Bmaina/vegformer-maasai-mara

# VegFormer: Cross-Modal Temporal Transformer for Vegetation Dynamics and Land Degradation Mapping > **Maasai Mara Ecosystem, Kenya | 2019–2024 | Sentinel-1/2 · CHIRPS · Self-Supervised Learning** --- ## Overview VegFormer is a deep learning framework for mapping vegetation dynamics and land degradation across the Maasai Mara ecosystem using multi-sensor satellite time series. The model fuses Sentinel-1 SAR backscatter, Sentinel-2 multispectral imagery, and CHIRPS rainfall data across 137 sixteen-day composites spanning 2019 to 2024, learning to distinguish vegetation types and identify degradation trajectories without requiring hand-labeled training data. The project builds on prior MODIS NDVI-based vegetation mapping work using Savitzky-Golay smoothing, statistical features, and Braun-Blanquet clustering, and advances it with modern deep learning architectures and self-supervised pre-training — producing richer, more interpretable, and more scalable outputs. --- ## Why This Matters The Maasai Mara ecosystem supports one of the world's most iconic wildlife migrations and sustains the livelihoods of hundreds of thousands of Maasai pastoralists. It faces severe and accelerating pressure from multiple directions. Agricultural encroachment from surrounding smallholder farms reduces the effective ecosystem area by an estimated 1 to 2 percent annually (Butt et al., 2011, Journal of Arid Environments). Overgrazing in community conservancies degrades grass cover and promotes woody encroachment (Groom and Harris, 2020, African Journal of Ecology). Rainfall variability driven by ENSO teleconnections creates multi-year drought cycles that push marginal land past ecological tipping points (Ogutu et al., 2008, Journal of Animal Ecology). Field-based vegetation surveys using Braun-Blanquet transects are expensive, infrequent, and spatially sparse, leaving large areas unmonitored between survey campaigns. Satellite-based vegetation monitoring offers a scalable alte …