FastSAM DINOv2 for Mine Vegetation Restoration Assessment Using UAV Multispectral and Terrain Data
Authors
Files
Abstract
Monitoring vegetation recovery on reclaimed mine surfaces requires delineating fragmented plant cover under variable substrate, illumination, and terrain conditions. This paper presents a FastSAM DINO framework that combines candidate region generation, frozen DINOv2 appearance descriptors, and independent multispectral and terrain representations. Reliability-gated fusion integrates the three information streams, while dense refinement supports vegetation boundaries and regions omitted by candidate masks. The framework reports vegetation cover, spectral vigor, and structural patch connectivity as separate components before combining them into a restoration screening score. The reported benchmark comprises six mine domains and five spectral bands, with geographically separated evaluation protocols. Full fusion achieves a mean intersection over union of 86.4%, a boundary F1 score of 82.7%, and a mean absolute cover estimation error of 4.8 percentage points. The corresponding appearance-only values are 80.2%, 75.9%, and 7.6 percentage points. The held-mine comparison yields a mean intersection over union of 82.1%, providing a distinct assessment of geographic transfer. Confidence-based deferral relates prediction error within accepted areas to retained valid area and the associated review demand. These comparisons describe mapping performance and the behavior of the proposed screening indicators. The framework connects region segmentation and heterogeneous information fusion to spatially explicit management outputs through observation validity records, spatial partitioning, and uncertainty reporting. Wider ecological interpretation requires independent observations of restoration condition, and operational deployment requires validation of acquisition controls, indicator weights, and processing variability.