Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates

Irene Iele¹, Giulia Romoli², Daniele Molino¹, Elena Mulero Ayllón¹, Filippo Ruffini^1,2, Paolo Soda^1,2, Matteo Tortora³

¹ University Campus Bio-Medico of Rome, ² Umeå University, ³ University of Genoa

Overview

Public release of the code for NDVI forecasting using satellite time series and weather data.

Contact

For questions and comments, feel free to contact me: irene.iele@unicampus.it

Citation

If you use this work, please cite:

@article{iele2026probabilistic,
  title={Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates},
  author={Iele, Irene and Romoli, Giulia and Molino, Daniele and Ayll{\'o}n, Elena Mulero and Ruffini, Filippo and Soda, Paolo and Tortora, Matteo},
  journal={arXiv preprint arXiv:2602.17683},
  year={2026}
}

Main Scripts

• dataset_builder/cache_builder_unified_monthly_noise.py: builds the cached dataset from the raw CSV files, adding monthly weather noise and saving the cache metadata and scaler.
• dataset_builder/torch_dataset.py: loads the cached samples as a PyTorch dataset, with optional feature engineering, scaling, and target discretization.
• dataset_builder/scaler.py: handles feature and target scaling, plus inverse transforms for evaluation.
• agrimatnet/model_quantile.py: defines the AgriMatNet quantile model and the pinball loss used during training.
• agrimatnet/train_quantile_ablation.py: trains the quantile model and supports ablation studies and time-weighted loss.
• agrimatnet/test_quantile_ablation.py: evaluates a trained checkpoint and exports metrics, predictions, and plots.
• agrimatnet/train_quantile.bash: SLURM wrapper that launches the training job on the cluster.
• agrimatnet/test_quantile_small_no_interp.bash: SLURM wrapper that launches the evaluation job on the cluster for the no-interpolation experiments.