Akin-Tunde/ml-water-quality

Satellite + in-situ ML pipeline for water-quality classification along the Nigerian coast (Lagos Lagoon and Niger Delta) # ml-water-quality **Machine-learning classification and prediction of coastal water quality along the Nigerian coast — Lagos Lagoon and the Niger Delta — from satellite ocean-colour observations and in-situ measurements.** Author: Akin-Tunde · B.Tech (FUTA, CGPA 4.25/5.00) · Marine science & AI-agent software --- ## 1. Results at a glance | Task | Best model | Accuracy | F1 (weighted) | ROC-AUC | | --- | --- | --- | --- | --- | | Satellite pixel-month water-quality classes (test 2023–2025, demo) | Random forest + SMOTE | 0.99998 | ~1.000 | ~1.000 | | Satellite log₁₀ chlorophyll-a regression (test 2023–2025, demo) | HGB | R² 0.999 | — | — | | In-situ water potability (held-out 20%) | Random forest + SMOTE | **0.652** | **0.537** | **0.679** | The **satellite numbers look too good to be true — because they are**. In the shipped *demo mode* the class labels are quantiles of `chlor_a`, which is itself one of the engineered features, and the synthetic ocean field has strong land/ocean/plume structure, so the task reduces to an almost deterministic readout. This is reported openly everywhere the numbers appear, and it is precisely why the pipeline also trains a **no-proxy variant** (random forest without `chlor_a` or derived optical-ratio features) that is the only model allowed to power maps and regressions. The **in-situ results are real and deliberately unflattering** — a held-out F1 of about 0.54 on a small, noisily measured survey dataset is the honest ceiling, and the project argues (notebook 06) that reporting that ceiling honestly is more scientifically valuable than tuning it away. SMOTE outperforms class-weighting on the in-situ task (F1 0.537 vs 0.504 for the random forest) but makes essentially no difference on the synthetic satellite task — the imbalance there is structural, not sample-based. ## 2. Why this matters Nigeria's coastal waters are among the most heavily stressed on the continent: the Lagos Lagoon receives untreated urban discharge f …