• Probabilistic NDVI forecasting tackles sparse, irregular satellite data due to cloud cover. • Transformer architecture separates historical NDVI dynamics from future weather covariates. • Temporal-distance weighted quantile loss aligns training for horizon-dependent uncertainty in predictions. • Cumulative and extreme-weather features capture delayed meteorological effects on vegetation. • Outperforms statistical, deep learning, and time-series baselines on European satellite datasets. • Ablation shows target history critical; meteorological covariates add complementary gains. • Code released publicly, enabling reproducibility and further research for the community.
Article Summaries:
- Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates
A new machine‑learning framework delivers short‑term, field‑level predictions of the Normalized Difference Vegetation Index (NDVI) using limited satellite observations and weather data. The method employs a transformer architecture that separates historical vegetation dynamics from future exogenous inputs, integrating past NDVI values with both historical and forecasted meteorological covariates. To handle irregular revisit intervals and horizon‑dependent uncertainty, a temporal‑distance weighted quantile loss is introduced, while cumulative and extreme‑weather feature engineering captures delayed climate effects. Experiments on European satellite datasets show consistent improvements over statistical, deep‑learning, and recent time‑series baselines in both point‑wise and probabilistic metrics. The code is publicly released.
Sources: