Understanding the role of non-structural carbohydrates (NSC) in tree-level carbon cycling crucially depends on the availability of NSC data in a sufficient temporal resolution covering extreme conditions and seasonal peaks or declines. Chemical analytical methods should therefore get complemented by less extensive retrieval methods. To this end, we explored the potential of diffuse reflectance spectroscopy for estimating NSC contents at a set of 180 samples taken from leaves, roots, stems and branches of different tree species in different biogeographic regions. Multiple randomized partitioning in calibration and validation data were performed with near-infrared (NIR) and mid-infrared (MIR) as well as combined data. With derivative spectra, NIR markedly outperformed MIR data for NSC estimation; mean RMSE for outer validation samples equalled 2.58 (in % of dry matter) compared to 2.90, r2 was 0.64 compared to 0.52. We found complementary information related to NSC in both spectral domains, so that a combination with high-level data fusion (model averaging) increased accuracy (RMSE decreased to 2.19, r2 equalled 0.72). Spectral variable selection with the CARS algorithm further improved results slightly (RMSE = 1.97, r2 = 0.78). On the level of tissue types, we found a marked differentiation concerning the appropriateness of datasets and approaches. High-level data fusion was successful for leaves, NIR data (together with CARS) provided the best results for wooden tissues. This suggests further studies with a greater number of samples per tissue type but only for selected (main) tree species to finally judge the sensitivities of diffuse reflectance spectroscopy (NIR, MIR) for NSC retrieval.
Keywords: Data fusion; Diffuse reflectance spectroscopy; Mid-infrared; Near-infrared; Non-structural carbohydrates; Spectral variable selection.
Copyright © 2023 Elsevier B.V. All rights reserved.