Multispectral Imaging and Convolutional Neural Network for Photosynthetic Pigments Prediction

Abstract

The evaluation of photosynthetic pigments composition is an essential task in agricultural studies. This is due to the fact that pigments composition could well represent the plant characteristics such as age and varieties. It could also describe the plant conditions, for example, nutrient deficiency, senescence, and responses under stress. Pigment role as light absorber makes it visually colorful. This colorful appearance provides benefits to the researcher on conducting a nondestructive analysis through a plant color digital image. In this research, a multispectral digital image was used to analyze three main photosynthetic pigments, i.e., chlorophyll, carotenoid, and anthocyanin in a plant leaf. Moreover, Convolutional Neural Network (CNN) model was developed to deliver a real-time analysis system. Input of the system is a plant leaf multispectral digital image, and the output is a content prediction of the pigments. It is proven that the CNN model could well recognize the relationship pattern between leaf digital image and pigments content. The best CNN architecture was found on ShallowNet model using Adaptive Moment Estimation (Adam) optimizer, batch size 30 and trained with 15 epoch. It performs satisfying prediction with MSE 0.0037 for in sample and 0.0060 for out sample prediction (actual data range -0.1 up to 2.2).