Reconstructing fresh green leaf spectra in the SWIR-2 region (2001–2500 nm) collected in a humid environment by referring to publicly available green leaf spectral databases

Abstract

AbstractLeaf spectra (reflectance and transmittance) are key parameters for land surface physical and biogeochemical modeling and are commonly measured using a portable spectroradiometer and an integrating sphere or contact probe with an artificial light source. However, spectral data may be obscured mainly because of water vapor and low signal-to-noise ratios, especially in the shortwave infrared-2 region (SWIR-2, 2001–2500 nm). This erroneous pattern is particularly pronounced in humid conditions, such as in many tropical and subtropical regions, making data unusable in SWIR-2. In this study, we proposed a statistical/mathematical spectral reconstruction approach to retrieve noise-free SWIR-2 fresh green leaf spectra by referring to the available previously published quality-controlled fresh green leaf reflectance and transmittance reference databases. We processed 896 pairs of fresh tea (Camellia sinensisvar.sinensis) leaf reflectance and transmittance data from Alishan in central Taiwan. The spectral data were acquired by a field spectroradiometer with an integrating sphere. We selected a subset (500–1900 nm) of the spectra in the visible, near-infrared, and SWIR-1 regions (VNS-1) that was relatively insensitive to atmospheric conditions. Then, we applied a Gaussian fitting function to smooth the spectral profile. We matched those spectra with publicly available, quality-controlled, and Gaussian fitting function smoothed reference green leaf spectral databases obtained from Italy (LOPEX), Panama (SLZ), and Puerto Rico (G-LiHT) (1694 reflectance and 997 transmittance samples) and selected the one that was most similar (yielding the highest correlation coefficient) to each smoothed Alishan VNS-1 spectrum. We then used multivariable linear regression, linear parameter multiplication, and spectral reversion to reconstruct SWIR-2 spectra based on VNS-1 spectra. To assess the validity of the proposed SWIR-2 reconstruction method, we acquired an independent set of green leaf spectral databases from France (Angers) with SWIR-2 of 2001– 2450 nm. We found that the performance of the SWIR-2 reconstruction approach was satisfactory, with mean (± standard deviation) root-mean-square errors (RMSEs) of 0.0041 ± 0.0019 (reflectance, 3.0% of the mean SWIR-2 of the test data) and 0.0054 ± 0.0027 (transmittance, 2.5%) for each spectrum and RMSEs of 0.0058 ± 0.0027 (reflectance, 4.2%) and 0.0055 ± 0.0043 (transmittance, 2.5%) for each SWIR-2 band. The proposed approach successfully modeled SWIR-2 of the test spectra, which could be further improved with the availability of a more comprehensive set of green leaf reference spectral databases.