Integration of leaf spectral reflectance variability facilitates identification of plant leaves at different taxonomic levels.

Plant identification is crucial for the conservation and management of natural areas. The shortwave spectral reflectance of leaves is a promising tool for rapidly identifying plants at different taxonomic levels. However, leaf spectral reflectance changes in response to biotic and abiotic conditions. Here we assess whether this variability in spectral reflectance affects the accuracy of classification methods currently used to predict plant taxonomy and identify factors that most influence leaf spectral signatures, as proxies for predicted biochemical and structural traits. We used leaf reflectance from 42 woody species from the living collection at the New York Botanical Garden across two sets of pairwise samplings (spring 2019/summer 2020 and spring 2019/winter 2021). We found that classification accuracy was poor when only spring samples were used to train models but improved when natural variation across all seasons was incorporated into classification models. To evaluate the influence of species relatedness or growth conditions (temperature, relative humidity, and daylength) on spectrally predicted traits, we applied Partial Least Squares Regression (PLSR) coefficients derived from NEON data to predict foliar traits including photosynthetic pigments, water content, leaf dry mass per area, and carbon and nitrogen content. Results showed that trait variation was not influenced by phylogeny but was significantly influenced by environment, except for water-related spectral bands for species remeasured in summer 2020. These results demonstrate that classification methods developed to handle datasets with large collinearities, such as leaf spectral reflectance, underperform when individual spectral variability is caused by environmental factors. This issue must be addressed for future development of remotely sensed taxonomy applications for evergreen broadleaf vegetation.