A Python tool for analyzing MapSPAM 2020 global crop production data. Given any administrative region (country, state, or district), it computes crop production statistics from GeoTIFF raster data.
- Location Analysis — Select any country/state/district and see top crops with irrigated vs rainfed breakdown
- Crop Rankings — Rank regions by production, harvested area, or yield for any crop, with choropleth maps
- Global Comparisons — Compare a crop across all countries with side-by-side production, area, and yield charts
- AI Assistant — Ask questions about the SPAM dataset and methodology (powered by Claude)
- 46 crops, 4 variables — Production, Harvested Area, Physical Area, Yield across all technology levels
| Level | Regions | Countries |
|---|---|---|
| Countries | 45 | 45 |
| States/Provinces | 1,309 | 55 |
| Districts | 26,727 | 57 |
Pre-built parquet indexes enable instant queries (sub-second) for all indexed regions.
- Python 3.11+ (tested with 3.14)
- Conda recommended for geospatial dependencies
# Create conda environment with geospatial packages
conda create -n geo python=3.12 rasterio rasterstats geopandas -c conda-forge
conda activate geo
# Install remaining dependencies
pip install typer[all] rich pygadm pyarrow chromadb sentence-transformers anthropic streamlit streamlit-foliumDownload the GeoTIFF ZIP files from mapspam.info/data and place them in data/2020/:
data/2020/
spam2020V2r0_global_production.geotiff.zip
spam2020V2r0_global_harvested_area.geotiff.zip
spam2020V2r0_global_physical_area.geotiff.zip
spam2020V2r0_global_yield.geotiff.zip
Optionally extract the ZIPs for faster processing (the tool uses extracted files when available).
Download GADM administrative boundary data into the local cache:
python -m src.cli init-boundariesPre-compute crop statistics for fast queries:
# Country level (fast, ~30 min)
python -m src.cli build-index --level 0 --parallel 8
# State level (~2 hrs)
python -m src.cli build-index --level 1 --parallel 8
# District level (~8 hrs, run overnight)
nohup ./build_remaining_l2.sh > build_level2.log 2>&1 &streamlit run app.pyOpen http://localhost:8501 in your browser.
# List all crops
python -m src.cli crops
# Analyze a location
python -m src.cli location India --top 10
python -m src.cli location Punjab --level 1 --var yield
# Show crop rankings
python -m src.cli ranking MAIZ --level 1src/
crops.py — 46 crop codes, tech levels, variables, filename parser
boundaries.py — GADM boundary cache + custom overrides
raster.py — GeoTIFF reading via /vsizip/, zonal statistics
index.py — Parquet index building (batch + parallel)
analyzer.py — Orchestrator: on-the-fly + index-first queries
formatter.py — Rich CLI output, CSV/JSON export
cli.py — Typer CLI with 6 commands
rag.py — RAG pipeline for AI assistant
faq.py — Curated FAQ content
app.py — Streamlit dashboard (4 tabs)
- Read from ZIP via
/vsizip/— avoids extracting 20GB of GeoTIFFs (falls back to extracted files when available) rasterio.maskwithcrop=True— reads only the bounding box, not the full 37MB global raster- Parquet indexes — pre-computed zonal stats for instant queries
- Batch zonal stats — one raster read per crop for all boundaries (700x fewer I/O operations)
- Weighted average yield —
sum(Y x H) / sum(H)per SPAM methodology - Pluggable boundaries — custom shapefiles override GADM per country
- Get an API key from console.anthropic.com
- Add it to
.streamlit/secrets.toml:ANTHROPIC_API_KEY = "sk-ant-..."
- Build the knowledge base:
python -m src.cli build-knowledge
make test # Run pytest
make lint # Run ruff
make check # Both (gate before commit)MapSPAM 2020 — Spatial Production Allocation Model by IFPRI. The GeoTIFFs contain gridded crop production data at ~10km resolution globally, covering 46 crops across irrigated and rainfed systems.
Data: MapSPAM is licensed under CC BY 4.0. Code: MIT.