Do Geospatial Foundation Models Transfer to Mining Safety? A Zero-Shot Computer Vision Benchmark on Tailings Storage Facilities

Foundation models such as DINOv2 and Prithvi-EO have become central to modern computer vision, with self-supervised pretraining promising strong zero-shot transfer to downstream tasks. However, a fundamental question remains unresolved: do domain-specific geospatial foundation models actually outperform general-purpose vision models on specialized Earth observation tasks, or does natural-image pretraining suffice? This work investigates this question in the safety-critical setting of satellite-based tailings storage facility (TSF) monitoring, motivated by recent catastrophic mining dam failures such as Jagersfontein (2022) and Brumadinho (2019). We construct the first labelled TSF benchmark from Anglo American's 2023 public disclosure, comprising 47 facilities across six countries with expert-assigned GISTM consequence class (5-class ordinal) and raising-method (7-class) labels, paired with quarterly cloud-masked Sentinel-2 composites at 10m resolution and six spectral bands (RGB, NIR, SWIR). We benchmark DINOv2 ViT-B/14 (self-supervised on 142M natural images) against Prithvi-EO-2.0-300M (masked autoencoder pretrained on multispectral NASA Landsat-Sentinel imagery) under a common linear-probe protocol, freezing the pretrained encoder and training a class-balanced logistic regression under 5-fold stratified cross-validation. Counter-intuitively, DINOv2 outperforms Prithvi-EO on raising-method classification (macro-F1 0.44 vs 0.34), correctly identifying Upstream construction, historically the highest-risk method, with F1 0.86 versus 0.44 for Prithvi. On GISTM consequence prediction, both models tie at macro-F1 0.30 and fail entirely on rare high-risk classes, collapsing predictions toward majority classes. This suggests that current geospatial pretraining corpora, which emphasize agricultural and natural scenes, do not encode the geometric priors needed for man-made industrial infrastructure, and that downstream infrastructure context beyond a 5km chip is required for consequence classification.