Prototypical network (ProtoNet) based 1-way 5-shot binary segmentation on LandCover.ai v1, with woodland as the novel class. 4 checkpoints published — one per experiment, each being the best-fold weights from 5-fold cross-validation.

Results

Experiment	Configuration	Checkpoint	fgIoU
1 — Backbone	Swin-T (frozen, ImageNet)	`best_ProtoNet_SwinT_none_fold1_seed-42.pt`	0.603
2 — Layer unfreezing (CNN)	ResNet-50 layer3	`best_ProtoNet_ResNet50_layer3_fold4_seed-42.pt`	0.587
3 — Stage unfreezing (ViT)	Swin-T stage3	`best_ProtoNet_SwinT_ImageNetV1_stage3_fold1_seed-42.pt`	0.605
4 — Pre-training ★	SeCo + ResNet-50 (frozen)	`best_ProtoNet_ResNet50_SeCo_none_fold1_seed-42.pt`	0.656

Key finding: Pre-training source dominates other factors — self-supervised SeCo outperforms all supervised sources, including domain-aligned Million-AID. Δ_pretrain = 0.103 (≈ 2× backbone effect, ≈ 8× fine-tuning effect).

Dataset

LandCover.ai v1 — 41 high-resolution aerial images (25–50 cm/pixel) over Poland. 4 labeled land-cover classes + background. Woodland (33.3%) is held out as the novel class; building, water, road are base classes used for episodic training.

Usage

python

import torch
from huggingface_hub import hf_hub_download
from src.model import ProtoNet  # from GitHub repo

# Backbone must match the checkpoint; pretrained init is overwritten by load_state_dict
model = ProtoNet(
    backbone_name="resnet50",
    pretrained="imagenet_v1",
    unfreeze_from="none",
)

ckpt_path = hf_hub_download(
    repo_id="zmgul/few-shot-woodland-segmentation",
    filename="experiments/checkpoints/best_ProtoNet_ResNet50_SeCo_none_fold1_seed-42.pt",
)
state = torch.load(ckpt_path, map_location="cpu")
model.load_state_dict(state["model_state_dict"])
model.eval()

# 5 support tiles + masks, 1 query tile → binary woodland mask
# See notebooks/05_qualitative_results.ipynb for end-to-end inference example.

Citation

bibtex

@mastersthesis{gul2026woodland,
  author  = {Gül, Zehra Merve},
  title   = {Woodland Segmentation in High-Resolution Aerial Imagery Using Few-Shot Learning},
  school  = {İstanbul University, Institute of Graduate Studies in Sciences},
  year    = {2026},
  type    = {Master's thesis}
}

License

MIT (code). LandCover.ai dataset and external pre-trained weights are subject to their original licenses.