![[2602.17683] Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.17683] Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates
Summary
This paper presents a probabilistic framework for forecasting NDVI from sparse satellite data and weather covariates, enhancing precision agriculture decision-making.
Why It Matters
Accurate NDVI forecasting is crucial for effective agricultural management, especially under varying climatic conditions. This research addresses the challenges posed by sparse data and offers a novel approach that integrates historical NDVI and meteorological data, potentially improving crop yield predictions and resource management.
Key Takeaways
- Introduces a transformer-based model for NDVI forecasting.
- Addresses data sparsity and irregular sampling from satellite observations.
- Incorporates weather covariates to enhance predictive accuracy.
- Utilizes a novel temporal-distance weighted quantile loss for training.
- Demonstrates superior performance over existing forecasting methods.
Computer Science > Machine Learning arXiv:2602.17683 (cs) [Submitted on 4 Feb 2026] Title:Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates Authors:Irene Iele, Giulia Romoli, Daniele Molino, Elena Mulero Ayllón, Filippo Ruffini, Paolo Soda, Matteo Tortora View a PDF of the paper titled Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates, by Irene Iele and 6 other authors View PDF HTML (experimental) Abstract:Accurate short-term forecasting of vegetation dynamics is a key enabler for data-driven decision support in precision agriculture. Normalized Difference Vegetation Index (NDVI) forecasting from satellite observations, however, remains challenging due to sparse and irregular sampling caused by cloud coverage, as well as the heterogeneous climatic conditions under which crops evolve. In this work, we propose a probabilistic forecasting framework specifically designed for field-level NDVI prediction under clear-sky acquisition constraints. The method leverages a transformer-based architecture that explicitly separates the modeling of historical vegetation dynamics from future exogenous information, integrating historical NDVI observations with both historical and future meteorological covariates. To address irregular revisit patterns and horizon-dependent uncertainty, we introduce a temporal-distance weighted quantile loss that aligns the training objective with the effective forecasting horizon. In...