*Abstract*
Satellite-based monitoring of smallholder agriculture is an important tool for food security but existing approaches are neither accessible nor effective for small plot field systems. To address these issues, crop type classification using representations generated by a global foundation model, TESSERA, is compared to best classification approaches in the literature. We present a novel approach to smallholder plots and compare representation based methods to raw data based methods for crop type classification in challenging environments. We find that our representation based approach offers a triple win: 1) consistent and statistically significant performance improvement over current methods, 2) greater simplicity due to the elimination of cloud masking and feature engineering, and 3) the reduction of computational cost. Our representation based approach achieves significantly higher F1 scores in the classification of 7 crop types for small fields in Austria for 5 classes (over 10% improvement in one case) and comparable F1 scores for two classes, and the best representation-based methods use 5% and 8% of compute compared to the best raw data method. These results indicate that representations are an effective approach for crop type classification tasks for small field systems.
*Bio*
Madeline Lisaius received BS and MS degrees in Earth Systems with a focus on environmental spatial statistics and remote sensing from Stanford University, Stanford, California, USA as well as MRes degree in Environmental Data Science from the University of Cambridge, Cambridge, UK. She is working towards the PhD in the Department of Computer Science and Technology at the University of Cambridge. She is focused on topics of food security and environmental justice, remote sensing, and machine learning.
