Performance comparison of the convolutional neural network optimizer for photosynthetic pigments prediction on plant digital image

Determination of photosynthetic pigments in intact leaves is an essential step in the plant analysis. Along with the rapid development of digital imaging technology and artificial intelligence, now determination of plant pigments can be done in a non- destructive and real-time manner. In previous research, a prototype of the non-destructive and real-time system has been developed by utilizing the Convolutional Neural Network (CNN) model to produce predictions of three main photosynthetic pigments, i.e., chlorophyll, carotenoid, and anthocyanin. The CNN model was chosen due to its ability to handle raw digital image data without prior feature extraction. In the near future, this ability will be useful for developing analytical portable devices. Input of the system is multispectral plant digital image, and the output are predicted pigment concentration. Convolutional Neural Network performance depends on several factors, among them are data quality, algorithm tuning (weight initialization, learning rate, activation function, network topology, batches and epochs, optimization and loss) and models combination. The focus of this research is to improve the accuracy of CNN model by optimizing the selection for updating CNN architecture parameters which are optimization method and the loss function. As it is already known, there is no single optimizer can outperform for all cases. The selection for the optimizer should be made by considering the variability of the data and the nonlinearity level of the relationship patterns that exist in the data. Because the theoretical calculation is not enough to determine the best optimizer, an experiment is needed to see at firsthand the performance of optimizers that allegedly matches the characteristics of the data being analyzed. Gradient descent optimization method is well known for its ease of computing and speed of convergence on large datasets. Here, 7 gradient descent-based optimizers were compared, i.e., Stochastic Gradient Descent (SGD), Adaptive Gradient (Adagrad), Adaptive Delta (Adadelta), Root Mean Square Propagation (RMSProp), Adaptive Momentum (Adam), Adaptive Max Pooling (Adamax), and Nesterov Adaptive Momentum (Nadam). From the eksperiment result it is known that Adam is the best optimizer to improve LeNet ability in handling a digital image-pigments content relationship. However, when the resources for experimentation is limited, using Adadelta and Adamax is a wise choice to minimize risk.