Estimating leaf chlorophyll contents by combining multiple spectral indices with an artificial neural network

Estimating leaf chlorophyll contents through leaf reflectance spectra is efficient and nondestructive, but the actual dataset always based on a single or a few kinds of specific species, has a limitation and instability for a common use. To address this problem, a combination of multiple spectral indices and a model simulated dataset are proposed in this paper. Six spectral indices are selected, including Blue Green Index (BGI), Photochemical Reflectance Index (PRI_5), Triangle Vegetation Index (TVI), Chlorophyll Absorption Ratio Index (CARI), Carotenoid Reflectance Index (CRI) and the green peak reflectance (R₅₂₅). Both stepwise linear regression (SLR) and back-propagation artificial neural network (ANN) are used to combine the six spectral indices for the estimation of chlorophyll content (Cab). In addition, to overcome the limitation of actual dataset, a “big data” is applied by a within-leaf radiation transfer model (PROSPECT) to generate a large number of simulated samples with varying biochemical and biophysical parameters. 30% of the simulated dataset (SIM30) and an experimental dataset are used for validation. Compared with linear regression method, NN yields better result with R² = 0.96 and RMSE = 5.80ug.cm^?2 for Cab if validated by SIM30, while R² = 0.95 and RMSE = 6.39ug.cm^?2 for SLR. NN also gives satisfactory result with R² = 0.80 and RMSE = 5.93ug.cm^?2 for Cab if validated by LOPEX dataset, however, the SLR only gets 0.72 of R² and 12.20ug.cm^?2 of RMSE. The results indicate that integrating multiple spectral indices can improve the Cab estimating accuracy with a better stability in different kind of species and the model simulated dataset can make up the shortfall of actual measured dataset.