Learning structural damage through hyperspectral images

Real-time ('snapshot') hyperspectral imaging is a recent advance in remote sensing technology that captures hundreds of spectral bands at a pixel within a scene of interest without scanning the scene linearly as it is in the traditional 'push-broom' hyperspectral imaging. This real-time mechanism enables the capability of collecting a large number of hyperspectral images in a short time; and also, it leads to the superb readiness when integrating with unmanned aerial vehicles (UAVs), especially the small multi-copter UAVs (colloquially drones). This capability provides the unique opportunity of performing 4-dimensional spatial-spectral reconstruction and hyperspectral-imaging based visual computing. More importantly, it provides the signal basis for the contextual information of any object in both the spectral dimensions and the local neighborhood. Hence, any pixel is equipped with more discriminative power in recognizing damage at materials level besides the spatial-variability based visual cues (e.g. edges and texture based on pixel intensities). In this paper, we will demonstrate this capability of hyperspectral imaging in identifying complex structural damage using traditional low-level feature extraction and classification methods (relative to deep learning methods). At the application level, different scenarios that challenge regular optical imaging will be attempted, which includes identifying structural concrete crack within complex scenes, identifying the degree of steel corrosion from visually similar corrosion artifacts, and pavement distress against pavement texture and contamination.