Trait prediction in rare marsh plants: A PLSR modeling framework using dominant families and sample size thresholds

Plant functional traits serve as crucial indicators for assessing the status of ecosystems. However, accurately predicting leaf traits of rare species using spectral reflectance remains challenging because the limited sample sizes of rare species make it difficult to construct reliable spectral prediction models for them directly. The study hypothesized that the leaf traits of rare marsh species could be inferred from the modeling results of the dominant species, even with a limited sample size. To investigate this, we measured leaf reflectance spectra and eleven leaf functional traits across different species and established Partial Least Squares Regression (PLSR) models based on five combinations of plant families (All-families, Dominant-families, Non-Cyperaceae, Cyperaceae-Poaceae, and Cyperaceae) and six different sample sizes (40, 80, 120, 160, 200, and 240 samples). Our findings revealed that the PLSR models exhibited higher predictive performance for nitrogen (N, R Val 2 =0.87) and leaf water content (LWC, R Val 2 =0.85) compared to the other nine leaf traits. Notably, models constructed using the Dominant-families dataset (excluding rare species) exhibited predictive accuracy statistically comparable to the All-families dataset (which includes rare species) for all eleven traits. Additionally, we found that a minimum of 160 samples in the Dominant-families dataset was required to achieve reliable prediction for most leaf traits. This finding demonstrates that Dominant-families models can not only effectively substitute for All-families models but also provide a valid pathway to infer rare species' leaf traits. By addressing the data scarcity challenge (i.e., insufficient samples of rare species) through dominant species sampling, prioritizing the collection of Dominant-families datasets meeting the 160-sample threshold offers a practical framework for predicting rare marsh species' leaf traits. Overall, our findings provide a foundational framework for advancing spectral inversion techniques to more accurately predict leaf traits in marshes, particularly for rare species where direct sampling is constrained. • Developed a framework enables accurate prediction of leaf traits in rare species of marsh herbaceous plant with limited samples through hyperspectral data and PLSR model. • PLSR Model constructed using Dominant-families achieve comparable validation accuracy to All-families combinations, with at least 160 samples. • N and LWC can be more accurate predicted by leaf hyperspectral data.