Improving Deep Learning Based UAV Thermal Images Classification By Taking Into Account Climate And Flight Parameters: The Case Of Fully Buried IEDs

---

Résumé

Improvised explosive devices (IEDs) pose a serious threat to military operations. Previous studies have shown that thermal imaging can effectively distinguish between undisturbed soil, disturbed soil, and disturbed soil containing or not containing a buried IED, using batch-normalized VGG11 and VGG16 networks respectively. This study focuses on assessing the added value of contextual data (namely flight altitude, ambient temperature, and sunlight exposure) in improving the classification of thermal UAV images. Flight altitudes were vectorized and processed through a fully connected network. Ambient temperature and sunlight data were modeled as time series and analyzed using Long Short-Term Memory (LSTM) networks. A batch-normalized VGG16 network served as the backbone for thermal image processing. The outputs from all modalities were combined and passed through a fully connected classifier. Experimental results, evaluated using precision and AUC-ROC, show that integrating contextual data significantly enhances the performance of thermal image classification for detecting fully buried IEDs.

Mots-clés

improvised explosive devices, infrared thermography, UAV imagery, thermal image classification, flight altitude, ambient temperature, sun exposure, VGG16, LSTM, time series modeling