Matthieu Gallet

Illustration by Mordillo

Since October 2025, I am Associate Professor at the University of Savoie Mont Blanc, LISTIC laboratory, teaching in the IDU (Informatics, Data and Usage) master’s program (machine learning, deep learning, databases). My research focuses on frugal machine learning frameworks, with an application to the remote sensing domain.

Between December 2024 and September 2025, I was a Research Engineer at CREA (Chamonix), developing vegetation-analysis pipelines from Lidar and optical data, deep learning for camera-trap and acoustic sensor data, and deploying the Mont Blanc massif’s weather-station and camera-trap network.

Since September 2024, I am a PhD student at LISTIC, University of Savoie Mont Blanc, supervised by Abdourrahmane Atto, Fatima Karbou and Emmanuel Trouvé.

My research sits at the crossroads of theory and practice: developing fast, robust and frugal algorithms for large-scale remote sensing data (time series, wide-footprint imagery), mainly applied to RADAR data (SAR and GPR).

My main contributions are novel algorithms — divergences between statistical distributions, fuzzy fusion of classifiers — for large-scale detection and classification. I apply them to the cryosphere, notably wet snow cover mapping in space and time, and glacier surface classification.

More recently, I focus on frugal approaches for energy efficiency and interpretability: deep learning on second-order statistics (covariance matrices) under geometric constraints, and self-supervised architectures adapted to non-Gaussian, causally-structured (temporal, volumetric) data. Building on my CREA work, I’m also interested in LIDAR-based classification of low, supra-forest vegetation.

I graduated from IPGP in 2021 (Master’s in Remote Sensing); my thesis addressed robust inversion and classification of Ground Penetrating Radar (GPR) data.

Selected publications

All publications are available via the Publications tab.

2026

ICPR
Aggregation of Ensemble of Classifiers with Fuzzy Learning: Application for Land Cover Classification on SAR Images

Gallet, Matthieu, Atto, Abdourrahmane, Karbou, Fatima and 1 more author

In Pattern Recognition 2026

Abs Bib

This paper presents a new aggregation framework based on learnable fuzzy operators for ensembles of classifiers of different natures. This framework is illustrated on land cover classification in mountainous environments using Synthetic Aperture Radar (SAR) imagery. This work is motivated by the complexity of labeling SAR data and the will of a optimal strategy to aggregate different classifiers, each one having its own strengths and weaknesses. One of the original features of this study is the novel application of fuzzy theory and its associated aggregators to a framework for classifying various surface types in mountain regions using SAR images and a forest of classifiers. This new approach involved adapting the Choquet integral to allow supervised learning, enabling us to learn the parameters of the constituent fuzzy measures. We propose a two-step framework for learning the aggregation parameters using probability estimates of a ensemble of classifiers and compare this approach to a standard aggregation method in machine learning: logistic regression. Through experiments on a SAR image dataset, we show that the proposed method outperforms standard methods, especially in the presence of noisy data, regardless of the number of training data.
@inproceedings{10.1007/978-3-032-31397-3_43, title = {Aggregation of Ensemble of Classifiers with Fuzzy Learning: Application for Land Cover Classification on {SAR} Images}, author = {Gallet, Matthieu and Atto, Abdourrahmane and Karbou, Fatima and Trouvé, Emmanuel}, booktitle = {Pattern Recognition}, address = {Cham}, publisher = {Springer Nature Switzerland}, year = {2026}, pages = {651--665}, isbn = {978-3-032-31397-3}, }
SigPro
On batch normalization for SPDnet

Gallet, Matthieu, Mian, Ammar, Bouchard, Florent and 1 more author

Signal Processing 2026

Abs Bib HTML

This paper deals with batch normalization for SPDnet, a deep learning architecture specifically designed to handle covariance matrices. Current SPDnet batch normalization relies on geometric means, which require expensive iterative computations and lacks formal backpropagation derivations. These limitations are addressed through two contributions: (i) formal backpropagation derivation for SPDnet batch normalization, and (ii) computationally efficient closed-form alternatives to the geometric mean: log-Euclidean, arithmetic and harmonic means, and the geometric mean of arithmetic and harmonic means. Simulation results on the batch normalization layer with different means demonstrate significant reductions in memory consumption and comparable training time by avoiding automatic differentiation, demonstrating the effectiveness of our approach. Numerical experiments on three real datasets show that simpler means can outperform the geometric mean with up to 5% accuracy improvements while reducing training time by a factor of 5.
@article{GALLET2027110793, title = {On batch normalization for {SPDnet}}, author = {Gallet, Matthieu and Mian, Ammar and Bouchard, Florent and Ginolhac, Guillaume}, journal = {Signal Processing}, volume = {250}, pages = {110793}, year = {2026}, issn = {0165-1684}, doi = {https://doi.org/10.1016/j.sigpro.2026.110793}, keywords = {SPDnet, Batch normalization, Riemannian geometry, Covariance matrices, Deep learning}, }

2024

ICASSP
Renyi Divergences Learning for explainable classification of SAR Image Pairs

Gallet, Matthieu, Mian, Ammar, and Atto, Abdourrahmane

In International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2024

Abs Bib HTML

We consider the problem of classifying a pair of Synthetic Aperture Radar (SAR) images by proposing an explainable and frugal algorithm that integrates a set of divergences. The approach relies on a statistical framework that takes standard probability distributions into account for modelling SAR data. Then, by learning a combination of parameterized Renyi divergences and their parameters from the data, we are able to classify the pair of images with fewer parameters than regular machine learning approaches while also allowing an interpretation of the results related to the priors used. Experiments on real multi-class data demonstrate the virtues of the suggested method when compared to both Random Forest and Convolutional Neural Networks (CNN) classifiers, showing its resilience to disturbances such as polluted labels and variations in the percentage of training data.
@inproceedings{10448227, title = {Renyi Divergences Learning for explainable classification of SAR Image Pairs}, author = {Gallet, Matthieu and Mian, Ammar and Atto, Abdourrahmane}, booktitle = {International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, year = {2024}, pages = {7445-7449}, keywords = {Training, Training data, Signal processing algorithms, Radar polarimetry,Probability distribution, Classification algorithms, Convolutional neural networks, SAR, Rényi divergence, explainable machine learning, classification}, doi = {10.1109/ICASSP48485.2024.10448227}, }

2023

TGRS
New Robust Sparse Convolutional Coding Inversion Algorithm for Ground Penetrating Radar Images

Gallet, Matthieu, Mian, Ammar, Ginolhac, Guillaume and 2 more authors

IEEE Transactions on Geoscience and Remote Sensing 2023

Abs Bib HTML PDF Code

In this paper, we propose two algorithms to enhance the interpretability of the hyperbola in B-scans obtained with a Ground Penetrating Radar (GPR). These hyperbolas are the responses of buried objects or cavities. To correctly detect and classify them, a denoising is typically necessary for GPR images as the signal-to-noise ratio is low, and the various interfaces naturally present in the earth have a strong response. Both algorithms are based on a sparse convolutional coding model plus a low rank component. It is solved through an Alternating Direction Method of Multipliers (ADMM) framework. In order to take into account the presence of outliers and the artifacts caused by the acquisition, the second algorithm is based on the Huber norm instead of the classic L2 -norm. These algorithms are tested on a real dataset labeled by geophysicists. The results show the denoising efficiency of this approach, and in particular the robustness of the second algorithm.
@article{gallet:hal-04075772, title = {New Robust Sparse Convolutional Coding Inversion Algorithm for Ground Penetrating Radar Images}, author = {Gallet, Matthieu and Mian, Ammar and Ginolhac, Guillaume and Ollila, Esa and Stelzenmuller, Nickolas}, journal = {IEEE Transactions on Geoscience and Remote Sensing}, year = {2023}, doi = {10.1109/TGRS.2023.3268477}, keywords = {Ground Penetrating Radar ; Sparse Inversion ; Convolutive Model ; Robust methods ; Dictionaries ; Radar ; Signal to noise ratio ; Radar imaging ; Signal processing algorithms ; Shape ; Radar antennas}, url = {https://hal.science/hal-04075772}, }