Integration of machine learning and deep learning models for hydro-meteorological data gap filling and prediction of daily rainfall in the Lake Abaya-Chamo Sub-basin, South Ethiopia

Abstract

Lake Abaya-Chamo Sub-basin, Ethiopia. Assessment of water resource development is hampered by hydro-meteorological data gaps at spatiotemporal scales and improved by using gridded data. The traditional gap-filling method was common; however, it fails in capturing spatiotemporal variability and nonlinear time-series data. In this study, advanced machine learning (ML), deep learning (DL), and spatial interpolation (SI) methods were applied to fill data gaps and predict rainfall. The performance of each algorithm was analyzed using performance matrices (R 2 , NSE, RMSE, MAE, and PBIAS). The study revealed that long-short term memory (LSTM) is superior to ML and SI methods for daily rainfall data infilling and prediction, due to its ability to capture non-linear and intricate long-term temporal dependencies. In stream flow data gap filling, Support Vector Regression (SVR) outperformed at Kulfo; Random Forest (RF) was superior at Bilate, Gidabo, Gelana, Hamesa, and Hare while decision tree (DT) had performed the least for all watersheds. Therefore, advanced ML and DL methods addressed the weakness of traditional gap fillings for hydro-meteorological data and prediction of rainfall in this sub-basin. This study suggested that ML and DL methods were better for accurate data gap filling and prediction for better water resources development in data-scarce regions. • Integration of SI, ML, and DL improves daily rainfall gap-filling in the Lake Abaya-Chamo Sub-basin. • LSTM performs well in low rainfall but underestimates extreme events based on 90th percentile analysis. • ML methods (RF, SVR, DT) provide reliable baseline for reconstructing missing daily streamflow data. • LSTM accurately predicts daily rainfall in the basin and supports water resource studies in Ethiopian basins.