Projects / Multimodal classification of natural disasters
Multimodal classification of natural disasters
A system that combines tabular seismic data and synthetic audio signals to detect natural disasters (earthquakes, tsunamis…) automatically.
- Period
- Oct. – Dec. 2025
- Context
- MSc Language & Computer Science (year 1), Sorbonne Université
- Supervisors
- Laurence Devillers, Nour Ben Chaabene
- Team
- with Mamadou Diouhé Barry
The problem
A natural disaster leaves traces of different kinds: structured seismic measurements and sound signals. Each source on its own is incomplete. The question: does combining these modalities improve detection over the best unimodal model?
The method
Seismic data+Audio signals→StandardScaler preprocessing→Unimodal models→Fusion concatenation · stacking · voting
- Unimodal modelling: Naive Bayes, SVM, Random Forest, Gradient Boosting and KNN, compared on each modality.
- Multimodal fusion: early fusion by concatenating features, then late fusion by stacking and voting.
Results
88.9%accuracy, best unimodal model (Random Forest)
0.886F1-score
Multimodal fusion improves class separability and outperforms the unimodal approaches.
Python · scikit-learn · pandas · NumPy