Physics-Aware Shadow Compensation for Hyperspectral Imagery Based on Spectral Unmixing and Data Fusion

Kurzfassung

Imaging spectrometers, also known as hyperspectral sensors, acquire reflectance spectra from targets on ground in up to hundreds of narrow spectral bands over a wide range of the electromagnetic spectrum. Having an increased spectral resolution with respect to other kinds of optical imagery, such as panchromatic and multispectral, hyperspectral data can discriminate materials more accurately and thus allow new applications in remote sensing. Nevertheless, the Earth’s surface topography causes the occlusion of incoming illumination for ground targets, leading to shadow effects in the acquired images. Shadows can considerably decrease the performance of image analysis algorithms, and have thus drawn growing attention in the literature. Although shadow issues have been discussed for some kinds of optical sensors, few works have addressed these effects and their unique challenges in hyperspectral imagery, demanding novel shadow-aware methods. This dissertation proposes three robust algorithms, built on one another, for tackling shadow effects in hyperspectral images. The dissertation first presents a shadow detection and removal framework based on physical assumptions and spectral unmixing. The main idea is to model the shadow formation using a few physically interpretable shadow-related parameters, and apply them in order to detect and remove shadows. Specifically, a novel physics- and shadow-aware spectral mixing model is proposed, which considers how material spectra and shadows contribute to the individual pixel spectrum measured by the sensor. The described spectral mixing model can tackle only simple scenarios because it assumes simplified optical interactions and illumination conditions on ground. Thus, the following work improves the model to handle more complicated and generalized cases. The improved model regards the entire radiative propagation process, from illumination sources to the backscattered signals recorded by the sensor, using a discrete-time stochastic process and physical assumptions. As the model’s complexity increases, resolving unknown parameters via spectral unmixing becomes an ill-posed problem. Hence, a novel spectral unmixing approach for a robust estimation based on the Alternating Direction Method of Multipliers (ADMM) and data fusion is proposed. The ADMM decomposes a complex optimization problem into subproblems, each of which is easier to solve. Digital Surface Models (DSMs) are also employed in this approach, as they are insensitive to shadow effects. In addition, spatial relationships between neighboring pixels are considered. The proposed methods have been extensively evaluated using several simulated and real datasets in small and large regions. Results demonstrate that the proposed research works are effective and superior to state-of-the-art methods, both qualitatively and quantitatively.