SVAMITVA AI System Architecture & Model Pipeline
Detailed technical workflow showing production U-Net / EfficientNet deep learning integration and prototype computer vision abstraction layer.
End-to-End Processing Workflow
Drone Orthophoto → AI Multi-Class Feature Extraction → GIS Delivery
Ingests ultra-high-resolution aerial drone imagery in JPG, PNG, TIFF, or GeoTIFF format. Extracts ground sampling distance (GSD), image dimensions, and CRS projection metadata.
Normalizes RGB reflectance channels and splits large gigapixel orthophotos into overlapping 512x512 pixel patches for deep learning batch inference without memory bottlenecks.
Multi-class deep convolutional network predicting per-pixel class masks for Buildings, Roads, and Waterbodies. Prototype uses HSV color space analysis, Canny edge detection, and morphological watershed.
Crops individual candidate building footprints and runs fine-grained roof material classification into RCC (Concrete), Tiled (Terracotta), Tin (Corrugated Sheet), or Other.
Converts pixel mask contours into clean spatial vector geometries (Polygon / LineString), computes footprint areas (m²), road corridor lengths, and centroid coordinates.
Manages asynchronous background processing jobs, serves real-time status polling endpoints, stores feature geometries, and manages human review approvals.
Presents interactive vector GIS map overlays, building detail inspector drawers, human-in-the-loop review queues, temporal change detection tools, and downloadable PDF reports.
Production Model Modular Plug-in Design
The backend architecture uses Python Abstract Base Classes (SegmentationModel and RoofClassificationModel). The prototype current runs DemoSegmentationModel and DemoRoofClassificationModel. When trained weights (e.g. PyTorch U-Net or EfficientNet weights trained on SVAMITVA village orthophotos) are placed in backend/app/ml/weights/, the engine seamlessly switches without modifying the API contracts or frontend.