Estimating Small Reservoir Evaporation Using Machine Learning Models for the Brazilian Savannah

Abstract

Small dams are infrastructures that regulate water supply for multiple users and play a key role in the agricultural development of the Brazilian savannah region known as the Cerrado. Evaporation is one of the major components of the hydrological cycle of small reservoirs, and should be better quantified. Studies based on machine learning techniques usually adjust models based on large datasets, which are frequently unavailable in developing countries. This study adjusted and evaluated the performance of different evaporation machine learning models that were regressed on a very small dataset and for restrictive scenarios. The performance of each model was assessed with five climatic input combinations. The performance of the random forest models was one of the better for the input combinations, and was considered to be one of the more robust machine learning techniques among those assessed for estimating evaporation from a small reservoir in the region. The Penman (benchmark) equation performed worse, as it overestimated evaporation by 14.7% on average. Strategies for improving the performance and applicability of models and overcoming data scarcity in remote areas are further discussed.