Machine Learning Techniques-Based Estimation of Monthly Reference Evapotranspiration in Uzbekistan Using Latitude, Longitude, and Elevation Data

Abstract

This study investigates the potential of machine learning algorithms to estimate monthly reference evapotranspiration (ETo) in Uzbekistan using limited input data. ETo values were calculated using the FAO-56 Penman-Monteith method for 10 meteorological stations (1971-2000). Elevation, latitude, longitude, and month number were used as inputs, while ETo was the target output. Multilayer Perceptron (MLP), Adaptive Neuro-Fuzzy Inference Systems (ANFIS), and Random Forest (RF) models were trained and tested using two data-splitting strategies: leave-one-out and 70/30 train-test split. Model performance was evaluated using R2, RMSE, and MAE. Among the models, RF achieved the highest accuracy and generalization capability. While MLP performed well in some locations, its performance was more variable. ANFIS showed sensitivity to membership function selection, with gaussmf performing best. Sensitivity analysis indicated that latitude and longitude were the most influential predictors. The results support the use of machine learning models for ETo estimation and future spatial mapping to assist in water resource management.