Invariant learning as a pathway to robust potato yield prediction

Abstract

Yield prediction is an essential task to sustain the food market and to ensure the food for the world in the upcoming decades. Potatoes (Solanum tuberosum L.) are a vital staple food for many countries in the world and the advancement of accurate yield prediction will aid in promoting the agricultural industry. Potato is one of the most exportable agricultural products in Cyprus. Artificial Intelligence (AI) and Remote Sensing (RS) based agriculture monitoring has showed a massive impact in yield estimation in recent years. Monitoring vegetation indices like Normalized Difference Vegetation Index during the phenological stages of potatoes can provide identical insights into crop growth and yield. In this study, our focus lies on robust yield prediction across varied spatial and temporal dimensions. Specifically, we explore two distinct regions in Cyprus (i.e seaside and interior), each characterized by unique local agroclimatic conditions. The dataset encompasses potato yield data, in-situ meteorological data and vegetation indices derived by Sentinel-2 for a 7-years period (2017)(2018)(2019)(2020)(2021)(2022)(2023). Thus, we test invariant learning against traditional ML methods in terms of spatial robustness and data drift issues.