Use, reduce, or omit radar? Locally calibrated mine-footprint mapping in Ghana and Côte d'Ivoire
Year: 2026
Type: preprint
Abstract: Can 32 locally labelled tiles per country tell us whether radar improves a mine-footprint map? We fitted optical-only and optical-plus-radar forests separately in Ghana and Côte d'Ivoire. We evaluated each country on two disjoint cohorts of 200 tiles, including tiles with no labelled mining. In Ghana, adding radar increased pooled footprint overlap (intersection over union, IoU) by 0.028 in the first cohort. The follow-up gain was smaller at 0.015, and its 95% geographic-block interval included zero. In Côte d'Ivoire, radar reduced overlap in both cohorts, by 0.030 and 0.034. Calibration selected the full radar model in Ghana and optical alone in Côte d'Ivoire. An exploratory equal blend increased follow-up IoU by 0.024 over optical alone in Ghana and lowered it by 0.018 in Côte d'Ivoire. Even the best Ivorian model detected only 12.6% of the labelled mine area. Both rounds reused the same fitted models and most geographic blocks. How well local calibration selects a model for a new region remains open. Ghana-derived imagery and combined imagery figures are licensed CC BY-SA 4.0. Ivorian-only derived imagery is licensed CC BY 4.0.
Author: Mamun, S M Abdulla Al
Institution: University of Manchester
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236746582
