Dynamics of target detection using drone based hyperspectral imagery

Abstract

Imagery from advances in imaging and acquisition platforms, such as Unmanned Aerial Vehicles (UAV) having finer spatial resolution helped solve problems in identifying plant species, agricultural monitoring, rock characterization, albeit it poses plausible challenges arising from complex object-sensor dynamics. Reflectance spectra from close-range imaging are found to be significantly affected by the shape of the object, illumination angle, and light source position. Assessment of close-range hyperspectral imaging for target detection is useful across different application scenario for drone-based hyperspectral imaging. Over an experimental study site in Bangalore, drone-based hyperspectral imagery was acquired with the goal of detecting and identifying various artificial targets placed in a complex imaging geometry with different target-backgrounds. The acquisition platform used a compact hyperspectral imaging sensor mounted on a drone at a flying altitude of 95 m. Different metallic sheet targets, painted with green color, and natural metallic color are used. Green metallic target is placed and imaged in various target-background environments whereas the other metallic target is placed on open soil background. Target detection algorithms: spectral angle mapper (SAM), matched filter (MF), adaptive cosine estimator (ACE), constrained energy minimization (CEM) are evaluated using the statistical performance indicators using ROC curves. For a partially visible target, with a probability of detection at 99%, the probability of false alarm for ACE, CEM, MF, and SAM is found to be 63%, 96%, 29%, and 18% respectively. For nearly camouflaged target PFA is found to be 75, 94, 59 and 79% for the ACE, CEM, MF and SAM detector with a probability of detection at 99%. Results warrant further refinements in existing hyperspectral target detection algorithms for close-range hyperspectral imaging.