LiDAR-Aided Total Variation Regularized Nonnegative Tensor Factorization for Hyperspectral Unmixing

Hyperspectral unmixing (HU) is an important research field in hyperspectral image processing. In recent years, Nonnegative Tensor Factorization (NTF)-based methods have gained great importance in remote sensing imagery, especially hyperspectral unmixing, regardless of any information loss. Nevertheless, NTF has some disadvantages, such as signal-to-noise ratio (SNR) and noncovexity conditions. Mentioned problem can be solved by introducing some spatial regularizations. On the other hand, LiDAR data provides Digital Surface Model (DSM) information gives accurate elevation information about the observed scene. Moreover, total variation (TV)-based regularization provides piecewise smoothness and it preserve edge structure information in the abundance maps. However, this property could be inappropriate for pixels located in edges. LiDAR-DSM alleviates this problem by contributing neighboring objects pixels differently. In this paper, we proposed a simple yet efficient HU framework that incorporates LiDAR data with TV regularized matrix–vector NTF method (LiMV-NTF-TV). Experimental studies are carried out on simulation data sets and demonstrate that the proposed framework can provide better abundance estimation maps.