Measurement of Photosynthetic Pigment Content using Convolutional Neural Network

Abstract

Estimation of photosynthetic pigment levels from leaves can be done using conventional methods using laboratory equipment such as spectrophotometers and using digital image processing from leaf images with a computational model. In digital image processing methods, various models are used, such as neural network, CNN, and linear regression. Measurement of photosynthetic pigment levels using image processing methods uses color value data from image data as input to the model used. In this study, we will analyze the effect of various types of color space and inpaint preprocessing settings on the accuracy of the CNN model in measuring leaf photosynthetic pigment levels. The color space types being tested are 4 single color spaces RGB, HSV, LAB, and YCbCr, as well as 6 color combination spaces RGB+HSV, RGB+LAB, RGB+YCbCr, HSV+LAB, HSV+YCbCr, and LAB+YCbCr. The choice of the type of color space takes into account the phenomenon of color constancy and the characteristics of the color space on the lighting elements. In addition, image data is divided into two types, namely through inpaint preprocessing and not, so that in total there are 20 types of input data. After the CNN model training process with various types of color spaces and different preprocessing settings as input data, observations were made on the accuracy values, namely the training MAE and the validation MAE for each model. From 20 types of input data, 3 types of input data are obtained which are recommended as input data that provide the best model accuracy value based on MAE validation with values ​​of 0.08761, 0.09252, and 0.09288. The three recommended input data from the sequence of accuracy values ​​are RGB+LAB without inpaint, RGB with inpaint, and LAB+YCbCr without inpaint.