Multi-Scale Vision Transformer Learning for Automated Plant Health Assessment in Urban Landscape Environments

Abstract

Accurate plant health assessment is essential for maintaining urban landscape quality, improving vegetation management, and reducing the operational costs of green infrastructure maintenance. However, traditional plant inspection methods rely heavily on manual observations and often struggle to identify early-stage physiological stress, disease symptoms, and environmental damage across heterogeneous landscape environments. This study proposes a multi-scale Vision Transformer framework for automated plant health assessment using RGB imagery and vegetation-related visual features. The proposed architecture integrates hierarchical image feature extraction, attention-based lesion localization, and adaptive feature fusion to identify disease symptoms and plant stress patterns across different spatial scales. To improve generalization under complex outdoor conditions, a domain adaptation module is introduced to reduce performance degradation caused by illumination changes, background interference, and seasonal variations. The framework is designed for evaluation using the PlantVillage dataset, containing approximately 54,000 labeled plant leaf images, supplemented by a proposed collection of 3,000 field images from urban landscape environments. Comparative experiments will include ResNet-50, EfficientNet, and standard Vision Transformer models. Evaluation metrics will include classification accuracy, macro-F1 score, precision, recall, and cross-domain generalization performance. The proposed framework aims to improve automated vegetation monitoring and support efficient, data-driven landscape maintenance practices.