Advanced evapotranspiration forecasting in Central Italy: Stacked MLP-RF algorithm and correlated Nystrom views with feature selection strategies

Abstract

Evapotranspiration is a key parameter in hydrology, particularly in the field of water resources management. Reference evapotranspiration (ETo) stands as a crucial metric, embodying the influence of climate on water loss from both soil and plant surfaces. The accurate forecasting of future ETo values is paramount for informed decision-making in agricultural practices and water supply planning. The anticipation of evapotranspiration variations supports optimized irrigation, drought assessment, and efficient water allocation. Employing innovative algorithms, specifically the Multilayer Perceptron-Random Forest (MLP-RF) Stacked Model and the Correlated Nystrom Views (XNV), this study focuses on predicting ETo up to 60 days ahead in Agro Pontino, an area in Central Italy known for its flourishing agricultural production in the Mediterranean Europe. A Radial Basis Function (RBF) Neural Network serves as a benchmark. Taking into account firstly all available weather-climatic variables and subsequently adopting two different strategies for reducing the variables, based respectively on Principal Component Analysis and feature selection using Particle Swarm Optimization, three distinct sets of input variables were considered. The models based on the complete set of exogenous climatic variables demonstrated superior accuracy. However, even models relying on only mean temperature, maximum relative humidity, and shortwave solar radiation as inputs produced good results, also for the 60-day forecasting horizon, with Kling-Gupta Efficiency (KGE) and Mean Absolute Percentage Error (MAPE) equal to 0.98 and 8.356 %, respectively, in the case of MLP-RF Stacked Model. The latter consistently outperformed XNV and RBF across various combinations of input variables and forecasting horizons. Notably, the reduction in accuracy with extended forecasting horizons was mild, suggesting the potential for accurate results over significantly more extended horizons. These forecasting models facilitate precise irrigation scheduling, minimizing water wastage and conserving resources. This targeted approach enhances crop yields, quality, and environmental sustainability, rendering agriculture economically viable and adaptable to climate variability.