Multi-Scale ViT Inference with Habitat-Fit Priors and kNN Retrieval for Multi-Species Plant Identification

This paper presents DS@GT ARC’s third-place solution for PlantCLEF 2026, where models must identify every plant species in roughly 3000×3000-pixel vegetation quadrat images while training data contains only single-label individual-plant images. The system fine-tunes DINOv2 ViT-L/14 and combines multi-scale tile inference, FAISS kNN retrieval, source-aware temporal fusion across repeated visits, habitat-fit demotion using geographic and altitude statistics, and a Southwestern Europe mask. The selected submission achieved a private-leaderboard macro-F1 of 0.43902 and a public score of 0.51096. Ablations identify habitat-fit demotion and multi-scale aggregation as the largest contributors.