Hello, is it me you're looking for? Searching for material targets in hyperspectral remote sensing imagery
Remote sensing refers to the collection of information about an object or surface by observing it at some distance; while the sensors may be used in close proximity, they are frequently deployed on airborne or satellite platforms. Hyperspectral remote sensing, in particular, collects very high-dimensional images at many narrow, contiguous wavelengths, or “colors.” While the human visual system views the world in three colors - red, green, and blue - hyperspectral imaging sensors view the world in hundreds of colors, often imaging in both the visible and non-visible (e.g., ultraviolet, infrared) portions of the electromagnetic spectrum. The advantage to imaging a scene at many spectral wavelengths is that materials that may appear visually similar, such as a green car in a grass field, will have very different spectral signatures in the hyperspectral domain. The ability to automatically discriminate between materials in a remotely-imaged scene has innumerable practical applications, including tracking traffic patterns, mapping natural disaster damage, and monitoring urban development; the more prominent categories of analysis are generally referred to as image classification, anomaly detection, change detection, and target detection. This presentation focuses specifically on target detection, and in particular on how spectral variability makes the problem significantly more challenging. If a single target material corresponds to a variety of spectral signatures, then traditional detection algorithms will generate a high number of false detections, i.e., pixels in the image that are improperly labeled as target-like. In this talk, an overview of both hyperspectral imaging and target detection will be provided, along with a presentation of cutting-edge detection algorithms that are specifically designed to handle this variability.