SEQUENTIAL TENSOR DECOMPOSITION FOR GAS TRACKING IN LWIR HYPERSPECTRAL VIDEO SEQUENCES

With the development of hyperspectral imaging instruments, hyperspectral video sequences (HVS) can now be acquired with both high spectral and high temporal resolutions, allowing dynamic monitoring tasks such as gas tracking. However, the effective use of such large-scale sequential data also raises some challenges. Directly processing these data requires dramatic needs in terms of memory and computational loads. In this paper, we propose a novel method for gas tracking in HVS, based on decomposing sequential tensors into low-rank and error components, respectively. The gas target can be revealed from the error components corresponding to each frame. The global information contained in each frame and the correlation between adjacent frames are exploited by this tensor decomposition. Experiments are conducted on real HVS, assessing the good performances of the proposed method.