Enhancing references evapotranspiration forecasting with teleconnection indices and advanced machine learning techniques

Abstract

After precipitation, reference evapotranspiration (ET_O) plays a crucial role in the hydrological cycle as it quantifies water loss. ET_O significantly impacts the water balance and holds great importance at the basin level because of the spatial distribution of managing water resources. Large scale teleconnection indices (LSTIs) play a vital role by influencing climatic variables and can be pivotal in determining ET_O and its predictive variables. This study aimed to model and forecast annual ET_O in Iran’s basins by utilizing LSTIs and employing various machine learning models (MLMs) such as least squares support vector machine, generalized regression neural network, multi-linear regression (MLR), and multi-layer perceptron (MLP). Initially, climate data from 122 synoptic stations covering six and 30, main and sub basins were collected, and annual ET_O values were computed using the Food and Agriculture Organization 56 (PMF 56) Penman–Monteith equation. The correlations between these values and 37 LSTIs were examined within lead times ranging from 7 to 12 months. Through a stepwise approach, the most influential predictor indices (LSTIs) were selected as input datasets for the MLMs. The findings revealed the significant influence of factors such as carbon dioxide (CO₂), Atlantic multidecadal oscillation, Atlantic Meridional Mode, and East Atlantic on annual ET_O. Overall, all MLMs performed well in terms of the Scatter Index during both training and testing phases across all sub-basins. Furthermore, the MLP and MLR models displayed superior performance compared to other models in the training and testing evaluations based on various assessment metrics.