Novel augmentation techniques using diffusion models for green wall plant health classification

Green walls, vertical plant-based structures, are increasingly popular due to their diverse environmental benefits, including aesthetic enhancement, temperature and humidity regulation, and air pollutant removal. These systems, typically consisting of modular plant units, require accurate condition prediction and timely replacement of withering plants, making plant health monitoring essential for effective maintenance. Despite advancements in deep-learning-based plant classification models, real-world data collection challenges persist, particularly for the "Wilted" state, which is significantly harder to acquire than the "Normal" state. Moreover, data scarcity for the "Slightly Wilted" state critical for proactive maintenance further exacerbates the challenge. The continuous nature of plant deterioration within the "Slightly Wilted" category introduces labeling ambiguities, making accurate classification more difficult. To address these challenges, this study proposes an innovative augmentation approach that synthetically generates "Slightly Wilted" data using Diffusion Models. Specifically, the method interpolates between "Normal" and "Wilted" states through Diffusion Models, assigning soft labels based on the synthesis ratio, thereby enhancing classification model performance. Experimental results demonstrate that the proposed augmentation methodology improves classification accuracy and F1 score by up to 4% compared to models initialized with ImageNet weights, highlighting its effectiveness. Additionally, the proposed method not only classifies plant health conditions but also provides a more granular assessment of health severity, offering enhanced precision and actionable insights for green wall maintenance.