PS-26015.
Geospatial Watershed Intelligence.
Application of geospatial techniques for visualization and analysis, interpreting geo-coded images to improve watershed development outcomes. India runs large watershed-development programs — check dams, percolation tanks, afforestation — in drought-prone rural areas. Verifying whether they're working currently relies on manual field visits: slow, expensive, and doesn't scale to thousands of sites.
Model 1 — U-Net, ResNet18 encoder
Reads a 6-channel stack (R, G, B, NIR, NDVI, NDWI) and classifies every 10m Sentinel-2 patch into one of 7 land-cover types. Trained on 4 real sites chosen to cover the classes any single site lacked.
Change Detection — Tier-1 rule matrix
A rule-based pixel-diff of two Model 1 passes across the same area — no secondary training needed, and every output class traces back to an auditable rule.
Recommendation Engine
Pure if-then rules over the change map, condition score, and NDVI trend. No dataset for "recommendations" exists, and a black-box output wouldn't be trusted by a government official — so every alert has a specific, inspectable reason.
Field Verification
In-browser EXIF GPS extraction from a geo-tagged photo, matched against the satellite-predicted class at that exact coordinate, with a human confirm/mismatch verdict logged alongside it.
Kadwanchi Watershed
Indo-German Watershed Development Programme site (1997–2002, 1888 ha). Primary site — has a real T1→T2 pair, drives the change-detection demo.
Tamhini Ghat Forest
Western Ghats dense-forest site, added to cover the dense-vegetation class. Single-date, training only.
Donimalai Iron Mine
Exposed-ground mine site, added to cover the barren/degraded class. Single-date, training only.
Jayakwadi Dam
Large reservoir/river site, added to cover rivers and large water bodies. Single-date, training only.