Nelvinebi/hybrid-ai-lstm-xgboost-interpretable-water-quality-prediction

A hybrid AI system combining LSTM, XGBoost, and K-Means to predict and interpret water quality using limited data. It computes WQI, identifies key pollution drivers, and reveals contamination patterns, enabling accurate, explainable, and scalable environmental decision-making for data-scarce regions like the Niger Delta. # 💧 Hybrid AI Model for Water Quality Prediction — Yenagoa, Bayelsa State, Nigeria > A production-ready **hybrid AI pipeline** that integrates **LSTM**, **XGBoost**, and **K-Means clustering** through a **stacking ensemble** to predict, classify, and spatially cluster water quality across 50 georeferenced samples from Yenagoa, Nigeria achieving **R² = 0.95** and **AUC = 0.96**, with full interpretability via SHAP and LIME. -0B3D91?style=for-the-badge) --- ## 📌 Problem The Niger Delta region of Nigeria home to dense communities, industrial activity, and significant petrochemical infrastructure faces severe water quality degradation. Groundwater sources in Yenagoa, Bayelsa State, are particularly vulnerable to contamination from heavy metals, nitrates, and dissolved solids, yet conventional monitoring is costly, temporally inconsistent, and lacks interpretability for local decision-making. Existing machine learning approaches to water quality prediction are typically single-model, ignore spatial heterogeneity, and fail to provide actionable explanations for their outputs. There is a critical need for a **scalable, interpretable, hybrid AI framework** that delivers accurate prediction and meaningful pattern discovery even under data-scarce conditions. --- ## 🎯 Objective - Develop a **hybrid stacking ensemble** combining LSTM, XGBoost, and K-Means for simultaneous regression, classification, and clustering of water quality - Adapt **LSTM to static physicochemical data** through pseudo-sequential encoding to leverage deep learning on small tabular datasets - Compute the **Water Quality Index (WQI)** using WHO weighted arithmetic standards with symmetric pH penalty weighting - Classify samples into five quality classes: **Excellent, Good, Fair, Poor, Unsuitable** - Identify **spatially coherent contamination clusters** using K-Means with optimal k via silhouette scoring and PCA visualization - Provide model **interpretability** via SHAP global …