Cross-domain evaluation of multi-species and herbicide-damage segmentation with DINOv2 and hierarchical inference

Résumé fourni par la source

We study how domain drift affects plant-species and herbicide-damage segmentation in field trials, and whether a DINOv2 foundation backbone with hierarchical taxonomic inference improves robustness relative to a CNN baseline. The contribution is twofold: (i) a cross-domain evaluation protocol and (ii) a controlled ablation of that DINOv2 + hierarchy stack against alternative backbones, decoders, and training regimes. The shift protocol is: train on multi-year Germany/Spain camera data (BASE, 2018–2020; 14 shared plant species plus damage labels); test under a moderate ground-camera shift (REALITY, 2023; Germany and Spain); then under an extreme aerial shift with three-country geography (DRONE, 2024; Germany, Spain, and the United States). Beyond viewpoint and sensor, DRONE also differs in illumination: flights were typically conducted at early morning and sunset, whereas BASE imagery was acquired under daytime ground-camera conditions. With hierarchical inference, DINOv2 raises species F1 from 0.45 to 0.73 in distribution and stays ahead under shift (camera: 0.56 vs. 0.29; drone: 0.37 vs. 0.18). Hierarchy is essential: without it, DINOv2 falls to 0.47/0.35/0.19 on BASE/REALITY/DRONE and nearly ties the baseline on drone data. Under the drone shift, coarser taxa (family ≈ 0.50, class ≈ 0.77) and binary healthy/damaged remain more usable than fine species or fine damage labels; residual errors are mostly vegetation–soil. Species-level model differences are significant (Stuart–Maxwell, p < 0.001). Foundation features plus hierarchical inference reduce—but do not remove—domain-drift losses in this herbicide-phenotyping setting, and indicate where coarser taxonomic or binary damage readouts are the safer operational outputs.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.