Active Layer Depth Characterization Using UAV-Based Hyperspectral Remote Sensing Université de Liège

Abstract

Permafrost degradation is one of the most significant consequences of climate warming in Arctic regions. Active Layer Depth (ALD), defined as the maximum seasonal thaw depth above permafrost, is widely used as an indicator of permafrost dynamics. How- ever, traditional monitoring methods mainly rely on field measurements that provide accurate but spatially limited observations. In this context, UAV-based hyperspec- tral remote sensing offers new opportunities for investigating environmental variables associated with ALD variability. This thesis evaluates the potential of hyperspectral imagery acquired by an unmanned aerial vehicle (UAV) for the characterization of ALD near Eight Mile Lake, Interior Alaska. A complete preprocessing workflow was implemented to transform raw hyper- spectral acquisitions into atmospherically corrected reflectance mosaics, which were combined with field measurements of ALD. Several complementary analyses were conducted, including spectral correlation anal- ysis, vegetation indices, Principal Component Analysis (PCA), and Random Forest machine learning models. Feature selection strategies, DEM integration, and model robustness were also investigated. The results demonstrate that hyperspectral imagery contains meaningful information related to ALD variability. The strongest spectral correlations reached approximately |r| = 0.60, while the optimized Random Forest model achieved a Pearson correlation coefficient of 0.801, a mean cross-validation R² of 0.584, and a global R² of 0.641. The inclusion of DEM information further improved predictive performance. In addi- tion, a recurrent importance of the 512 nm wavelength was identified across multiple independent analyses. Overall, the results highlight the potential of UAV-based hyperspectral remote sensing as a promising tool for local-scale permafrost monitoring and Active Layer Depth characterization.