Evaluating the Applicability of Convolutional Neural Networks for Tree Classification in Donggwoldo

Background and objective: This study establishes a foundational modeling framework for quantitative research by developing and evaluating a baseline convolutional neural network (CNN) to classify tree representations in Donggwoldo, a 19th-century Korean court painting. Rather than focusing solely on performance optimization, the objective is to construct an initial methodology for identifying tree types based on their pictorial characteristics and converting traditional visual records into structured digital data. This research illustrates the potential of deep learning as a methodological bridge between artistic depictions and quantitative analysis within cultural heritage studies.Methods: A dataset of 580 high-resolution tree images was extracted from the Dong-A University and Korea University versions of Donggwoldo. These were manually categorized into six types based on art-historical classifications. To address data limitations and imbalance, augmentation techniques—including grayscale conversion, zooming, horizontal flips, and random color adjustments—were applied to enhance diversity while preserving stylistic integrity. A transfer learning model based on ResNet50V2 was implemented, utilizing pretrained layers as feature extractors. Two fully connected layers with ReLU activation and dropout regularization were added to prevent overfitting, with early stopping employed to ensure stable convergence.Results: The model achieved an overall classification accuracy of approximately 98% on a 150-image test set. Confusion matrix analysis indicated that rare misclassifications were primarily due to low resolution, background interference, and incomplete segmentation. Despite these challenges, the CNN effectively distinguished between diverse depiction styles, confirming the feasibility of applying deep learning to traditional paintings.Conclusion: This study demonstrates that CNNs can effectively transform traditional pictorial information into structured digital data. By establishing a baseline model, the research provides a methodological foundation for the quantitative analysis of historical landscape imagery. These results highlight deep learning’s potential as a complementary tool for cultural heritage studies, with future research directions including automated segmentation and interdisciplinary integration with landscape architecture.