Zero-Shot Neural Architecture Search for National-Scale Forest Segmentation From Sentinel-2 Imagery

Abstract

Large-scale forest monitoring from Sentinel-2 imagery is constrained by the high computa- tional cost of deep-learning model selection on multi-terabyte datasets. This work evaluates zero-shot neural architecture search as a training-free strategy for semantic segmentation in Earth observation. An ensemble of proxy metrics (SynFlow, Fisher Information, Gradient Norm) is applied to a search space of 8,640 Attention U-Net variants, enabling efficient architectural pruning without full training. Validation on a 4.5 TB national Sentinel-2 L1C dataset (Romania) demonstrates a strong rank correlation ( $\rho =0.80$ ) between zero-shot scores and trained performance for the candidate configurations evaluated during HPO. The selected architecture achieves a pixel-wise F1 score on reconstructed maps of spatially disjoint holdout tiles of 0.87, outperforming U-Net, DeepLabV3Plus, and SegFormer. In addition, a systematic analysis of reconstruction and thresholding shows that adaptive thresholds (Yen, Entropy) improve segmentation consistency over fixed heuristics. Overall, the results establish zero-shot NAS as a computationally efficient paradigm for large-scale Earth observation segmentation.