pangruncheng/HiASTF

HiASTF: Seismic Dataset for ASTF Inversion

Dataset Summary

HiASTF is a seismic dataset designed for studying earthquake source time function (ASTF) inversion using Empirical Green’s Functions (EGFs). It provides paired waveform data suitable for deep learning models, where the goal is to recover ASTFs from observed seismic signals.

The dataset includes:
• Empirical Green’s Functions (EGFs)
• Target waveforms (convolution of EGF and ASTF)
• Ground-truth ASTFs

Two main dataset variants are provided: EGFs_new_3_with_ASTFs_vr_ratio2_min0: Original dataset without duration balancing.
EGFs_new_3_with_ASTFs_vr_ratio2_min100: Augmented dataset where ASTFs are balanced such that each duration bin contains at least 100 samples.

Additionally, a test_set is provided for evaluation.

Dataset Structure

1. Training and Validation Sets

Each dataset folder contains:
• Training files: Train_new_3_pairs_*.h5
• Validation files: Validation_new_3_pairs_*.h5

Each .h5 file stores paired data:
• Input: EGF waveform + target waveform
• Output: ASTF (source time function)

2. File Naming Convention

Example: Train_new_3_pairs_Samezero256_normalize_vr_ratio2_min0_2.h5
Explanation:
• Samezero256 → Fixed-length waveforms (256 samples, zero-padded)
• normalize → ASTF is area-normalized
• (no normalize) → ASTF is raw (non-normalized)
• ratio2 → ASTF:EGF=2:1
• min0 / min100 → Dataset version

3. Test Set

The test_set folder contains three subsets:
• Test_level1_*
• Test_level2_*
• Test_level3_*

Each level represents different testing conditions or difficulty levels.

Dataset Variants

Dataset Description min0 Original dataset without ASTF duration balancing min100 Balanced dataset with ≥100 samples per duration bin

The min100 dataset is constructed to alleviate the imbalance of ASTF duration distribution, especially improving coverage of long-duration events.

Normalization
• Files with normalize: ASTFs are area-normalized
• Files without normalize: ASTFs are raw (non-normalized)

Normalization is applied only to ASTFs and does not affect waveform alignment.

Usage in Paper

The experiments in the associated study use:
• Training & Validation: EGFs_new_3_with_ASTFs_vr_ratio2_min0
• Testing: Normalized datasets from test_set

The min100 dataset is used as an augmented dataset for additional experiments on improving model generalization.

Task Description

Input:
• Empirical Green’s Function (EGF)
• Target waveform

Output:
• ASTF (source time function)

This is a supervised learning problem for seismic source inversion.

Data Characteristics
• All waveforms are aligned and zero-padded to fixed length.
• ASTFs exhibit significant variability in duration.
• The dataset is designed for convolution-based physical modeling: Target waveform ≈ EGF * ASTF
• The min100 variant reduces duration imbalance across samples.

⚠️ Notes and Recommendations
• Use normalized ASTFs (normalize) when training neural networks for stability.
• The min0 dataset reflects natural data distribution.
• Ensure consistency between training and testing normalization settings.