Few-Shot Classification on EuroSAT Sentinel-2 Imagery

Fuente: Zenodo
Gespeichert in:
Bibliographische Detailangaben
1. Verfasser: Fouilloux, Anne
Format: Recurso digital
Veröffentlicht: Zenodo 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866901705414672384
author Fouilloux, Anne
author_facet Fouilloux, Anne
contents <h2>Few-Shot EuroSAT Within-Domain v0.1.0</h2> <p>First release of the within-domain few-shot classification experiment on EuroSAT Sentinel-2 imagery.</p> <h3>What's included</h3> <ul> <li>Prototypical Networks implementation (PyTorch, MPS-compatible)</li> <li>Jupytext notebook with full experiment</li> <li>Snakemake pipeline for reproducibility</li> <li>Dockerfile for containerized execution</li> <li>Research software metadata (CITATION.cff, codemeta.json)</li> </ul> <h3>Results (5-way, 600 episodes)</h3> <p>| Setting | Accuracy | |---------|----------| | 1-shot | 71.5% ± 1.0% | | 5-shot | 82.1% ± 0.8% | | 20-shot | 84.3% ± 0.7% | | Novel-only 3-way 5-shot | 53.8% ± 0.7% |</p> <h3>Replication context</h3> <p>Tests whether Prototypical Networks (Snell et al., NeurIPS 2017) can transfer from common to rare land cover types on Sentinel-2 imagery, simulating a Natura 2000 rare habitat monitoring scenario.</p> <p>Part of the <a href="https://platform.sciencelive4all.org">Science Live</a> FORRT replication initiative.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19607663
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Few-Shot Classification on EuroSAT Sentinel-2 Imagery
Fouilloux, Anne
few-shot learning
prototypical networks
remote sensing
Sentinel-2
EuroSAT
land cover classification
meta-learning
replication study
<h2>Few-Shot EuroSAT Within-Domain v0.1.0</h2> <p>First release of the within-domain few-shot classification experiment on EuroSAT Sentinel-2 imagery.</p> <h3>What's included</h3> <ul> <li>Prototypical Networks implementation (PyTorch, MPS-compatible)</li> <li>Jupytext notebook with full experiment</li> <li>Snakemake pipeline for reproducibility</li> <li>Dockerfile for containerized execution</li> <li>Research software metadata (CITATION.cff, codemeta.json)</li> </ul> <h3>Results (5-way, 600 episodes)</h3> <p>| Setting | Accuracy | |---------|----------| | 1-shot | 71.5% ± 1.0% | | 5-shot | 82.1% ± 0.8% | | 20-shot | 84.3% ± 0.7% | | Novel-only 3-way 5-shot | 53.8% ± 0.7% |</p> <h3>Replication context</h3> <p>Tests whether Prototypical Networks (Snell et al., NeurIPS 2017) can transfer from common to rare land cover types on Sentinel-2 imagery, simulating a Natura 2000 rare habitat monitoring scenario.</p> <p>Part of the <a href="https://platform.sciencelive4all.org">Science Live</a> FORRT replication initiative.</p>
title Few-Shot Classification on EuroSAT Sentinel-2 Imagery
topic few-shot learning
prototypical networks
remote sensing
Sentinel-2
EuroSAT
land cover classification
meta-learning
replication study
url https://doi.org/10.5281/zenodo.19607663