Near-Infrared Spectroscopy for Moisture Content Measurement of Newly Fallen Deciduous Leaves: A Comparison of Partial Least Squares Regression and Random Forest with Wavelength Importance

Near-infrared spectroscopy (NIRS) was applied to measure the moisture content (MC) of newly fallen deciduous leaves from four broadleaf tree species, namely Zelkova serrata (Keyaki), Quercus serrata (Konara), Quercus acutissima (Kunugi), and Fagus crenata (Buna), collected in 2016 and 2017. Leaves were sampled immediately after abscission, an ecologically important condition that has received direct measurement, providing initial MC data for litter decomposition research. Two multivariate calibration approaches were compared, partial least squares regression (PLSR) and random forest regression (RFR), each applied with and without spectral preprocessing (moving average, multiplicative scatter correction, second-order Norris Gap derivative, and mean centring), and both evaluated by leave-one-out cross-validation (LOOCV) for direct comparability. PLSR with preprocessing achieved R2cv = 0.75–0.86 and an RPD (the ratio of performance to deviation) = 2.01–2.70 across species and years; RFR without preprocessing yielded R2cv = 0.47–0.75 (RPD = 1.39–2.03), while RFR with preprocessing improved markedly to R2cv = 0.78–0.88 (RPD = 2.13–2.94). Wavelength importance, assessed by variable importance in projection (VIP) for PLSR under both preprocessing conditions, showed that preprocessed models concentrated importance sharply in the water combination band near 1900 nm, whereas non-preprocessed models distributed importance more broadly but still situated their single strongest wavelength within this same band in six of seven datasets. RFR feature importance shifted from the water absorption region (without preprocessing) toward the cellulose/lignin band near 2100–2300 nm (with preprocessing); a model-independent correlation analysis indicated that this shift corresponds, in most datasets, to a genuine change in the underlying MC–spectrum relationship rather than solely reflecting how RFR handles collinear wavelengths. Independent, bidirectional year-to-year prediction (2016-calibrated models tested on 2017 data, and vice versa) showed that the high within-year accuracy summarised above did not reliably transfer across years—most severely for Z. serrata, for which cross-year prediction failed in both directions—indicating that annual recalibration is advisable for operational use of this method.