Earthquake Prediction Model

Date: June 19, 2025

Overview

This project develops a neural network model to predict earthquake Magnitude (Richter scale, ~0–10) and Depth (~0–700 km) using spatial-temporal features (Timestamp, Latitude, Longitude). The model aims to support seismic hazard assessment, a critical task in geophysics.

Objectives

• Predict Magnitude and Depth as continuous variables (regression task).
• Achieve target performance:
  • Standardized: MSE < 0.3, R² > 0.7.
  • Original scale: MSE < 0.04 for Magnitude (~0.2 units error), < 2500 for Depth (~50 km error).
• Visualize prediction errors spatially to guide improvements.

Model Description

The model is a deep neural network implemented in TensorFlow/Keras, with separate architectures for Magnitude and Depth:

• Architecture: Functional API with three hidden layers (128 → 64 → 32 neurons, ReLU activation), batch normalization, dropout (0.2), and a linear output layer (1 neuron).
• Input Features: Timestamp (normalized to [0,1]), Latitude, Longitude.
• Preprocessing:
  • Features standardized with StandardScaler.
  • Targets scaled separately for Magnitude and Depth.
• Training:
  • Optimizer: Adam.
  • Loss: Mean squared error (MSE).
  • Hyperparameters tuned via GridSearchCV (neurons=128, batch_size=64, epochs=100).
  • Callbacks: Early stopping (patience=10), learning rate reduction (factor=0.5, patience=5).
• Evaluation Metrics: MSE and R² in standardized space; MSE in original scale.

Installation

pip install -r requirements.txt

Dataset

Data: Earthquake dataset (database.csv)

Usage

1. Clone the repository:

git clone https://github.com/chaw-thiri/earthquake_detection_-_crack_depth_estimation.git

2. Place database.csv in the specified path (C:\Users\chawt\Desktop\earthquake detection).
3. Run the script:

python main.py

Performance

Magnitude:

Best Negative MSE (validation): -0.9909 (MSE ≈ 0.9909 standardized). Test MSE (standardized): 1.0252 Test R² (standardized): 0.0227 MSE (original): 0.1817 (~0.426 magnitude units error)

Depth:

Best Negative MSE (validation): -0.3474 (MSE ≈ 0.3474 standardized). Test MSE (standardized): 0.2481 Test R² (standardized): 0.7541 MSE (original): 3722.4633 (~61.0 km error)

Comparison with Previous Iterations

Iteration	Magnitude MSE (std)	Magnitude R² (std)	Magnitude MSE (orig)	Depth MSE (std)	Depth R² (std)	Depth MSE (orig)
Initial (3-layer, GridSearchCV)	0.9233	N/A	N/A	0.9233	N/A	N/A
Deeper (3-layer, GridSearchCV)	0.8769	0.1504	0.1834	0.8769	0.1504	10789.8740
4-layer (GridSearchCV)	1.0276	0.0203	0.1822	0.4116	0.5920	6176.4036
Simple (2-layer, Sequential)	1.0252	0.0227	0.1817	0.3126	0.6901	4690.6951
Current (4-layer, GridSearchCV)	1.0252	0.0227	0.1817	0.2481	0.7541	3722.4633

Magnitude: Persistently poor performance (MSE ≈ 1.0, R² ≈ 0.02–0.15), no better than predicting the mean. Original MSE (~0.18, ~0.42 units error) is acceptable but far from the target (<0.04). Depth: Significant improvement, surpassing standardized targets (MSE < 0.3, R² > 0.7) and approaching original MSE target (<2500). Outperforms the simpler Sequential model (MSE: 0.3126, R²: 0.6901).

Analysis

Strengths

Depth Prediction: Excellent performance (MSE ≈ 0.25, R² ≈ 0.75) due to spherical coordinates capturing spatial patterns (e.g., subduction zones). The deeper architecture with batch normalization and dropout generalizes well. Preprocessing: Normalizing Timestamp, using spherical coordinates, and separate scalers ensured stability and improved Depth results. Training: GridSearchCV, early stopping, and learning rate scheduling optimized Depth convergence.

Weaknesses

Magnitude Prediction: Near-zero R² (0.0227) and high MSE (1.0252) indicate Timestamp and spherical coordinates lack predictive power for Magnitude, requiring geological features (e.g., fault stress). Depth Limitation: Original MSE (3722.4633, ~61.0 km error) is slightly above the target (<2500, ~50 km). Feature Insufficiency: Lack of geological context limits Magnitude and constrains further Depth improvement.

Comparison to Targets

Magnitude: Far below targets (MSE < 0.3, R² > 0.7, original MSE < 0.04). Depth: Meets standardized targets (MSE < 0.3, R² > 0.7), close to original MSE target (<2500).