Noise-Resistant Spectral Features for Retrieving Foliar Chemical Parameters

Foliar chemical constituents are important indicators for understanding vegetation growing status and ecosystem functionality. Provided the noncontact and nondestructive traits, the hyperspectral analysis is a superior and efficient method for deriving these parameters. In practice, the spectral noise issue significantly impacts the performance of the hyperspectral retrieving system. To systematically investigate this issue, by introducing varying levels of noise to spectral signals, an assessment on noise-resistant capability of spectral features and models for retrieving concentrations of chlorophyll, carotenoids, and leaf water content was conducted. Given the continuous wavelet analysis (CWA) showed superior performance in extracting critical information associating plants biophysical and biochemical status in recent years, both wavelet features (WFs) and some conventional features (CFs) were chosen for the test. Two datasets including a leaf optical properties experiment dataset (n = 330), and a corn leaf spectral experiment dataset (n = 213) were used for analysis and modeling. The results suggested that the WFs had stronger correlations with all leaf chemical parameters than the CFs. According to an evaluation by decay rate of retrieving error that indicates noise-resistant capability, both WFs and CFs exhibited strong resistance to spectral noise. Particularly for WFs, the noise-resistant capability is relevant to the scale of the features. Based on the identified spectral features, both univariate and multivariate retrieving models were established and achieved satisfactory accuracies. Synthesizing the retrieving accuracy, noise resistivity, and model's complexity, the optimal univariate WF-models were recommended in practice for retrieving leaf chemical parameters.