CPAZMaL: A New Dataset of X-band SAR Time Series in the French Alps for Machine Learning
Abstract
High-resolution SAR time-series datasets remain scarce for high-mountain land cover applications, limiting the development of advanced machine learning methods for temporal monitoring of mountain surfaces in complex Alpine terrain. This paper introduces CPAZMaL: Cryosphere analysis with PAZ images for Machine Learning, a comprehensive benchmark dataset combining 3 years of X-band SAR acquisitions from PAZ satellites over the Mont-Blanc massif (Western European Alps). The dataset encompasses 8 terrain classes, including rarely studied glacier types such as ice aprons and hanging glaciers alongside well-characterized accumulation and ablation zones, as well as non-glacial surfaces such as rock outcrops, Alpine grasslands, and forested areas, enabling the study of landscape transitions driven by climate change. All the SAR images were processed through a standardized radiometric calibration and terrain correction workflow using a custom 1.5-m LiDAR DEM, ensuring geometric and radiometric consistency across dualpolarization (HH/HV) and both orbit (ascending/descending) configurations. Data are organized in a hierarchical HDF5 structure with group-based sampling to minimize spatial autocorrelation. We establish baseline performance across 4 machine learning scenarios: temporal classification, time-series forecasting, and domain adaptation, addressing polarimetric and cross-sensor transfer. Results demonstrate the dataset's utility for developing robust classification methods that handle geometric distortions, temporal variability, and the limited training samples inherent to high-resolution SAR data in steep Alpine environments, with a strong emphasis on glacial dynamics.
Related
- Tropical Cyclone-Following Dataset for Limited-Area Machine Learning Applications
- GLORIF1, a global river flow dataset created by integrating process-based modelling and machine learning
- Distribution dataset of chondrichthyan species occurring in French metropolitan waters and across the North-East Atlantic
- French urban tree dataset with age, dimensions and species
- A mangrove-based dataset to study the ecology of Neotropical mosquitoes (Diptera, Culicidae) in French Guiana
BibTeX
@article{gallet2026cpazmal,
title = {{CPAZMaL: A New Dataset of X-band SAR Time Series in the French Alps for Machine Learning}},
author = {Matthieu Gallet and Suvrat Kaushik and Christophe Lin-Kwong-Chon and Ludovic Ravanel and Emmanuel Trouvé},
journal = {IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
year = {2026},
month = jul,
doi = {10.1109/jstars.2026.3713622},
url = {https://hal.science/hal-05743098v1},
note = {CC BY-NC-SA 4.0},
}
RIS
TY - JOUR
TI - CPAZMaL: A New Dataset of X-band SAR Time Series in the French Alps for Machine Learning
AU - Matthieu Gallet
AU - Suvrat Kaushik
AU - Christophe Lin-Kwong-Chon
AU - Ludovic Ravanel
AU - Emmanuel Trouvé
PY - 2026
DA - 2026/07/15
JO - IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
DO - 10.1109/jstars.2026.3713622
UR - https://hal.science/hal-05743098v1
AB - High-resolution SAR time-series datasets remain scarce for high-mountain land cover applications, limiting the development of advanced machine learning methods for temporal monitoring of mountain surfaces in complex Alpine terrain. This paper introduces CPAZMaL: Cryosphere analysis with PAZ images for Machine Learning, a comprehensive benchmark dataset combining 3 years of X-band SAR acquisitions from PAZ satellites over the Mont-Blanc massif (Western European Alps). The dataset encompasses 8 terrain classes, including rarely studied glacier types such as ice aprons and hanging glaciers alongside well-characterized accumulation and ablation zones, as well as non-glacial surfaces such as rock outcrops, Alpine grasslands, and forested areas, enabling the study of landscape transitions driven by climate change. All the SAR images were processed through a standardized radiometric calibration and terrain correction workflow using a custom 1.5-m LiDAR DEM, ensuring geometric and radiometric consistency across dualpolarization (HH/HV) and both orbit (ascending/descending) configurations. Data are organized in a hierarchical HDF5 structure with group-based sampling to minimize spatial autocorrelation. We establish baseline performance across 4 machine learning scenarios: temporal classification, time-series forecasting, and domain adaptation, addressing polarimetric and cross-sensor transfer. Results demonstrate the dataset's utility for developing robust classification methods that handle geometric distortions, temporal variability, and the limited training samples inherent to high-resolution SAR data in steep Alpine environments, with a strong emphasis on glacial dynamics.
SN - 1939-1404
ER -