One mine, two maps: label conventions change which satellite segmentation model wins, and a convention-conditioned network serves both

Public inventories of mining land disagree about where a mine ends. The two global polygon sets most often used to train satellite models follow different delineation conventions: Maus et al. (2022) enclose the land between mine features, and Tang and Werner (2023) trace each feature. We ask what this difference does to learning and to evaluation, and whether one model can serve both conventions. On 722 Sentinel-2 tiles that carry both label sets, the two conventions overlap at a median IoU of 0.655, and the Tang area is 78% of the Maus area. A U-Net trained on one convention loses when it is scored on the other, and the ranking of the two single-convention models reverses with the reference: +0.062 IoU for the Maus-trained model on Maus labels and −0.027 on Tang labels (five seeds; both intervals exclude zero). We then test a remedy that was specified before the runs. A single network receives the requested convention, either as a second output channel or through feature-wise modulation. On 113 tiles that no model had seen and that were used for no design choice, the conditioned model beats an unconditioned model trained on the same mixed labels (+0.031 IoU), stays within the pre-specified margin of −0.02 against the single-convention model (−0.004), and its matching output beats its other output (+0.075). Its predicted area is within 6% of the label area in each convention, where the unconditioned model is 11% low or 17% high. The gain was detectable on the 86 Maus-only tiles. On the 27 Tang-only tiles the interval included zero. An exploratory zone analysis suggests why: the network uses the convention as an inclusiveness setting, not as a second object definition. Under a score that ignores disputed pixels, the ranking reversal is not detected. We close with reporting recommendations for mining benchmarks. Third-party material retains the licences identified in the manuscript.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24369292
Primary Topic
Remote-Sensing Image Classification
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

One mine, two maps: label conventions change which satellite segmentation model wins, and a convention-conditioned network serves both

S. M. Abdulla Al Mamun
Zenodo (CERN European Organization for Nuclear Research)
Remote-Sensing Image Classification
preprint

One mine, two maps: label conventions change which satellite segmentation model wins, and a convention-conditioned network serves both

S. M. Abdulla Al Mamun
preprint en

Abstract

Public inventories of mining land disagree about where a mine ends. The two global polygon sets most often used to train satellite models follow different delineation conventions: Maus et al. (2022) enclose the land between mine features, and Tang and Werner (2023) trace each feature. We ask what this difference does to learning and to evaluation, and whether one model can serve both conventions. On 722 Sentinel-2 tiles that carry both label sets, the two conventions overlap at a median IoU of 0.655, and the Tang area is 78% of the Maus area. A U-Net trained on one convention loses when it is scored on the other, and the ranking of the two single-convention models reverses with the reference: +0.062 IoU for the Maus-trained model on Maus labels and −0.027 on Tang labels (five seeds; both intervals exclude zero). We then test a remedy that was specified before the runs. A single network receives the requested convention, either as a second output channel or through feature-wise modulation. On 113 tiles that no model had seen and that were used for no design choice, the conditioned model beats an unconditioned model trained on the same mixed labels (+0.031 IoU), stays within the pre-specified margin of −0.02 against the single-convention model (−0.004), and its matching output beats its other output (+0.075). Its predicted area is within 6% of the label area in each convention, where the unconditioned model is 11% low or 17% high. The gain was detectable on the 86 Maus-only tiles. On the 27 Tang-only tiles the interval included zero. An exploratory zone analysis suggests why: the network uses the convention as an inclusiveness setting, not as a second object definition. Under a score that ignores disputed pixels, the ranking reversal is not detected. We close with reporting recommendations for mining benchmarks. Third-party material retains the licences identified in the manuscript.

Zenodo (CERN European Organization for Nuclear Research)
University of Manchester (GB)
Remote-Sensing Image Classification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon