ahmdtrdi/water-quality-prediction

Machine Learning Models for prediciting water at South Africa # EY Water Quality Prediction — AI Data Challenge 2026 Predicting **3 water quality parameters** (Total Alkalinity, Electrical Conductance, Dissolved Reactive Phosphorus) for **unseen river locations** across South Africa. > **Evaluation metric:** Mean R² across all 3 targets. ## 🏗️ Project Structure ``` water-quality-prediction/ ├── config/ │ ├── base.yaml # Pipeline configuration │ └── feature_sets.yaml # Per-target selected features (from notebook 03) ├── data/ │ ├── 01-raw/ # Immutable source CSVs (from EY) │ ├── 02-processed/ # Cleaned parquet files │ └── 03-external/ # External features (generated on Kaggle) ├── docs/ │ ├── DEVLOG-DS.md # Experiment log (R² per experiment) │ ├── AGENTS-DataScientist.md # Data science methodology guidelines │ ├── 2026_EY_AI_... # Challenge guidance │ └── ey_winners_approach.md # Past winners' reference ├── notebook/ │ ├── kaggle/ # 8 Kaggle-compatible notebooks (main pipeline) │ ├── ey_provide/ # EY-provided benchmark notebooks (reference) │ └── custom_snowflake_archive/ # Archived Snowflake-based notebooks ├── src/ │ ├── data_ingestion.py # API calls, checkpoint logic │ ├── feature.py # Feature engineering utilities │ ├── spatial_cv.py # Spatial cross-validation (LeaveStationGroupOut) │ ├── ensemble.py # OOF stacking ensemble framework │ └── utils.py └── requirements.txt ``` ## 🧪 Experiment Batching Strategy Every dataset must **earn its place** through measured R² improvement: | Exp | Name | Datasets Added | Purpose | |-----|------|----------------|---------| | 0 | Baseline | Water Quality only | Floor: naïve baselines | | 1 | EY Core | + Landsat + TerraClimate | Provided satellite features | | 2 | External APIs | + SoilGrids + Weather + Elevation + OSM | Domain-driven features | | 3 | Spatial Co …