Unexploded Ordnance Detection in Hyperspectral Images by Using Deep Neural Networks

Unexploded Ordnance (UXO) is a major threat affecting the lives of people in more than 60 countries. This work tests deep neural networks to automatically detect UXO in Hyperspectral Images (HSI). Initially, we constructed our own dataset of 134 HSI cubes divided into three folds: two for training and one for validation. U-Net was selected through preliminary experiments as the most promising detection method among those compared. Customised loss functions were designed for the U-Net, resulting in 3 different models. These models were trained and validated in a supervised manner on our data. The results obtained are very promising with a UXO detection rate of around 70% and an F1 score above 0.8.