Description
Due to the variability introduced by physical and chemical properties, acquisition conditions, and spectral mixing, interpreting soil reflectance spectra from hyperspectral imaging is challenging. This variability alters the position and shape of absorption bands. Studying the impact of water content, grain size, and acquisition geometry requires defining parameters that are physically based and interpretable. Tools like Tetracorder rely on fixed spectral libraries. In complex environments where variability distorts absorption shapes, this can be limiting.
Spectral deconvolution is a relevant framework for this purpose because it separates the continuum associated with physical effects, such as illumination, scattering and surface roughness, from absorptions linked to mineralogical signatures. The process involves four steps: continuum removal, absorption detection, parameter optimization and mineral identification. In this study, we present a new absorption detection method based on continuous wavelet transform (CWT).
In the solar domain (400–2500 nm), electronic transitions produce broad absorptions within the visible spectrum. In contrast, SWIR vibrational features are numerous and variable in shape. They also often overlap. This variability complicates automatic detection. While CWT-based approaches are effective for isolated absorption peaks, their performance is compromised when absorption bands overlap. These approaches also rely on manually tuned filtering thresholds that must be adjusted for each spectrum or sensor. Currently, there is no automated method capable of robustly detecting all mineral absorptions in a complex context.
The primary objective of my work is to propose an absorption band detection method adapted to the challenging conditions of hyperspectral imaging to facilitate mineral identification. The workflow automatically detects absorption features and identifies minerals in synthetic, laboratory, and airborne hyperspectral data. It accurately retrieves absorption features and reliably identifies minerals in pure samples and mixtures. Applying the method to the Cuprite image further demonstrates its robustness and relevance through qualitative comparison with a reference tool, such as Tetracorder.
These results suggest several directions for future research. One approach would be to analyze absorption parameters other than the position of the absorption band. This would allow us to evaluate mineral detection limits in different contexts and assess how factors such as water content, grain size, and acquisition geometry affect peak detectability.
| Speaker information | PhD 3rd year |
|---|