Open-Earth-Monitor Global Workshop 2026

Savanna ecosystems cover approximately one fifth of Earth's land surface and provide critical ecosystem services to over one billion people, yet their dynamic vegetation layer remains difficult to monitor consistently at scale. Spaceborne lidar from the Global Ecosystem Dynamics Investigation (GEDI) mission provides vegetation structure measurements, such as canopy height and cover, but its spatially sparse sampling necessitates extrapolation using satellite remote sensing data. Temporal consistency of these wall-to-wall mapping products remains a key challenge, particularly in heterogeneous savanna systems characterized by pronounced seasonality and complex disturbance dynamics.
This study compares two approaches for mapping GEDI-derived canopy height and cover across Kruger National Park, South Africa. The first uses hand-crafted Sentinel-1/2 features derived from phenology-informed time series analysis. The second uses TESSERA foundation model embeddings (pixel-wise representations of annual Sentinel-1/2 time series) as open, analysis-ready features with lightweight regression heads. Both approaches use phenology-aligned GEDI samples anchored to leaf-on conditions as training data, and are evaluated using temporal cross-validation and independent airborne lidar data acquired across multiple sites in the study area, with particular focus on temporal transferability and label efficiency.
The comparison addresses a question of growing practical relevance: does explicit phenological knowledge embedded in task-specific feature engineering outperform the implicit temporal representations learned by large-scale foundation models, and under what conditions? The results will inform scalable, open, and reproducible approaches to vegetation structure monitoring in African savannas, with direct relevance for biodiversity conservation and carbon stock assessment.