Skip to Main Content
Hyperspectral change detection has been proved to be a promising technique for detecting indiscernible targets in different background. However, in the case of dense industrial and urban areas the complexity of the terrain and the multi-temporal images, which include positional deviation, radiant and atmospheric variation, shadows and spatial structure alteration, severely affects the automation of the change detection. This paper develops and enlarges four clustering based methods to detect man-made changes in VNIR and TIR hyperspectral scenes. The first applied method is Covariance-Equalisation (CE) multivariate statistical techniques, which detects differences between linear combinations of the spectral bands from the two acquisitions. The other three methods perform clustering of a reference image and then detect changes in a target image using a class-conditional distance detector: (a) class-conditional CE (QCE), (b) bi-temporal QCE and (c) Wavelength Dependent Segmentation (WDS). For the detection of small changes in industrial and urban areas, data from two flight campaigns were used: AHS-160 over the port of Antwerp and over the city of Kalmthout (Belgium). It was found that the use of a spatially adaptive detector greatly increases change-detection performance for both target detection and false alarm reduction. Moreover, WDS clustering based methods demonstrated a substantial improvement in change detection when applied on combined-wavelengths (as MWIR and LWIR or VNIR and TIR) hyperspectral data sets with respect to a single-wavelength data set.