Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates

2026-06-26 · Source: stat.ML updates on arXiv.org · Field: Agriculture & Food Systems — Precision Agriculture & Smart Farming, Crop Science & Plant Technology, Agricultural Sustainability & Climate · Depth: Advanced, extended

Summary

A new probabilistic forecasting framework has been developed for short-term Normalized Difference Vegetation Index (NDVI) prediction, specifically addressing challenges of sparse and irregular satellite observations due to cloud cover. This transformer-based architecture integrates historical NDVI data with both historical and future meteorological covariates, explicitly separating these information streams. Key innovations include a temporal-distance weighted quantile loss, which aligns training with the effective forecasting horizon, and cumulative/extreme-weather feature engineering to capture delayed meteorological effects. Experiments on European satellite data demonstrate that this approach consistently outperforms diverse statistical, deep learning, and recent time series baselines across point-wise and probabilistic evaluation metrics. The model predicts NDVI quantiles up to 14 days ahead, specifically q∈{0.1,0.5,0.9}, and its code is available on GitHub. Ablation studies confirm the central role of target history and complementary gains from meteorological covariates.

Key takeaway

For agricultural scientists or precision agriculture engineers developing crop monitoring systems, you should consider integrating this probabilistic NDVI forecasting framework. Its ability to handle sparse satellite data and incorporate future weather covariates significantly improves short-term vegetation predictions up to 14 days. This allows for more accurate, uncertainty-aware decision support for irrigation, fertilization, and stress mitigation, enhancing proactive farm management strategies.

Key insights

Transformer-based probabilistic NDVI forecasting improves accuracy by integrating sparse satellite data with weather covariates and handling temporal irregularity.

Principles

• Decouple historical and future data processing for time series.
• Weight loss by temporal distance for irregular series.
• Engineer cumulative weather features for delayed effects.

Method

A transformer-based architecture processes historical NDVI and weather, and future weather covariates separately. It uses a temporal-distance weighted quantile loss and cumulative/extreme-weather feature engineering to predict multi-step NDVI quantiles.

In practice

• Use Sentinel-2 data for clear-sky NDVI.
• Apply time-aware linear interpolation for missing NDVI.
• Perturb future weather covariates to model uncertainty.

Topics

• NDVI Forecasting
• Precision Agriculture
• Transformer Models
• Satellite Remote Sensing
• Weather Covariates
• Probabilistic Time Series

Code references

• arco-group/ndvi-forecasting
• timeseriesAI/tsai
• Nixtla/neuralforecast

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• EO-WM: A Physically Informed World Model for Probabilistic Earth Observation Forecasting — EO-WM uses physically informed diffusion transformers to improve probabilistic Earth Observation forecasting by better modeling weather impacts.
• Zero-shot Multivariate Time Series Forecasting Using Tabular Prior Fitted Networks — Multivariate time series forecasting can be reframed as scalar regression for zero-shot tabular foundation model application.
• Hybrid Probabilistic Forecasting of Under-Five Malaria Admissions in Ghana: A Gaussian Process Regression with Holt-Winters Smoothing — Combining GPR with Holt-Winters provides robust probabilistic malaria forecasting, capturing non-linearity and seasonality.
• Exploring the Potential of Probabilistic Transformer for Time Series Modeling: A Report on the ST-PT Framework — Probabilistic Transformer (PT) maps Transformer operations to Mean-Field Variational Inference on a Conditional Random Field (CRF).
• Pointwise is Pointless? A Multimodal Ablation Study for Precipitation Nowcasting with Graph Neural Networks — Multimodal GNNs improve precipitation nowcasting, but point observations require careful integration to impact dense spatial fields effectively.
• Enhancing AI and Dynamical Subseasonal Forecasts with Probabilistic Bias Correction — Probabilistic Bias Correction (PBC) significantly improves subseasonal weather forecast accuracy and extreme event prediction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.