Inspiration : Farmers and agronomists make decisions every planting season with imperfect, fragmented data. Satellite imagery, soil surveys, and weather records contain huge signals about land health — but those signals are rarely packaged into a simple, actionable tool for individual fields.
Turn public satellite data + local metadata into a simple “fertile / not fertile” signal plus practical next steps that farmers, extension agents, insurers, and policy teams can act on.
What it does : AgroRemote ingests satellite imagery (e.g., Sentinel-2), computes vegetation and soil proxies (NDVI, temporal summaries), creates compact vector representations (embeddings), and indexes them in TiDB Serverless for fast semantic search.
How we built it: Data ingestion — Datasets (DEM, soil maps) are downloaded and stored as GeoTIFFs for a chosen region.Tiles are cloud-masked, atmospherically corrected, and NDVI/time-series indices are computed.NDVI rasters are tiled into fixed-size chips and summarized.Each chip is converted into a fixed-size vector using a lightweight encoder (PCA / normalized feature vector or CNN-based embedding). Minimal Streamlit/React UI allows drawing polygons, querying, viewing chips and reports.
Challenges we ran into : Small-sample embeddings,Vector schema & DB compatibility & Noisy satellite data
Accomplishments that we're proud of : Built a full end-to-end PoC pipeline: satellite → NDVI → chips → embeddings → TiDB vector index → searchable UI.
What we learned : Vector search in cloud relational DBs (TiDB Serverless) is powerful for hybrid queries (spatial + semantic), but schema design must be precise (vector dims, index timing).
What's next for AgroRemote : Time-series embeddings and multi-vector columns (image, temporal, soil) to improve search relevance. LLM agent for richer, structured agronomic reports (planting windows, fertilizer quantities). Build a mobile capture feature: farmers can upload phone images to find similar fields and get immediate guidance. Run field pilots with extension agents to collect labeled outcomes and iterate the model. Multi-crop recommendation engine
Built With
- pymysql
- python
- streamlit
Log in or sign up for Devpost to join the conversation.