Research

• Under review

  Divergent and asymmetrically reversible accumulation-zone failure across western North American glaciers

  Tarka, M., Baraër, M., Aubry-Wake, C.

  Nature Climate Change

  Under review.

• Preprint

  2026

  Benchmarked and leakage-validated deep-learning glacier segmentation: a 42-year, five-sensor snow-fraction record for western North America

  Tarka, M., Baraër, M., Aubry-Wake, C.

  SSRN

  Submitted to the ISPRS Journal of Photogrammetry and Remote Sensing. The code and the data are open.

  Cite this work

  @misc{tarka2026glaciersegmentation,
    author    = {Tarka, Maxime and Bara{\"e}r, Michel and Aubry-Wake, Caroline},
    title     = {Benchmarked and leakage-validated deep-learning glacier segmentation: a 42-year, five-sensor snow-fraction record for western North America},
    year      = {2026},
    publisher = {SSRN},
    doi       = {10.2139/ssrn.7418104},
    note      = {Preprint}
  }

  Copy BibTeX

  Preprint on SSRNCode on GitHub

• Published dataset

  2026

  Data for: Benchmarked and leakage-validated deep-learning glacier segmentation: a 42-year, five-sensor snow-fraction record for western North America

  Tarka, M., Baraër, M., Aubry-Wake, C.

  Zenodo, version 1.0, CC BY 4.0

  The trained models, the 540 annotated scenes, four baselines, the analyses and the yearly files of each glacier, all that goes with the preprint.

  Cite this work

  @dataset{tarka2026glaciersdata,
    author    = {Tarka, Maxime and Bara{\"e}r, Michel and Aubry-Wake, Caroline},
    title     = {Data for: Benchmarked and leakage-validated deep-learning glacier segmentation: a 42-year, five-sensor snow-fraction record for western North America},
    year      = {2026},
    version   = {1.0},
    publisher = {Zenodo},
    doi       = {10.5281/zenodo.22311126},
    license   = {CC-BY-4.0}
  }

  Copy BibTeX

  doi 10.5281/zenodo.22311126

• In preparation

  Data article on the weather-station dataset of the Sainte-Marthe experimental catchment, Quebec, Canada

  Tarka, M. (first author), et al.

The same idea appears as truncation in the Geotop talk, decapitation in the thesis and the CGU talk, and accumulation-zone failure in the manuscript under review.

Satellite scenes of the same glacier look alike. If scenes are split at random, the same glacier can land on both sides, which happened for 35.9% of glaciers in my data. That split scores 0.777 on validation against 0.762 for a split by glacier, which looks like a gain and is not. Inside that run, scenes of glaciers seen in training score 0.696 and scenes of glaciers never seen only 0.575. Every number I report comes from a split by glacier, with a held-out test score of 0.705. The full analysis is in the SSRN preprint.

Article-based thesis defended on 26 June 2026, rated Excellent, and recommended by the jury for the ÉTS Board of Directors Excellence Award.

Glacier decapitation and divergent behavioural trajectories: multi-sensor U-Net segmentation of 22,155 glaciers in western North America (1984–2025)

Original title, in French

Décapitation glaciaire et trajectoires comportementales divergentes : segmentation U-Net multi-capteurs de 22,155 glaciers dans l’Ouest de l’Amérique du Nord (1984–2025)

Dataset on ZenodoSSRNglacier-fsnow-unetGlacierScopeHugging FaceORCID