Résumé
Coffee is a highly valued commodity and a widely consumed beverage, playing an important role in global trade. However, coffee farming landscapes are increasing transitioning into smaller-scale agricultural setups. This transformation highlights the critical need for accurate classification and mapping of coffee cropping systems (CS), especially in countries like Uganda, where dense vegetation and complex terrain present substantial challenges to traditional land survey methods. Moreover, understanding the spatial distribution of Robusta coffee (Coffea canephora) CS is essential for developing site-specific management strategies, guiding extension services, and informing evidence-based policy decisions. To address this gap, the present study aimed to enhance the discrimination and mapping capabilities of Robusta coffee CS at a sub-pixel scale using a two-step classification approach and multi-date Sentinel-2 (S2) data. In the first step, the random forest (RF) classification algorithm was used to map the major land use and land cover (LULC) classes in Google Earth Engine platform. Then, the Robusta coffee cropland class was masked, and a sub-pixel multiple endmember spectral mixture analysis (MESMA) was employed to discriminate Robusta coffee CS using three endmembers (EMs) obtained from in-situ hyperspectral data collected in 2023 from: (i) Robusta coffee with agroforestry, (ii) Robusta coffee with banana, and (iii) Robusta coffee under full sun. The result showed 93.5% overall accuracy for the major LULC and 89.9% for all Robusta coffee CS classes, disentangled as follows: 91.3% accuracy for Robusta coffee with agroforestry, 88.5% for Robusta coffee with banana, and 91.2% for Robusta coffee under full sun. Moreover, the MESMA sub-pixel algorithm demonstrates credible performance in discriminating the Robusta coffee CS within each S2 pixel at the heterogeneous landscape in the study area. The findings of this study can inform site-specific interventions (e.g., pest management, fertilizer application, etc.) tailored to the type of CS.