dprindfu/EY-Water-Quality-Data-Challenge

Geospatial ML pipeline predicting river water quality across South Africa using satellite, climate, and terrain data. EY Open Science AI & Data Challenge 2026 — top 10%, R² = 0.44. # Predicting River Water Quality Across South Africa Geospatial ML pipeline for predicting three water quality indicators - Total Alkalinity (TA), Electrical Conductance (EC), and Dissolved Reactive Phosphorus (DRP) - at ungauged river monitoring stations using only publicly available satellite, climate, and terrain data. Built for the **2026 EY Open Science AI & Data Challenge**. Placed in the top 10% out of 300+ teams internationally. Leaderboard R² of 0.44 using zero in-situ water chemistry features. --- ## The Problem South Africa monitors water quality at ~186 river stations, but predictions are needed at locations with no historical chemistry data. Given only a station's coordinates and sample date, predict three targets that are fundamentally driven by geology and land use — information the model has to infer indirectly from remote sensing and terrain. The core difficulty is **spatial extrapolation**. All validation stations sit in the Eastern Cape, geographically separated from training data. Standard random CV gives misleadingly high scores; the real test is whether the model generalizes to entirely unseen catchments with different geology. ## Approach The problem splits naturally into *where* a station is and *when* a sample was taken: **Spatial signal** -- Features that don't change over time: elevation, distance to coast, catchment hydrology (Pitman model parameters, mean annual runoff), Human Development Index, and irrigation extent. These set the baseline chemistry at each location. **Temporal signal** -- Date-specific features: Landsat 8 surface reflectance bands and indices, TerraClimate monthly climate variables, and CHIRPS rainfall. These capture seasonal dilution and concentration dynamics. Each target gets its own model because the three indicators respond to different physical drivers: - **TA → Gradient Boosting.** Alkalinity is geology-dominated. GB's sequential residual fitting extrapolates more smoothly than bagging methods when t …