The effect of parameter adjustment in sago palm classification-based convolutional neural network (CNN) model

In our study location, Merauke Regency, the easternmost city in Indonesia, the sago palm is associated with different types of ecosystems and other non-sago vegetation. During the harvesting season, the white flowers blossoming between the leaves on the tops of palm trees may be distinguished manually. Four classes were determined to address the visual inspections involving different parameters that were examined through the metric evaluation and then analysed statistically. The computed Kruskal-Wallis test found that the parameters vary in each network with a P-value of 0.00341, with at least one class being higher than the others, i.e., non-sago with a P-value of 0.044 with respect to precision, recall, and F1-score. Thus, the general linear model (GLM) was tested specifically in trained Network-15 and Network-17, which have similar parameters except for the batch size. It indicated the two networks' differences based on their prediction results, classes, and actual images. Accordingly, a combination of learning rate (Lr) and batch size improved the reliability of the training and classification task.