Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates

A research team from multiple institutions has developed a breakthrough AI system for predicting crop health using sparse satellite imagery. The paper 'Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates' introduces a transformer-based architecture specifically designed to overcome the major challenge of irregular satellite sampling caused by cloud coverage.

The technical innovation centers on a dual-stream approach that separates historical vegetation dynamics from future exogenous information. The model integrates historical NDVI observations with both past and forecasted meteorological data, using a novel temporal-distance weighted quantile loss function that aligns training objectives with effective forecasting horizons. This addresses the irregular revisit patterns of satellites, where cloud cover can create data gaps of weeks. The system also incorporates cumulative and extreme-weather feature engineering to capture delayed meteorological effects on vegetation response.

In extensive experiments using European satellite data, the proposed approach consistently outperformed diverse baselines including statistical methods, deep learning models, and recent time series techniques. The framework achieved superior performance across both point-wise and probabilistic evaluation metrics, with ablation studies confirming the critical role of target history while showing meteorological covariates provide complementary gains when jointly exploited.

This represents a significant advancement for precision agriculture, where accurate short-term vegetation forecasting enables data-driven decision support for irrigation, fertilization, and harvest timing. The open-source availability of the code means agricultural technology companies and research institutions can immediately implement this approach to improve crop yield predictions and resource allocation.