Data-driven procedures for environmental portable and remote spectral sensing

Abstract

The significance of in situ environmental monitoring is increasing due to climate change. Recent advancements in portable spectrometers, the widespread availability of digital imagery, and the high-resolution capabilities of next-generation satellites present excellent opportunities for out-of-laboratory environmental monitoring. Although portable spectrometers and cameras can be easily transported to measurement sites, their use presents various challenges. Changing environmental and light conditions, environmental interferences, and the reduced size of the instruments often result in data of poorer quality compared to benchtop instrumentation. This thesis explores the preprocessing and modeling of recent in situ and remote sensing methods to develop reliable environmental monitoring techniques.

The research focuses on providing statistical methodologies for environmental monitoring using five types of instrumentation: portable digital and hyperspectral imagery, multi- and hyperspectral remote sensing data, portable Raman spectroscopy, and Fourier transform infra-red (FT-IR) spectroscopy. The dissertation includes various novel methodologies illustrated in five included research articles. (i) A novel method is proposed for enhancing the spatial resolution of hyperspectral images, called spectral profile partial least-squares (SP-PLS). SP-PLS is able to maintain good spectral fidelity after resolution enhancement and comes with spatial and spectral uncertainty maps. (ii) Another novel method, kernel partial least-squares optimized with kernel flows (KF-PLS) is proposed for hyperspectral retrieval models. This method is performant when compared to various linear and nonlinear methods, and with K-PLS optimized with benchmark optimisers. The discussion is extended to the optimisation of kernel principal component regression (KF-PCR). (iii) An innovative colorimetric procedure is proposed, based on color profiles, to evaluate color similarity, the appearance of new tones and color profile distribution changes. The method is extended to vegetal samples.

In addition to remote sensing data, experimental data were acquired in-house. The proposed procedures make out-of-laboratory data to be used for monitoring vegetation health, evaluating similarities, and identifying changes due to contaminants and stressors in this thesis.