Habitation Boundary Extraction using Geospatial Artificial Intelligence on HPC

Abstract. Artificial Intelligence (AI), High-Performance Computing (HPC) and high-resolution satellite imagery have collectively enhanced the capability to extract meaningful geospatial information quickly and accurately. In this paper, we have proposed an automated workflow for extracting habitation boundary using Deep Learning (DL). The workflow is built on deep learning-based semantic segmentation, augmented by an adaptive contour-refinement loss, optimized to detect small and scattered habitation from satellite images. The pipeline incorporates Spatial Pyramid Dilation Convolution (SPDConv) and Effective Squeeze-and-Excitation (EffectiveSE) in the backbone, combined with a boundary-aware hybrid loss (Dice, weighted BCE, L1 boundary loss). The model was trained and validated on C-DAC PARAM Siddhi-AI HPC system using 6274 samples. The satellite image was sourced from ISRO Bhuvan and the annotations were created on QGIS. Both the satellite image and its corresponding raster annotation were tiled with resolution of 640 x 640 pixels. The training workload was distributed using PyTorch’s Distributed Data-Parallel (DDP) framework, enabling efficient scaling across multiple GPUs and optimizing experimentation workflow on large datasets. The final model attained 0.757 precision, 0.669 recall, and 0.7 Intersection-over-Union (IoU) on the validation set, indicating reliable performance in extracting habitation boundaries. The predicted segmentation masks are post-processed to create geospatial polygons of habitation boundaries. The output can be seamlessly integrated with national geospatial programmes like PMGSY National GIS.