Groundwater Fluoride Prediction Using Machine Learning & Fuzzy Logic

A data-driven, intelligent, and scalable framework to analyze groundwater fluoride contamination across India using Machine Learning, Regression Models, and a Fuzzy Inference System (FIS). This system supports early detection of fluoride-vulnerable regions and helps government agencies and water-resource managers make informed decisions.

---

Project Objectives

• Large-Scale Analysis: Evaluates over 16,776 groundwater samples across various Indian states and districts.
• Predictive Modeling: Utilizes advanced regression techniques to estimate precise fluoride concentrations.
• Automated Classification: Categorizes water quality into Safe, Moderate, and High-risk zones using optimized classifiers.
• Interpretability: Employs Mamdani Fuzzy Logic to convert technical data into human-readable risk scores.
• Spatial Visualization: Generates state-level heatmaps and regional analysis for spatial awareness.

---

Dataset Architecture

The dataset comprises physicochemical parameters that significantly influence fluoride mobility within aquifers.

Feature Category	Parameters Included
Physicochemical	pH, EC, TDS, Na⁺, Ca²⁺, Mg²⁺, K⁺, Cl⁻, SO₄²⁻, NO₃⁻, HCO₃⁻
Target Variable	Fluoride concentration (mg/L)
Geospatial	State and District identifiers

---

Data Preprocessing Pipeline

1. Standardization: Normalization of column nomenclature (e.g., standardizing “EC µS/cm” to “EC”).
2. Imputation: Conversion of invalid entries to null values followed by Median Imputation to maintain numeric stability.
3. Risk Labeling: Implementation of WHO drinking water standards for classification.
   • Class 0 (< 1.5 mg/L): Safe
   • Class 1 (1.5–2.5 mg/L): Moderate Risk
   • Class 2 (> 2.5 mg/L): High Risk
4. Scaling: Application of Min-Max scaling to a standard 0–1 range.
5. Class Balancing: Utilization of SMOTE (Synthetic Minority Over-sampling Technique) to resolve dataset imbalances and achieve perfect class parity.

---

Machine Learning Performance

Seven distinct models were evaluated to determine the most effective classifier for fluoride risk.

Model	Classification Type	Accuracy
Random Forest	Ensemble Learning	93% (Top Performer)
XGBoost	Gradient Boosting	High Accuracy
LightGBM	Boosting	Efficiency at Scale
ANN	Neural Network	Pattern Recognition
SVM (RBF)	Kernel-based	Nonlinear Mapping

Regression Analysis (Continuous Prediction)

Model	R² Score	RMSE
Random Forest Regressor	0.273	0.684
Linear Regression	0.218	0.709
SVR	0.174	0.729

---

Fuzzy Logic Inference System (FIS)

The system utilizes a Mamdani-type FIS to handle environmental uncertainty and provide interpretable results.

• Input Memberships: Very Low, Low, Normal, High, Very High.
• Output Risk Scores: Low Risk (< 33), Medium Risk (33–66), High Risk (>= 66).

---

Limitations & Future Scope

Current Constraints:

• Absence of seasonal temporal data.
• Limited to fluoride without accounting for heavy metal or nitrate interactions.
• Exclusion of complex spatial hydrogeological layers.

Future Directions:

• Implementation of GIS-based real-time heatmaps.
• Integration of Deep Learning for enhanced predictive precision.
• Incorporation of SHAP/LIME for model explainability and transparency.

---

Installation & Usage

# Clone the repository
git clone [https://github.com/codemuggle09/AquaRisk](https://github.com/codemuggle09/AquaRisk)

# Navigate into project folder
cd AquaRisk

# Install dependencies
pip install -r requirements.txt

# Launch the dashboard
python -m streamlit run webapp.py