| Abstract [eng] |
Hyperspectral imaging is a remote sensing technique that collects data across the entire electromagnetic spectrum to obtain spectral data for each pixel in a scene. This enables a non-destructive method for gathering data for material identification from spectral signatures, object detection, and other tasks related to material analysis. With the growing popularity of hyperspectral imaging being used on UAVs, the need for improved hyperspectral analysis algorithms increases. One of the research areas in the hyperspectral imaging field is hyperspectral unmixing, which combines one or multiple algorithms that extract material information from individual pixels in the hyperspectral image. This study focuses on solving the three main problems of hyperspectral unmixing: lack of available open UAV-gathered hyperspectral datasets; absence of standardised testing for hyperspectral unmixing algorithms; paucity of research on hyperspectral unmixing methods for hyperspectral data gathered by UAVs. An open UAV-gathered hyperspectral unmixing dataset was created and published for use in further study. A benchmarking methodology was proposed for use in the evaluation of new and existing hyperspectral unmixing algorithms. A new U-Net-based deep neural network model was created that achieved, on average, a 12% better hyperspectral unmixing performance compared to the state-of-the-art transformer-based neural network model. |
