Hybrid U‑Net Benchmark Suite

Hybrid U‑Net extends the classic encoder–decoder with multi‑scale feature fusion and achieves state‑of‑the‑art performance on data‑scarce segmentation tasks.
This repository accompanies the paper

Nabiee S. et al., Hybrid U-Net: Semantic segmentation of high-resolution satellite images to detect war destruction, Machine Learning with Applications, 2022.

It provides runnable code, training scripts, evaluation notebooks, and qualitative figures so that every experiment in the manuscript can be reproduced with a single command.

Why another segmentation zoo

• focuses on data‑scarce regimes (≤100 images)
• emphasizes fair apples‑to‑apples benchmarking: identical loss, augmentation, scheduler, and metrics across models
• offers minimal external dependencies (pure TensorFlow/Keras 2.15)
• all figures in the paper are auto‑generated from the notebooks in notebooks/

Models implemented

key	paper name	params	default backbone
hybrid_u	Hybrid U‑Net (ours)	3.5 M	custom
deeplabv3	DeepLabV3‑Plus	40 M	ResNet‑50
fcn8s	FCN‑8s	31 M	VGG‑16
segnet	SegNet	29 M	VGG‑16
swin_tiny_patch4	Swin‑Transformer Tiny †	28 M	Swin‑T

† The Swin‑Transformer variant is loaded through timm. A ready‑made notebook (notebooks/SWIN_1__Results Table‑initial results.ipynb) demonstrates how to fine‑tune it and export weights for the CLI.

Installation

python -m pip install -e .

Running an experiment

# train Hybrid U‑Net
segbench train /path/to/dataset --model hybrid_u --epochs 80

# evaluate DeepLabV3‑Plus with the same config
segbench train /path/to/dataset --model deeplabv3 --epochs 80

Single‑image demo

segbench infer sample.png \
        --weights runs/hybrid_u_20250509-1123.h5 \
        --out sample_mask.png

Quantitative results on the XYZ‑Skin‑Lesion dataset

model	Dice	IoU	inference fps (RTX 4090)
Hybrid U‑Net	0.912	0.848	185
Swin‑T	0.908	0.842	110
DeepLabV3‑Plus	0.905	0.836	72
FCN‑8s	0.871	0.795	148
SegNet	0.864	0.782	160

Qualitative comparison

Zoom‑in heat‑map for Hybrid U‑Net and other models segmentation results.

Citing this repository

@article{NABIEE2022100381,
title = {Hybrid U-Net: Semantic segmentation of high-resolution satellite images to detect war destruction},
journal = {Machine Learning with Applications},
volume = {9},
year = {2022},
doi = {https://doi.org/10.1016/j.mlwa.2022.100381},
url = {https://www.sciencedirect.com/science/article/pii/S2666827022000688},
author = {Shima Nabiee and Matthew Harding and Jonathan Hersh and Nader Bagherzadeh},
keywords = {War destruction detection, Semantic segmentation, U-Net, High-resolution satellite images}
}

License

Copyright (c) 2025, The Regents of the University of California. Released under the BSD 3‑Clause license. See the LICENSE file for the full text.