Enhancing plant health classification via diffusion model-based data augmentation

Vertical green wall systems-integrated assemblies of plant modules-have witnessed rapid market proliferation due to their broad spectrum of environmental benefits, including aesthetic enhancement, regulation of ambient temperature and humidity, and the mitigation of airborne pollutants. Given that these systems rely on modular plant units, precise forecasting and timely replacement of deteriorating specimens become imperative, rendering plant health monitoring a pivotal component of their maintenance. In response, a substantial body of research has advanced the application of deep-learning image classification techniques to evaluate plant health conditions. However, practical data collection is impeded by the relative scarcity of instances representing the ’Wilted’ condition compared to the ’Normal’ state, with the acquisition of ’Slightly Wilted’ data-crucial for proactive maintenance-being even more challenging. The intrinsic variability within the ’Slightly Wilted’ category further exacerbates the difficulties of accurate annotation and classification. To surmount these obstacles, this study introduces an innovative data augmentation framework that synthesizes ’Slightly Wilted’ examples from existing ’Normal’ and ’Wilted’ state data via diffusion model techniques. Specifically, the proposed methodology interpolates between the two states using diffusion models and assigns soft, probabilistic labels to the generated images based on the interpolation ratios, thereby enhancing the precision of the classification model training. Experimental evaluations reveal that this augmentation strategy substantially improves predictive performance, while also enabling the model to discern not only categorical plant health conditions but also to quantify the degree of health decline-thereby offering a more nuanced and actionable approach to managing plant vitality in vertical green wall systems.