Determination of lettuce nitrogen content using spectroscopy with efficient wavelength selection and extreme learning machine

Measuring the nitrogen content of plants is useful for nitrogen fertilizer management. The aim of this work is to explore the use of spectroscopy for estimating lettuce leaf nitrogen content. Leaf reflectance spectra were measured using a spectroradiometer with a range of 350-2500 nm, and 160 fresh lettuce leaves given five different nitrogen treatments were used for spectra acquisition and total nitrogen determination. Interval partial least squares (iPLS), synergy interval partial least squares (siPLS), and siPLS combined with genetic algorithms (GA-siPLS) were used to select the optimal intervals or variables. Partial least squares (PLS) and extreme learning machine (ELM) methods were used to develop calibration models. One characteristic wavelength region (470-590 nm) using 120 variables was selected by iPLS; four characteristic wavelength regions (440-530, 530-620, 620-710 and 890-980 nm) using 362 variables were selected by siPLS, and 56 wavelength variables were selected by GA-siPLS. Six different regression models were established for nitrogen content by PLS and ELM based on optimal intervals or variables. The results imply that GA-siPLS is an efficient variable selection algorithm and ELM is a successful nonlinear regression tool. Furthermore, GA-siPLS combined with ELM is a feasible method for measuring nitrogen content in lettuce, as it performed better than other models. The optimal results achieved a root mean square error of prediction (RMSEP) = 0.2890% and correlation coefficient Rp = 0.9218.