Challenging Global Generalizations: Superior Land Cover Mapping in Botswana with a Locally Trained Transformer Model

Tshepho Manyothwane^*Gizaw Mengistu Tsidu^*

• Botswana International University of Science and Technology, Palapye, Botswana

Accurate and high-resolution land use and land cover (LULC) classification remains a critical challenge in ecologically diverse and spatially heterogeneous dryland environments, particularly in data-scarce regions. This study presents a Transformer-based deep learning model trained on Landsat 8 OLI imagery to classify LULC across Botswana, a region characterized by complex environmental gradients and dynamic land cover transitions. This work provides one of the first applications of Transformer architectures for national-scale LULC mapping in Botswana, integrating field observations, Dynamic World-derived labels, and Google Earth validation to construct reliable training datasets in data-limited regions. The model achieved an overall accuracy of 95.31%, with a Total Disagreement (TD) of 4.69%, primarily driven by Allocation Disagreement (AD= 3.44%) rather than Quantity Disagreement (QD = 1.25%). This indicates that the model accurately estimates class proportions but exhibits the misplacement of some classes. Despite this, F1-scores of 0.80 or higher for most land cover categories, reflect strong thematic performance. Compared to the global Dynamic World (DW) product, the Transformer model demonstrated superior spatial detail, class-wise accuracy, and robustness, particularly in urban settings and ecologically sensitive zones such as the Makgadikgadi Pans and Okavango Delta. The Transformer-based land cover classification model demonstrated superior performance compared to widely used global LULC products. While global datasets such as ESA WorldCover (65% OA), Dynamic World (72% OA), and Esri Land Cover (75% OA) provide valuable baseline products, their accuracies remain significantly lower than our locally trained Transformer model, which achieved an overall accuracy of 95.31% on the testing dataset. Temporal LULC trajectories reconstructed for 2014, 2019, and 2024 effectively captured major land change processes, including cropland expansion, grassland regeneration, and seasonal floodingoffering a valuable tool for environmental monitoring and sustainable land management in semi-arid regions.