Leaf Biochemistry Parameters Estimation of Vegetation Using the Appropriate Inversion Strategy

Abstract

Biochemistry parameters of vegetation are important indicators of the photosynthetic process and provide a substantial amount of data about the status of ecosystems. Estimation of these parameters are greatly affected by the correlations of spectral bands and the sensitivity of each biochemistry parameter to inversion models. Hence, reducing the spectral dimension and inefficient computation process using an appropriate inversion strategy is significant for biochemistry parameters’ estimation. In this work, we used band-selection-based artificial neural networks (ANNs) combined with feature weighting (FW) and principal component analysis (PCA) process to reduce the sensitive spectral correlations and to improve the inversion model predictability for four biochemistry parameters: chlorophyll a and b (Cab), carotenoid (Car), equivalent water thickness (EWT), and leaf mass per area (LMA). We analyzed the model performance by conducting different inversion strategies, including: (1) linking reflectance (R), transmittance (T), and R&T spectral properties in different numbers of band to four biochemistry parameters; (2) simultaneously and then separately inverting them using FW- and PCA-ANNs considering their sensitivity to the ANN model; and (3) choosing a spectral subset from R, T spectrum for EWT, and LMA inversion successively. The results show that: (i) the FW- and PCA-ANN models exhibit efficient improvements by selecting less spectral characteristics; (ii) concurrently inverting EWT and LMA can achieve a satisfactory R2, while it is inappropriate for Cab and Car whose optimal R2 are obtained by separately inverting all four biochemicals; (iii) the properties of R, T, and R&T spectra exhibit various performances on parameters inversion.