Sentinel-2 forest typology mapping in Central Africa: assessing deep learning and image preprocessing effects

Introduction Central African forests are key reservoirs of carbon and biodiversity. Developing a detailed, spatially explicit typology of forest types is essential for monitoring and conservation. However, this task remains challenging due to limitations inherent to optical satellite imagery, especially disturbances caused by two major sources of noise: (i) atmospheric effects and (ii) Bidirectional Reflectance Distribution Function (BRDF) distortions, which introduce spectral inconsistencies across image collections. Even after standard corrections, residual errors often persist, masking the subtle ecological signals required for accurate classification. In this study, we evaluate whether recent deep learning models can implicitly learn to account for such distortions, potentially reducing the need for traditional preprocessing steps. Methods We produced a 10-m resolution vegetation typology map of the highly heterogeneous TRIDOM landscape (∼180,000 km 2 ) spanning Cameroon, Gabon, and the Republic of Congo, using Sentinel-2 imagery. We compared the performance of Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and self-supervised ViTs trained with DINOv2. Results Our results show that CNNs achieved the highest classification accuracy (OA = 0.91, Kappa = 0.84), outperforming both ViTs and DINOv2-based models (OA ≈ 0.70) on preprocessing images. When uncorrected imagery was used, CNN accuracy dropped to 0.76 (Kappa = 0.59), while ViTs exhibited also a decline (Kappa falling from 0.54 to 0.24). Discussion These findings highlight the partial ability of deep learning models to compensate for image noise, but emphasize that traditional preprocessing remains necessary for reliable classification. Our results also demonstrate that CNNs consistently outperform self-supervised Vision Transformers in large-scale forest mapping, providing accurate classification of forest typologies. This work offers new insights into the robustness and current limitations of deep learning architectures when applied to complex tropical landscapes.