Nondestructive measurement of total nitrogen in lettuce by integrating spectroscopy and computer vision

Abstract

This study was conducted to develop and assess a method of estimating the nitrogen (N) content of lettuce (Lactuca sativa) canopy using a combination of spectroscopy and computer vision for nondestructive N detection. In the experiment, 90 lettuce samples with five N treatments were collected for data acquisition by two different techniques. On the spectroscopy side, canopy spectral reflectance was measured in the wavelength range of 350 to 2500nm at 1-nm increments. Four spectral intervals (376 variables) were selected by synergy interval partial least squares and were further reduced to 73 wavelength variables, chosen using a genetic algorithm applied to first-order derivatives of the canopy reflectance. On the computer vision side, 11 plant features were extracted from images, including top projected canopy area as a morphological feature; red, green, blue, hue, saturation, and intensity values as color features; and contrast, entropy, energy, and homogeneity as textural features. Next, principal component analysis was implemented on the spectral variables and on the image features, and extreme learning machine modeling was used to fuse the two kinds of data and construct a model. For the optimum model achieved, the root-mean-square error of prediction=0.3231% and the correlation coefficient of prediction=0.8864. This work demonstrates that integrating spectroscopy and computer vision with suitable efficient algorithms has high potential for use in the nondestructive measurement of N content in lettuce, considerably improving accuracy over that using a single sensor modality.