Deep Noisy-Label Learning Based Cross-Resolution Land-Cover Mapping in Kenya

Abstract. High-resolution land-cover information is critical for environmental monitoring across rapidly changing African landscapes, yet national-scale 10 m mapping remains limited by scarce training data and inconsistencies among existing products. This study presents a scalable framework for generating a 10 m land-cover map of Kenya for 2022 by integrating multi-temporal Sentinel-2 composites, historical land-cover datasets, and volunteered geographic information. A confidence-weighted fusion of GlobeLand30, FROM-GLC, and FROM-GLC10 produces large-scale synthetic labels that reduce temporal misalignment and systematic biases. A ConvNeXt–UPerNet segmentation network trained with a noise-aware bootstrap focal loss effectively captures Kenya’s strong multi-scale heterogeneity, from cropland mosaics and rangelands to riparian corridors and expanding urban areas. The resulting map shows clear improvements in spatial coherence and thematic detail over existing 10–30 m products. The proposed approach offers a practical pathway for routine 10 m national mapping in data-sparse regions and provides timely, reliable information for ecological monitoring, agricultural assessment, and sustainable land-use planning in Kenya and East Africa.