Training strategy over architecture: A systematic ablation study for Spartina detection from aerial imagery with small geospatial datasets

Adrien Le Guillou, Manon Brehier, Jérôme Ammann, Xavier Dauvergne, Pierre Stéphan · 2026

Read Training strategy over architecture: A systematic ablation study for Spartina detection from aerial imagery with small geospatial datasets by Adrien Le Guillou, Manon Brehier, Jérôme Ammann, Xavier Dauvergne, Pierre Stéphan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/plos-pone-0358464.txt.

Description

Adrien Le Guillou leads a team of five with a lesson for everyone training image models on small maps, in PLOS ONE in 2026. Their task was spotting invasive cordgrass in aerial photographs of tidal marsh, where labeled examples are inherently scarce. Three respected segmentation architectures were each trained under twelve configurations, varying how the training itself was run, for many runs judged by cross validation.

Training strategy crushed architecture. The spread between the best and worst training recipes was about ten points of overlap accuracy; the spread between the three architectures was under one point. Two popular tricks, freezing the encoder and a fashionable loss function, hurt every architecture they touched. The winning recipe mapped the whole Bay of Brest, estimating some forty-four hectares of cordgrass. Elevation data, assumed precious, contributed little beside the color imagery. Data scarce fields should tune their teaching before shopping for bigger models.

Catalog record

CREATORAdrien Le Guillou, Manon Brehier, Jérôme Ammann, Xavier Dauvergne, Pierre Stéphan
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE112703
LICENSECC_BY
RIGHTSCC BY
SOURCEPLOS

Source: https://doi.org/10.1371/journal.pone.0358464

Browse the full library · About the Tower of Babel