Within the hyperspectral community, change detection is a continued area of interest. Change detection refers to the problem of identifying interesting changes that occur to a spatial area over which imagery has been collected on multiple occasions. Interesting changes in imagery typically correspond to changes in material reflectance associated with pixels in the scene, due e.g. to objects entering or leaving the scene. Complicating the problem of change detection is the presence of shadow, illumination, and atmospheric differences, as well as misregistration and parallax error, which often produce the appearance of a material change. Current change detection methods all suffer from similar weaknesses related to these nuisance changes.
To address these change detection difficulties, this dissertation proposes a model-based approach to change detection in hyperspectral imagery. Using a physical model describing the sensor-reaching radiance, the change detection problem is formulated as a statistical hypothesis test in this work. This model-based change detection method incorporates model terms to account for both direct and diffuse shadow fractions to help mitigate false alarms associated with shadow differences between scenes. The resulting generalized likelihood ratio test provides an indicator of change at each pixel. This change detection approach is the first to utilize a physical model, which allows for the use of additional information beyond the statistics of the data for improved detection performance over baseline methods.
Using knowledge of hyperspectral data collection and calibration procedures, the data model and alternating optimization technique are extended for application to uncalibrated and relatively calibrated data. The model-based change detection method is applied to synthetically generated data, tower data, and airborne data with differing levels of calibration applied. The model-based approach demonstrates improvement in detection performance over baseline change detection algorithms for the data sets considered. Additionally, the model-based method provides estimates of model parameters, which can potentially be used for further data exploitation purposes.
Hyperspectral noise estimation is a sub-problem associated with the model-based change detection method. Hyperspectral data collected using charge-coupled devices or other photon detectors have sensor noise that is directly dependent on the amplitude of the signal collected. Additionally, the statistics of the noise can vary both spatially and spectrally as a result of camera characteristics and the calibration process applied to the data. Supervised and unsupervised noise estimation techniques are presented for estimating the noise statistics using calibration data or directly from the imagery if calibration data is unavailable.