Classification of Trees in <Donggwoldo (東闕圖)> Using CNN Deep Learning - Focusing on Tree Representation Techniques -

Abstract. This study explores the potential of applying Convolutional Neural Network (CNN)-based deep learning techniques to the analysis of vegetation landscapes depicted in traditional pictorial records. The research focuses on <Donggwoldo (東闕圖)>, a 19th-century Korean court painting that visually documents the architectural and natural landscape of the Joseon royal palaces. Drawing upon prior studies and the Chinese painting manual 『Jieziyuan Huazhuan (芥子園畵傳)』, six distinct types of tree representation were defined to serve as classification categories for the analysis. To construct a training dataset, a total of 580 tree images were manually cropped from two versions of <Donggwoldo>, held respectively by the Korea University Museum and the other by the Dong-A University Museum. These images were then augmented using techniques such as rotation, flipping, and zooming to enhance dataset diversity and reduce overfitting. Seven pre-trained CNN models (including ResNet50V2, InceptionV3, and DenseNet121) were tested in a CPU-based Google Colab environment. Among them, MobileNetV2 was selected as the most suitable model based on its balance of high accuracy, low computational demand, and relatively fast training time. The selected model achieved 100% validation accuracy and a minimal validation loss of 0.3% by epoch 65. Final performance on the test set reached 95.5% accuracy, with precision, recall, and F1 score each measuring 86.6%. These findings demonstrate that deep learning can effectively classify tree types in historical paintings and offer valuable tools for digital heritage preservation, landscape interpretation, and cultural informatics. The approach also underscores the potential for machine learning to assist in the systematic study of historical visual culture.