Hierarchical Vision Learning for Plant Disease Recognition Under Variable Field Conditions

Abstract

Automated plant disease recognition is important for improving agricultural productivity and reducing crop losses. Despite advances in computer vision, disease classification under natural field conditions remains difficult because of variations in illumination, leaf morphology, background complexity, and disease severity. This paper proposes a hierarchical vision learning framework that combines multiscale feature extraction with attention-based region localization for plant disease identification. The proposed method is evaluated in an illustrative experimental setting involving 62,400 plant images covering 12 crop species and nine disease categories, collected across four agricultural regions. The framework achieves a macro-averaged F1-score of 0.908 and an overall classification accuracy of 93.6%, compared with 0.846 and 88.1%, respectively, for a conventional convolutional baseline. For early-stage disease symptoms, the model improves recall from 0.741 to 0.834. Additional cross-region evaluations indicate improved robustness to changing lighting and environmental conditions. These results demonstrate the potential of hierarchical visual representations for developing more reliable automated plant health monitoring systems in precision agriculture.