Seismic Event Detection in the Valencian Community Seismic Network Using the Wavelet Scattering Transform, PCA, and SVM: A Robust Lightweight Alternative to Deep Learning for Local Seismic Networks

Southeastern Spain is a region of moderate-to-high seismic risk, where reliable seismic event detection is a prerequisite for earthquake early warning systems. This paper presents a seismic event detection pipeline based on the Wavelet Scattering Transform (WST) with a ZNE component-feature ensemble, Principal Component Analysis (PCA), and a Support Vector Machine (SVM), developed for the Valencian Community Seismic Network (SISCOVA) in Spain. The pipeline is evaluated on a local dataset of labeled three-component traces spanning January 2025 through mid-March 2026 and benchmarked against a classical recursive STA/LTA detector and three state-of-the-art pretrained deep learning models evaluated in a zero-shot transfer setting. Under known signal-to-noise ratio conditions, the proposed approach achieves an accuracy and F1-score of 98.82%, outperforming both the STA/LTA baseline and all evaluated deep learning models. Qualitative validation on continuous full-day recordings from the SISCOVA network further confirms its operational viability. These results demonstrate that WST-based feature extraction, combined with classical machine learning, provides an effective and computationally efficient alternative for seismic event detection in local seismic networks.