Abstract
Accurate and timely monitoring of Sub-Saharan forests is critical for global carbon accounting, biodiversity conservation, and mitigating the impacts of climate change. However, traditional remote sensing approaches often struggle with persistent cloud cover, irregular temporal sampling, and the complex, non-linear dynamics of forest degradation. In this paper, we introduce a novel Spatiotemporal Graph Attention Network (ST-GAT) framework designed for real-time canopy cover estimation and rapid deforestation detection in Sub-Saharan African forests. By modeling forest regions as dynamic spatiotemporal graphs—where nodes represent localized forest patches and edges capture spatial and ecological dependencies—our model adaptively aggregates information across both space and time. We utilize multi-spectral imagery from Sentinel-2 and Landsat-8, integrated with meteorological covariates, to train and evaluate the ST-GAT model over several heterogeneous ecological zones in Central and East Africa. The experimental results demonstrate that the proposed ST-GAT outperforms state-of-the-art convolutional and recurrent architectures, achieving a coefficient of determination (R²) of 0.92 for canopy cover estimation and an F1-score of 0.89 for early deforestation detection. Crucially, the model exhibits robust performance under simulated heavy cloud cover, leveraging spatial attention to reconstruct missing temporal data. This research provides a scalable, highly interpretable, and computationally efficient tool to support real-time forest management and conservation policy decisions in ecologically vulnerable regions.
