Seismic Risk Zonation for Disaster Preparedness

ML pipeline combining DBSCAN clustering and SVM classification to identify seismic hotspots and classify risk zones across the Indian subcontinent.

---

Dataset

Property	Value
Source	USGS Earthquake Catalog API
Region	India + bordering fault zones (Nepal, Afghanistan, Myanmar, N. Indonesia)
Period	2014 – 2024 (10 years)
Filter	Magnitude ≥ 3.5
Records	5,933 earthquake events

Bounding box includes neighboring countries deliberately — fault lines in Afghanistan's Hindu Kush and the Nepal Himalayan belt directly affect North India, and northern Indonesia covers the Andaman fault line.

---

Pipeline

extractData.py          → Raw CSV from USGS API
preprocessData.py       → Deduplication, time parsing, date features
featureEngineering.py   → 5 physics-grounded features
dbscan_clustering.py    → K-distance graph + DBSCAN
riskLabelGeneration.py  → Weighted composite risk score → Low/Medium/High
svm_classification.py   → 3-kernel SVM comparison, 5-fold CV
risk_map.py             → Interactive Folium maps

---

Feature Engineering

All features are computed at event level from raw properties — no group-level aggregation that would cause leakage into the test set.

Feature	Formula / Method	Rationale
energy_index	log₁₀(E) = 5.24 + 1.44×M (Gutenberg-Richter), normalized	Physics-grounded energy release
depth_score	1 / (depth + 1), normalized	Shallow quakes cause more surface damage
frequency_index	BallTree haversine count within 111 km radius, normalized	Local fault-line activity density
mag_variability	Rolling std (window=50), normalized	Regional seismic unpredictability
composite_score	Weighted blend (energy 35%, depth 25%, frequency 25%, variability 15%)	Pre-SVM risk signal

---

DBSCAN Clustering

• Features used: latitude, longitude only — geographic distance, not a mixed feature space
• Eps selection: K-distance graph (k=5) with elbow detection; validated eps=0.05 (≈30 km radius in standardised space)
• Results: 46 clusters, 453 noise points (7.6%)

Noise points are NOT auto-labeled "Low" — a magnitude-threshold rule assigns their risk independently.

---

Risk Label Generation

Cluster-level composite score (0–100):

Weight	Metric
30%	Average magnitude
20%	Energy index (G-R based)
20%	Cluster size (recurrence)
15%	Depth risk (inverted depth)
10%	Frequency index
5%	Max magnitude

Noise point fallback: M ≥ 6.0 → High, M ≥ 5.5 or depth < 15 km → High, M ≥ 5.0 → Medium, else Low.

---

SVM Classification

Three kernels trained and compared:

Kernel	Test Accuracy	5-Fold CV
Linear SVM	75.65%	74.15% ± 1.26%
Polynomial SVM (deg=3)	78.26%	74.86% ± 0.73%
RBF SVM	78.01%	75.62% ± 1.43%

Best model selected by cross-validation mean: RBF SVM with class_weight="balanced".

Why RBF?

Seismic risk is non-linear — magnitude and depth interact in ways no straight hyperplane can capture. RBF handles non-linear boundaries while remaining robust with gamma="scale".

---

Results

Metric	Value
Clusters found	46
Noise points	453 (7.6%)
High risk events	4,098
Medium risk events	1,572
Low risk events	263
SVM Accuracy	78.01%
SVM CV Score	75.62% ± 1.43%

---

Output Files

File	Description
seismic_risk_map.html	All 5,933 points coloured by risk level, with legend and popups
seismic_heatmap.html	Magnitude-weighted density heatmap
kdistance_graph.png	Eps selection plot
dbscan_clusters.png	Cluster scatter plot
confusion_matrix.png	SVM confusion matrix
kernel_comparison.png	Bar chart comparing all three kernels
cluster_risk_stats.csv	Per-cluster risk scores and labels

---

Technologies

Python · Pandas · NumPy · scikit-learn · Folium · Matplotlib · Seaborn