Evaluating Temporal Transferability of Imaging Spectroscopy-based Functional Trait Models in Alpine Grasslands

Imaging spectroscopy (IS) is widely used to estimate plant functional traits, yet the temporal transferability of spectral-trait models remains a critical issue. This is because models calibrated under specific environmental and phenological conditions may fail to generalize when applied beyond the conditions represented in their calibration data. This study evaluates the temporal transferability of IS-based plant trait models representing plant productivity, vegetation structure, and phenological development across four acquisition years (2010–2013) in alpine grasslands of the Swiss National Park (SNP). We evaluated how partial least squares regression models (PLSR) developed with Airborne Prism Experiment (APEX) IS data predict functional traits such as above-ground biomass (fresh and dry; AGB, g m -2 ), carbon (C%), nitrogen (N%), acid detergent lignin (ADL%), acid detergent fiber (ADF%), and neutral detergent fiber (NDF%). Plant trait models were developed for both single- and multi-year calibrations and evaluated using temporally held-out cross-year predictions. We found that within-year models’ validation performance ranged from R 2 ≈ 0.10 to 0.55. Despite this low-to-moderate within-year validation performance, temporal transferability was predominantly negative, with 76% of predictions yielding cross-year R 2 < 0. Multi-year models yielded positive transfers more frequently than single-year models (29.5% versus 18.6%), although this difference was not statistically significant, and positive transfers remained concentrated within specific calibration-prediction year combinations. Across the 148 transfer cases, the median Pearson correlation was 0.3, and the median regression slope was 0.2, with 86% of slopes below 0.5, indicating widespread compression of predicted trait variation. Cross-year R 2 was negatively associated with phenological mismatch between calibration and prediction datasets, quantified by absolute differences in accumulated growing degree days (|ΔAGDD|; Spearman ρ = −0.305, p = 0.0003), although this descriptive relationship largely distinguished temporal transfers involving 2010 and could not be separated from other year-specific differences. As the temporal transfer cases shared calibration models and included overlapping target datasets, case-level frequencies, confidence intervals, and correlations were interpreted descriptively. Traits such as fresh AGB and ADL exhibited the highest positive transfer frequencies, with 40% and 36% positive transfers, respectively, whereas N showed no positive transfers. The NIR and SWIR spectral regions contributed prominently to positive transfer models, although their relative contributions varied among traits and calibration designs. These findings demonstrate that temporal transferability cannot be assumed from within-year model performance and highlight the importance of temporally held-out evaluation. Robust trait modeling across years requires explicit consideration of phenological dynamics, trait-specific sensitivity, and environmental variability beyond the conditions represented in the calibration data.