EFFECTS OF TARGET CLASSIFICATION ON AI-BASED UNEXPLODED ORDNANCE DETECTION PERFORMANCE

This thesis aims to reduce the safety risks for warfighters in an area of operations where unexploded ordnance (UXO) may be present, and lessen the number of training opportunities due to malfunctioning munitions in a controlled environment. The thesis leverages the advancement in unmanned technologies and artificial intelligence (AI) development to complete dull, dirty, and dangerous tasks more effectively. Specifically, the thesis attempts to improve a trained AI detector's performance using different data-labeling methods as applied to the electro-optical images. The thesis describes the efforts conducted to train a UXO detector for a proposed deep learning convolutional neural network followed by validating its performance. To further enhance UXO detection capabilities, the research explores how the optimal target classification method developed and verified for a single-spectrum sensor can also be applied for a multispectral sensor. As such, the thesis outlines a development of a prototype of a real-time UXO detection system composed of a commercial-off-the-shelf (COTS) multi-spectral sensor and a small COTS unmanned aerial system.Military Expert 5, Singapore ArmyApproved for public release. Distribution is unlimited