IndiRoad-Net: An AI-based Architecture for Automatic Road Network Extraction from Indian Satellite Imagery

Segmentation of roads from satellite imagery is particularly challenging in dense Indian urban areas due to canopy occlusion, shadows, and severe class imbalance (road pixels below 5% per tile, rising to 50:1 background-to-road ratio in peri-urban scenes). Existing models trained on Western benchmarks such as DeepGlobe and Massachusetts Roads do not generalise well to Indian imagery conditions including monsoon cloud cover and dense deciduous canopies lining arterial roads. To address this, we propose IndiRoad-Net: a five-stage deep learning pipeline comprising an EfficientNet-B3 encoder, Attention U-Net++ decoder, Partial Convolution-based occlusion handling, topology correction via Zhang-Suen skeletonization and Douglas-Peucker simplification, and GeoJSON export for GIS integration. Trained on 850 DeepGlobe tiles and evaluated on 150 held-out tiles at threshold ? = 0.3, the model achieves Recall of 0.823 and AUC-ROC of 0.950, prioritising road detection completeness for routing applications. Zero-shot transfer to Sentinel-2 imagery over Bengaluru correctly identifies major arterial roads without Indian fine-tuning.