Publications

Duhan, D; Singh, MC; Singh, D; Satpute, S; Singh, S; Prasad, V (2023). Modeling reference evapotranspiration using machine learning and remote sensing techniques for semiarid subtropical climate of Indian Punjab. JOURNAL OF WATER AND CLIMATE CHANGE, 14(7), 2227-2243.

Abstract
A study was carried out to develop and evaluate the performance of different machine learning (ML) models for predicting reference evapotranspiration (ET0). The models included multiple linear regression (MLR), least square-support vector machine (LS-SVM), artificial neural networks (ANNs) and adaptive neuro-fuzzy inference system (ANFIS). The daily meteorological data for 50 years (1970-2019) were used to estimate ET0 using FAO-ET calculator. The FAO-ET calculator was compared with ML models to investigate the best-fit ML model for predicting ET. Thereafter, ET predicted by the best-fit ML model was compared with satellite (MOderate resolution Imaging Spectroradiometer - MODIS) ET, which was finally mapped to a larger landscape (over entire Punjab and Haryana). Modeling of ET0 was best performed through LS-SVM followed by ANN2, ANN1, ANFIS10, ANFIS2, MLR and ANFIS9 models. Among developed models, coefficient of determination (R-2) value varied from 0.800 to 0.998, being highest (0.998) under LS-SVM model. MODIS overestimated ET when compared with LS-SVM having R-2 and root mean square error (RMSE) values of 0.73 and 3.95 mm, respectively. After applying the bias correction factor, R-2 and RMSE were 0.74 and 1.19 mm, respectively. The ML and satellite-based ET estimation would be useful for timely water budgeting to manage the water scarcity problems from local to regional levels.

DOI:
10.2166/wcc.2023.003

ISSN:
2408-9354