Bangladesh Groundwater Arsenic: Mechanism Partitioning and Screening

Code and model weights for:

Phosphate Dominance, Geochemical Bistability, and Seasonal Dynamics in Bangladesh Groundwater Arsenic: National-Scale Mechanism Partitioning and Rapid Risk Screening

Md Rakib Hasan, Mst Anika Khatun Rupa, Fazla Zawadul Arabi, Shoumik Zubyer, Anwar Zahid

Water Research (2026)

Overview

This repository contains the analysis code and trained model for a national-scale study of groundwater arsenic mechanisms in Bangladesh. We analysed 1,807 groundwater samples from 988 locations across all seven physiographic zones, collected in both dry and wet seasons.

Key findings:

• Phosphate competition is the dominant national-scale arsenic driver (rho = 0.326 vs Fe rho = 0.164)
• Mechanism shifts from phosphate to reductive dissolution below ~80 m (Holocene-Pleistocene boundary)
• Monsoon amplifies arsenic risk in 77% of zone-depth combinations (+13.6 million people exposed)
• Geochemical bistability with a phosphate tipping point at 0.37 mg/L
• ArseniCFM screening model achieves AUC = 0.781 with spatial block CV; 50 samples match full training

Repository Structure

src/
  h1_geochemistry/     # H1: Multi-mechanism analysis (phosphate vs reductive)
  h2_depth/            # H2: Depth-stratified geochemical transitions
  h3_seasonal/         # H3: Seasonal forcing, SEM, entropy diagnostics
  h4_arsenicfm/        # H4: ArseniCFM multi-modal screening model
models/
  arsenicfm_best.pt    # Trained ArseniCFM weights (94K params)
embeddings/
  prithvi_pca64_embeddings.npy      # Prithvi-100M embeddings (1807 x 64)
  sentinel_statistical_embeddings.npy  # Sentinel-2 statistical features (1807 x 64)
  prithvi_pca_model.pkl             # Fitted PCA model (768 -> 64 dim)
figures/
  make_figures.py      # Scripts to reproduce all manuscript figures

Data Availability

The geochemical dataset (bangladesh_groundwater_arsenic_1807samples.csv, 1,807 samples x 45 features) is archived on Zenodo:

DOI will be updated upon acceptance.

Installation

git clone https://github.com/rakibhhridoy/bangladesh-groundwater-arsenic.git
cd bangladesh-groundwater-arsenic
pip install -r requirements.txt

Requirements

• Python 3.11+
• PyTorch 2.0+
• scikit-learn, shap, libpysal, esda
• earthengine-api (for satellite data extraction)
• See requirements.txt for full list

Reproducing Results

H1: Geochemical Mechanism Analysis

python src/h1_geochemistry/main.py

H2: Depth-Stratified Analysis

python src/h2_depth/main.py

H3: Seasonal Forcing Analysis

python src/h3_seasonal/main.py

H4: ArseniCFM Model Training and Evaluation

python src/h4_arsenicfm/main.py

Generate Prithvi-100M Embeddings

python src/h4_arsenicfm/prithvi_encoder.py --install  # first time only
python src/h4_arsenicfm/prithvi_encoder.py --test

ArseniCFM Architecture

Three-encoder multi-modal fusion model (~94,000 parameters):

• Satellite encoder: Prithvi-100M 64-dim PCA embeddings
• Geochemical encoder: 18-feature MLP (32-dim output)
• Context encoder: 4-feature MLP (depth, season, lat, lon; 16-dim output)
• Fusion: Cross-modal attention -> 112-dim representation
• Outputs: As concentration (MSE), WHO exceedance (BCE), mechanism class (CE), depth zone (CE)

Citation

@article{Hasan2026_arsenic,
  title   = {Phosphate Dominance, Geochemical Bistability, and Seasonal Dynamics
             in Bangladesh Groundwater Arsenic: National-Scale Mechanism
             Partitioning and Rapid Risk Screening},
  author  = {Hasan, Md Rakib and Rupa, Mst Anika Khatun and Arabi, Fazla Zawadul
             and Zubyer, Shoumik and Zahid, Anwar},
  journal = {Water Research},
  year    = {2026}
}

License

MIT License. See LICENSE for details.