Quantifying chlorophyll-aandbcontent in tea leaves using hyperspectral reflectance and deep learning

To improve the quality of green tea, low light stress has been used to increase the chlorophyll-a(chl-a) content of tea leaves, although shading treatments sometimes lead to early mortality of tea trees. Therefore, in situ measurement of chl-aand chlorophyll-b(chl-b), which are markers for evaluating light stress and response to changing environmental conditions, can be used to improve tea tree management. Chlorophyll content estimation is one of the most common applications of hyperspectral remote sensing, but most prior studies have evaluated samples grown under relatively low stress. Therefore, the results of prior studies are not applicable for estimating chl-aand chl-bcontents of shade-grown tea. Machine learning algorithms have recently attracted attention as an approach for evaluating biochemical properties. In the present study, three different common machine learning algorithms were compared, including random forests, support vector machines and deep belief nets. The ratios of performance to deviation (RPDs) of deep belief nets (DBN) were always larger than 1.4 (the ranges of RPD were 1.49-4.92 and 1.48-5.10 for chl-aand chl-b, respectively), suggesting that DBN is a unique algorithm that can reliably be used for estimation of chl-aand chl-bcontents.