SDFFNet A Spatial-Frequency Domain Feature Fusion Network for Road Extraction From Very High-Resolution Satellite Images

Road extraction from very high-resolution satellite images is difficult due to occlusions, cluttered backgrounds, and the thin elongated shape of roads. This project proposes two deep learning segmentation models – DeepAttnLab and PSPNet – for binary road segmentation from 256×256 image patches. Both models share a ResNet‑50 backbone with stride modification to preserve spatial detail. DeepAttnLab uses attention mechanisms to suppress irrelevant background regions, while PSPNet captures multi‑scale global context through a pyramid pooling module. A combined loss function (Binary Cross‑Entropy, Dice Loss, and Focal Loss) addresses severe class imbalance between road and background pixels. An existing network (SDFFNet) is also trained as a baseline under identical conditions. All models are trained using the AdamW optimizer with cosine annealing. After evaluation, the best performing model is deployed in a Flask web application. The web app includes user registration, login, a relevance classifier to filter non‑satellite uploads, and real‑time road mask prediction. Experimental results show that both proposed models achieve accurate and continuous road extraction, outperforming the baseline and traditional methods.

Keywords: Road extraction, very high-resolution satellite imagery, DeepAttnLab, PSPNet, attention mechanism, pyramid pooling, class imbalance, combined loss, Flask web application.