Using feature engineering and machine learning in FAO reference evapotranspiration estimation

Abstract

The authors of this study investigated the use of machine learning (ML) and feature engineering (FE) techniques to accurately determine FAO reference evapotranspiration (ETo) with a minimal number of climate variables being measured. The recommended techniques for areas with insufficient measurements are based solely on daily temperature readings. Various ML methods were tested to evaluate how sophisticated an ML algorithm is for this task necessary. The main emphasis was on feature engineering, which involves converting raw variables into inputs better suited for ML algorithms, resulting in improved results. FE methods for estimating evapotranspiration include approximations of clear-sky solar radiation based on altitude and Julian day, approximate relative humidity and wind velocity, a categorical month variable, and variables interactions. The authors confirmed that the ability of ML in such tasks is not solely dependent on choosing the suitable algorithm but also on this frequently ignored step. The results of computational experiments are presented, accompanied by a comparison of the proposed method against standard ETo empiric equations. Machine learning methods, mainly due to the transformation of raw variables using FE, provided better results than traditional empirical methods and sophisticated ML algorithms without FE. In addition, the authors tested the applicability of the developed models in the broader area to evaluate the possibility of their generalizability. The potential of this approach to deliver improved predictions, reduced input requirements, and increased efficiency holds interesting promise for optimizing water management strategies, irrigation planning, and decision-making within the agricultural sector.

Keywords

• reference evapotranspiration
• input data reduction
• machine learning
• feature engineering.