Machine Learning Outperforms EEflux METRIC Model on Point Estimates

Abstract

For a vast number of countries, especially in the Mediterranean area, water shortage is among the most important agricultural and environmental hazards. As water consumption rises, the situation worsens due to a deficient supply of resources because of unnecessary exploitation, deforestation, unfair distribution, water wars, and ill management. More than half of the water supply is used for agricultural purposes and the goal was always to grow more crops to satisfy the demands of a growing population. Not only does agriculture demand the highest water supply, but it also has the highest potential for improving quality. Hence, efficient and smart irrigation would be a necessary step for water preservation. In this thesis, we developed a full-fledged utility-based regression module using point-wise, probabilistic, conformal, and quantile regression trained on data gathered from flux towers across America and Europe. Our module aided in predicting evapotranspiration, a metric that allows farmers to know how much water to use for crop irrigation. Adding to that, we implemented a data oversampling module - SMOGN - which helped in up-sampling data in our regression problem based on rare versus none rare values. Using our probabilistic models, we were able to quantify the uncertainty in our predictions. We also performed several feature selection techniques and studied model interpretability using SHAP and LIME. We developed our research in a way to allow farmers and agricultural experts to choose between a point-wise prediction, a probabilistic and certain prediction, and a prediction interval. We were able to achieve the best results using our conformal model, yielding an 81% coverage rate, meaning that 81% of the actual data lie in its corresponding prediction interval.