Cross-Regional Transferability of AI Crop-Type Mapping: Insights and Challenges

Abstract

Accurate crop discrimination is vital for effective agricultural planning and sustainability management, especially in regions like SubSaharan Africa (SSA), where small-scale farming predominates and ground data is scarce. Conducting field surveys in SSA is challenging due to labor and cost constraints, as well as logistical and political barriers. This paper explores the feasibility of transferring crop-type classification models between regions with similar crops. Utilizing data from Baraouéli and Karamoja, collected from Source Cooperative, we trained multi-layer perceptron (MLP), random forest (RF), and support vector machine (SVM) classifiers using Sentinel-2 imagery. These models were then evaluated and applied to cross-crop type classification between Baraouéli (Mali) and Karamoja (Uganda) to assess the transferability of machine learning models. While the models demonstrated strong local performance, achieving high overall accuracy in their respective regions, their performance declined when transferred between regions. However, focusing solely on specific crops such as maize and sorghum improved the models’ performance, albeit with reduced accuracy compared to local classifications. The study suggests that incorporating additional features such as texture, DEM, crop height, and weather data could enhance the adaptability of classifiers between regions. These findings highlight the potential for developing transferable models within SSA to address challenges related to limited ground surveyed data, providing valuable insights for researchers and policymakers