IndiaSAT LULC v4
10 views
Skip to first unread message
Aaditeshwar Seth
Feb 25, 2026, 2:21:06 AMFeb 25
to learning, ac...@googlegroups.com, CoRE stack: NRM, core-st...@googlegroups.com, Raman Kumar
Hi all,
Raman and team had been working on this since a while and this is finally done. We'll be pulling these new improved annual LULCs into the CoRE stack soon. The earlier v3 would make mistakes between crops (often single cropped areas) and shrubland, a common error that occurs even in other products like Dynamic World. The approach we adopted is to sample hi-res (1m resolution) tiles from across India and run computer vision algorithms on it which have gotten really good now at segmentation and object detection tasks.
- We used these to detect farm boundaries and to distinguish between areas in farming plots vs. shrubs.
- Similarly we used object detection to identify agro-horticulture areas like mango orchards.
- We then used these samples from across India to train agroecological zone wise models that work on satellite data to produce wall-to-wall maps.
- The result is much better crops vs shrubs distinction, and a breakup of the tree class into trees in agro-horticulture vs. trees elsewhere (forests, farm bunds, road side, etc.).
- The crops vs shrubs distinction done this way is able to do better than the global pasture maps that came out last year and other global LULCs.
We are also planning to use the ONE maps by Madhu, Pradeep, and team to carve out finer classes in the shrubs category.
And looking forward to new maps from IOLN for long-term historical changes.
Adi
Co-founder, Gram Vaani; Co-founder, CoRE Stack
Reply all
Reply to author
Forward
0 new messages