Accurately estimating and revealing the patterns of leaf C:N:P stoichiometry with remote sensing and neural network methods in a karst area

The carbon (C), nitrogen (N), and phosphorus (P) concentrations, along with the C:N:P stoichiometry in plants, play a vital role in regulating nutrient use efficiency in terrestrial ecosystems. However, fast and accurate estimation of the leaf C: N:P stoichiometry remains challenging, especially in complex karst areas. This study addresses this challenge by combining ground-based hyperspectral remote sensing technology with advanced neural network methods to estimate leaf C:N:P stoichiometry, based on 301 samples collected from nine typical karst areas in southern China. Our results showed that the PLSR (partial least squares regression) + BPNN (back propagation neural network), PLSR + GRNN (generalized regression neural network), and S-Transformer (simplified transformer) models achieve high performance in estimating leaf C:N:P stoichiometry in karst areas in southern China (R-2 was 0.71 0.88). Notably, we introduce fractional differentiation as a novel preprocessing step, which significantly improves model accuracy. Particularly, the predictive performances were effectively enhanced and stable when the fractional order was larger than 1.6. Furthermore, we identify distinct spectral sensitivity ranges for C and P are more sensitive to reflectance in the 400 800 nm wavelength ranges, while N exhibits sensitivity in both the 400 800 nm and 1500 2500 nm ranges after fractional differentiation. Our results also reveal that N plays an essential role in coordinating the C and P cycles in plants, and increasing leaf N concentrations could alleviate P stress on plant growth in karst areas. Compared to leaves C and N, P is more susceptible to environmental factors. Our results indicate a significant advancement in the application of groundbased hyperspectral remote sensing and neural network methods for ecological stoichiometry estimation in karst ecosystems. By successfully applying PLSR + BPNN, PLSR + GRNN, and S-Transformer models, we provide a robust framework for rapid and accurate hyperspectral inversion of leaf C:N:P stoichiometry. Our findings not only contribute to the understanding of nutrient dynamics in karst areas, but also provide practical tools for precision agriculture and ecosystem monitoring in similar environments.