Answers are grounded in this paper and cite their sources.
Forecasting vegetation dynamics over decades requires learning from both seasonal cycles and long-term trends. Accurate forecasting over such timescales is challenging because ecosystems carry legacy effects of past disturbances, climatic variability, and recovery that shape future vegetation responses. However, existing deep learning methods for satellite image time series operate on short temporal windows of weeks to months. They do not capture the multidecadal patterns that encode this ecological memory. We present Seasonal Vision Transformer (SeasonalViT), a transformer-based model that processes 140 seasonal Landsat NDVI observations per location, collected over 35 years. The model separates vegetation dynamics into two temporal scales: within-year phenological transitions and across-year trends spanning decades. Rather than predicting absolute NDVI values, it forecasts deviations from the most recent same-season observation. The prediction horizon is flexible, ranging from one season to multiple years. This anchoring reduces sensitivity to differences among the four Landsat sensors. Training combines masked reconstruction pretraining with self-supervised forecasting fine-tuning, requiring no manual annotations. Alongside the model, we introduce Global Seasonal NDVI Archive, a dataset of 250,000 globally distributed locations with seasonal NDVI composites covering 1990–2024 at 30[m] resolution. We evaluate against eight baselines spanning statistical, machine learning, and deep spatio-temporal methods. SeasonalViT achieves the best performance under spatial, temporal, and combined distribution shift. The advantage is the greatest for land covers with structured interannual dynamics, such as forests and cropland.
No takes yet. Share an insight, caveat, or question.
Faran et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: