The Generalisation of Machine Learning Models for Predicting Irrigated Agriculture in Heterogeneous Landscapes: An Examination of Model Transferability in Mozambique

This study investigates whether machine learning models for mapping irrigated agriculture can be effectively transferred between regions to save resources compared to collecting new training data. We examine how models trained in one region of Mozambique perform when applied to another region with different landscape characteristics. The results show that simple transfers of models between regions are generally ineffective due to regional variations in weather conditions, landscapes, and farming practices. However, incorporating diverse data from multiple regions improves classification performance. The study highlights the importance of targeted data collection in new areas and demonstrates approaches for identifying areas with high prediction errors to guide efficient field data collection efforts.

Preprint

Under Review