Estimation of SPAD value in tomato leaves by multispectral images

Abstract

To improve the prediction accuracy of tomato chlorophyll content based on multispectral images, three preprocessing methods that can weaken illumination influence, i.e., self-adaptive gamma correction, multiscale Retinex and reflectance reconstruction, are compared and analyzed, and corresponding estimation models for the tomato Soil-Plant Analysis Development (SPAD) value are established. The correction coefficients are set according to the deviations between the grayscale values of the pixels in the highlighted and shaded areas and the average grayscale values in the area of uniform illumination to realize the self-adaptive gamma correction. The SPAD estimation model is constructed by the corrected image, with input parameters of nir, RVIb,g, RVIg,nir, RVIr,nir, NDVIr,g, and NDVIb,r, and the Rc2 and Rv2 of the model are 0.87 and 0.8, respectively. The original image is convolved with three different Gaussian functions to obtain a multiscale Retinex corrected image, and the SPAD estimation model is constructed. The input parameters are b, RVIr,g, RVIr,nir, RVIg,b, RVIg,nir, RVIb,nir, and NDVIr,g, and the Rc2 and Rv2 of the model are 0.91 and 0.84, respectively. The accuracy of the above two models depends on the selection of the corrected gamma value and the convolution kernel, and improper selections could seriously affect the accuracy of the model. For the SPAD value estimation model based on the reflectance reconstruction, its input parameters are nir, RVIg,nir, RVIr,g, NDVIg,b, NDVIr,nir, and NDVIb,r, and the Rc2 and Rv2 of the model are 0.90 and 0.88, respectively. The preprocessing procedure is simple, and the universality is high, making it suitable for the application of the digital field management of the crop.