Bangladesh Geospatial Analysis Platform

Satellite-based geospatial analysis of Bangladesh built on Google Earth Engine, covering water and flooding, river morphology, urbanisation, land cover, air quality, climate, agriculture, coastal change, and a set of socio-economic proxies. Analyses run at national scope, per division, or per district, over records reaching back to 1975.

The platform ships 28 Earth Engine analysis modules behind a single CLI, two more that run offline on downloaded rasters, and a validation subsystem that scores selected indicators against independent ground truth and assigns each one a quality tier.

Not every output carries the same evidential weight, and the distinction matters more than the module count:

• Measured. Quantities read more or less directly off the sensor: water extent, built-up area, NDVI, land surface temperature, pollutant column densities, nighttime radiance.
• Proxy. Composite indices with no direct satellite observable behind them: the poverty index, slum index, health risk index, energy potential, erosion susceptibility. These are weighted combinations of measured layers. They are useful for ranking and screening, not for reporting levels.
• Literature-based. Arsenic hotspots, cyclone landfall points, and known slum locations are digitised from published sources, not detected from imagery.

Only indicators listed in validation/registry.py have been scored against independent reference data. Everything else is uncalibrated. See Validation.

Requires Python 3.10 or newer and a Google Cloud project with the Earth Engine API enabled.

git clone https://github.com/deluair/bd_gis.git
cd bd_gis
pip install -r requirements.txt

Earth Engine authentication, needed once per machine:

# 1. Register a Cloud project for Earth Engine at
#    https://code.earthengine.google.com/register
# 2. Authenticate:
earthengine authenticate
# 3. Point the pipeline at your project:
export GEE_PROJECT=your-cloud-project-id

GEE_PROJECT defaults to the original author's project, which you will not have access to, so setting it is effectively required. Scope can also be preset with BD_GIS_SCOPE.

python run_pipeline.py                              # default: quick water test
python run_pipeline.py --scope sylhet --test        # fastest scope, good smoke test
python run_pipeline.py --nightlights                # a single module
python run_pipeline.py --scope sylhet --floods      # a single module, scoped
python run_pipeline.py --district Comilla --poverty # single district
python run_pipeline.py --full                       # water modules
python run_pipeline.py --full-extended              # 23-module pipeline

Start with --scope sylhet. National-scope runs cover roughly 148,000 km2 and can take hours.

National scope switches water thresholding from Otsu to fixed thresholds, because Otsu times out at that extent.

Reference geography carried in config.py: 16 rivers, 19 wetlands and haors, 10 urban centres, 7 economic zones, 14 coastal districts, 8 extreme flood years (1987, 1988, 1998, 2004, 2007, 2017, 2020, 2022), 9 known slum areas, 8 arsenic hotspots, and 8 cyclone landfall points.

--full-extended runs 23 of these in three dependency-ordered parallel waves. It does not include cyclones, aquaculture, chars, timelapse or alerts; run those individually.

These parse household survey and census microdata and do not call Earth Engine. Input files are not distributed with the repository.

validation/ scores an indicator's per-unit output against independent reference data and writes a JSON validation card recording the statistics, the reference citation, and any caveats.

Quality tiers, from validation/registry.py:

Registered indicators:

The registry records a standing caveat on the optical comparison: JRC is itself Landsat-derived, so agreement there is partly circular and is not a fully independent validation. The SAR comparison is genuinely cross-sensor.

python -m validation.run gis_poverty_index \
  --predicted-csv outputs/poverty/poverty_by_division.csv \
  --unit-col division --value-col poverty_index

A test guard (tests/test_coverage.py) requires every registered indicator to have an implemented reference loader.

Everything lands in outputs/, which is gitignored. One subdirectory per domain (rivers/, floods/, haors/, changes/, nightlights/, urbanization/, vegetation/, landcover/, airquality/, climate/, poverty/, infrastructure/, crops/, slums/, coastal/, soil/, health/, energy/, report_maps/), holding CSV time series, GeoTIFFs, interactive HTML maps, and 300 DPI PNG figures.

Water detection. NDWI, MNDWI and AWEI combined by majority vote, with Otsu auto-thresholding at sub-national scope and fixed thresholds nationally.

SAR flood detection. Sentinel-1 VV backscatter below a conservative -17 dB threshold. The threshold biases toward under-detection.

Urban heat island. Urban core MODIS LST minus surrounding rural ring LST, with QA masking applied first, since LST fill values otherwise corrupt every statistic.

Rice phenology. Monsoon flooding (LSWI > 0) followed by NDVI greening (> 0.4), intersected with a cropland mask, per season: aman (Jul–Nov), boro (Dec–May), aus (Mar–Aug).

Erosion susceptibility. RUSLE-shaped combination of rainfall erosivity, soil erodibility, slope and vegetation cover. The output is a relative index, not a quantitative soil loss rate.

Composite indices. Poverty, slum, health risk and energy indices are normalised weighted sums of their inputs. Weights are heuristic and labelled as such in the source.

• DMSP-OLS digital numbers (0–63) and VIIRS radiance are not comparable across the 2013/2014 boundary. compute_light_change raises rather than silently comparing them.
• GHSL epochs are 5-year; a request for 2017 snaps to 2015 or 2020, and the snap is logged.
• WorldPop ends at 2020, Sentinel-5P starts late 2018, Dynamic World starts 2015, GRACE mascon ends 2017.
• Slum mapping at 30 m Landsat resolution is a proxy, not identification of specific settlements.
• Arsenic zones are literature-based buffers, not satellite-derived.
• The pollutant stack mixes incomparable units and is a relative index only.
• estimate_buildup_density measures built-up area, not road length.
• FAO GAUL administrative boundaries do not exactly match official Bangladesh boundaries.
• Timeouts use signal.SIGALRM, which is Unix and macOS only.
• The largest braided rivers (Padma, Jamuna, Meghna, Brahmaputra) can return zero erosion because their channels exceed the analysis buffer width.

pip install -r requirements-dev.txt
pytest -q          # 26 tests, no Earth Engine credentials required
ruff check .

CI runs ruff, the test suite on Python 3.11 and 3.12, and an install-and-import check that imports every module against a clean requirements.txt.

The layout is deliberately flat: analysis modules live at the repository root and import config as cfg. See CLAUDE.md for conventions and docs/CODEBASE_LEDGER.md for architecture truths, known quirks, and open issues.

@software{hossen_bd_gis,
  author  = {Hossen, Md Deluair},
  title   = {Bangladesh Geospatial Analysis Platform},
  url     = {https://github.com/deluair/bd_gis},
  license = {MIT}
}

Cite the underlying datasets separately. Each carries its own attribution requirements, and the Earth Engine catalogue entry for each collection states them.

MIT, see LICENSE. The license covers this source code only, not the satellite datasets it reads or any survey microdata you supply.