Hyperspectral estimation of rice pigment content based on band depth analysis and BP neural network

The estimation accuracy of plant pigment content is low under higher pigment content since conventional vegetation indices tend to be less sensitive to the variance of pigment content. In order to improve estimation accuracy of rice carotenoid and chlorophyll contents with canopy reflectance during all growth stage, we explore the feasibility and effectiveness of combining the band depth analysis (BDA) and back propagation (BP) neural network to solve the problem of vegetation index saturation. With canopy hyperspectral data (400-750 nm), four band indices — band depth (BD), band depth ratio (BDR), normalized band depth index (NBDI) and band depth normalized to band area (BNA) — were calculated via continuum removal processing. Principal component analysis (PCA) was used to reduce the dimensions of hyperspectral data, and determined 10 principle components, which were introduced into BP neutral network as input variables. In the study, canopy hyperspectral reflectance and pigment content measurements were conducted in Meichuan Town of Hubei Province, China. Eight treatments of nitrogen fertilization (0, 45, 82.5, 127.5, 165, 210, 247.5 and 292.5 kg·hm^-2) were applied to generate various indices of vegetation and pigment content. Linear and nonlinear regression models were used to quantitatively analyze the vegetation indices and measured pigment content. In addition, coefficient of determination (R²) and root mean square error (RMSE) were used to evaluate the models. All the hyperspectral indices were comparatively analyzed. As a result, BDA showed the differences in spectral absorption characteristics and revealed more potential information to enhance spectral difference. The estimation model combined band index BD and BP had the highest estimation accuracy for carotenoid content in rice leaves, with R² = 0.61 and RMSE = 0.128 mg·g^-1; while the estimation model combined band index BNA and BP had the highest estimation accuracy for chlorophyll content in rice leaves, with R² = 0.73 and RMSE = 0.343 mg·g^-1. Further comparison between BDA & BP models with the best regression model for vegetation index indicated that BP neutral network model based on BDA provided a better solution to saturation problem and a higher estimation precision of rice leaf pigment content.