Time Series Forecasting in Agriculture: Explainable Deep Learning with Lagged Feature Selection

Reference evapotranspiration is a crucial metric in agricultural contexts, characterizing the evapotranspiration rate from a well-hydrated surface and serving as a fundamental benchmark for water management and crop irrigation, especially in arid regions. This study applied neural networks that integrates a Temporal Selection Layer to enhance the prediction of reference evapotranspiration levels through feature selection. This approach not only aims to refine the accuracy of time-series forecasting but also enhances the interpretability and efficiency of the model by identifying the most relevant features and periods. Additionally, the study incorporates the SHAP technique to determine and explain the contribution of individual features to the model outputs. The data for this research was collected from several meteorological stations of the Sistema Agrometeorológico para a Gestão da Rega no Alentejo in Portugal between 2012 and 2022. The results show that the Temporal Selection Layer effectively identifies important features in agricultural data, and by employing the SHAP technique, we enhance the understanding of how these features influence predictions to improve farm management decisions.