📊 Analysis Overview

This notebook performs a complete machine learning pipeline for flood risk prediction using geospatial raster features. The analysis is structured into 6 major parts.

---

Part 1 – Data Loading & Exploration

The dataset (Training_Data_16k.csv, ~16k samples) is loaded and explored. Features are categorized into continuous (Elevation, Slope, Rainfall, TWI, SPI, etc.), categorical (LULC, Geomorphology, Lithology, Soil), and one circular feature (Aspect). The binary target variable represents flood risk (0 = no flood, 1 = flood)

---

Part 2 – Data Preprocessing

Invalid/special values (-128, -1, NA, 65535) are replaced with NaN and rows with missing values are removed. Features are then processed based on their type:

• Continuous → Z-score normalization (StandardScaler)
• Categorical → One-hot encoding (with drop_first to avoid the dummy variable trap)
• Circular (Aspect) → Sin/Cos transformation to preserve angular continuity

All processed feature groups are concatenated into a final dataset and saved to CSV files, along with visualizations of distributions before and after transformation.

---

Part 3 – Correlation Analysis

Three correlation methods are applied:

• Pearson correlation between continuous features (flagging pairs with |r| > 0.7)
• Point-Biserial correlation between all features and the binary target — top 20 most correlated features visualized
• Chi-Square test + Cramér's V for categorical features vs. the target, with statistical significance testing

---

Part 4 – Collinearity & Feature Selection

Collinearity analysis identifies highly correlated feature pairs (|r| > 0.8) and computes VIF (Variance Inflation Factor) for all features, categorizing them as no / moderate / high / severe multicollinearity.

Feature selection is performed using four independent methods:

• Mutual Information — measures dependency between each feature and the target
• Random Forest Feature Importance — built-in importance from a trained RF model (100 trees, max depth 10)
• Permutation Importance — measures accuracy drop when each feature is randomly shuffled
• ANOVA F-Statistic — tests statistical difference in feature means between classes

All four methods are normalized and averaged into a Combined Ranking, producing recommended feature sets of 10, 15, and 20 top features. A consensus analysis identifies features that appear in the top 10 across multiple methods.

---

Part 5 – Accuracy Assessment

Twelve different feature sets are benchmarked (Top 10/20 from each individual method + combined sets + all features as baseline). For each set, a Random Forest classifier is trained and evaluated using:

• Accuracy, Precision, Recall, F1-Score
• 5-fold Cross-Validation

Results are compared to identify the best trade-off between the number of features and model performance.

---

Part 6 – Hyperparameter Tuning

The best-performing feature set from Part 5 is used to fine-tune the Random Forest model in two stages:

1. Randomized Search (50 combinations, 3-fold CV) — broad exploration of the hyperparameter space
2. Grid Search (5-fold CV) — fine-grained search around the best parameters from step 1

The baseline, randomized search, and grid search models are compared across all metrics. A confusion matrix and full classification report are generated for the best model, and the final optimized parameters are saved.

---

📁 Output Files

The analysis produces a comprehensive set of saved files including processed datasets, feature ranking CSVs, accuracy assessment results, best model parameters, and 15+ visualization PNGs covering distributions, correlations, VIF, feature importances, and model comparisons.