BBG52: A New Dataset for Plant Species Recognition in the Balikpapan Botanical Gardens, Borneo Island

Abstract

The Balikpapan Botanical Garden serves as conservation area in Indonesia for preserving biodiversity, particularly the endemic species of Kalimantan. Accurate and efficient identification and classification of plant species are crucial for conservation efforts. However, traditional methods are often time-consuming and require an expert necessitating the need for an automated approach. In this study, we manually collected a dataset of natural images in the Balikpapan Botanical Garden that simulate real-world conditions which contains 5200 image samples of 52 different plant species named BBG52. We compared the manual train-test data splitting by considering intra-class variants againts random splitting to evaluate the performance differences. To construct classification model, we employed ResNet variants as pre-trained models—ResNet-34, ResNet-50, and ResNet-10—and examined the effect of the hidden layer in the classification part of the model. Our empirical results demonstrate that manual data splitting yields better performance than random splitting. Furthermore, the ResNet-50 model without additional hidden layers achieved the best performance with an accuracy of 96.88% and F1-score of 0.9689. The computational analysis provided empirical evidence that the model requires 0.1379 seconds on a CPU and 0.0861 seconds on a GPU demonstrating the model’s efficiency in the constrained device. The BBG52 dataset is openly accessible at https://github.com/inidhanii/BBG52