copyright policy of the publisher/author.