Apple, as a globally vital economic crop, often suffers severe yield reductions due to leaf diseases such as rust, powdery mildew, and brown spot. Traditional disease identification methods heavily rely on manual expertise, requiring professional technicians to visually inspect leaf morphology, color, and other characteristics for diagnosis. However, this approach is time-consuming, labor-intensive, and prone to errors, as visual differences in early disease stages are often subtle and easily confused. In recent years, machine learning algorithm-based disease recognition technologies have enabled automated analysis of leaf images, rapidly locating lesions and classifying disease types with high accuracy in laboratory settings. However, in complex field environments, factors like varying lighting conditions and background interference often degrade accuracy. The critical question arises: How can algorithms achieve both “clear vision” (high precision) and “fast computation” (efficiency) in such scenarios?
Professor Hui Liu’s team from school of traffic and transportation engineering, central south university, addressed this challenge by developing the Incept_EMA_DenseNet model. By integrating multi-scale feature analysis and an attention mechanism, this study elevated the accuracy of apple leaf disease identification to 96.76%, significantly outperforming mainstream models. This innovation effectively balances algorithmic precision and practicality. The study has been published in Frontiers of Agricultural Science and Engineering (DOI: 10.15302/J-FASE-2024583).
Starting with the limitations of traditional models, this study found that single-scale feature extraction struggles to capture both global distribution and local details of diseases. For instance, while rust’s yellow spots and gray spot’s brown patches may share similar local textures, their overall patterns differ. To address this, they embedded a multi-scale fusion module into the model’s shallow layers, enabling simultaneous analysis of fine textures and overall leaf morphology.
To further prioritize critical information, they introduced an Efficient Multi-scale Attention (EMA) mechanism, which automatically identifies disease regions and assigns higher weights. For example, when analyzing powdery mildew, the algorithm focuses on the density of white powdery substances rather than healthy green areas. Compared to conventional attention methods, EMA simplifies computations, reduces parameters by 50%, and improves classification accuracy by an additional 1.38%, achieving true “intelligent focusing”.
To ensure field applicability, this study optimized the classic DenseNet_121 network through lightweight modifications. The refined model can run smoothly on standard smartphones, allowing farmers to diagnose diseases in real time by simply photographing leaves, eliminating reliance on expensive equipment.
The technology’s efficacy was validated on a dataset of 15,000 images. In mixed tests involving eight common diseases and healthy leaves, the model achieved over 94% accuracy in distinguishing easily confused diseases (e.g., brown spot vs. gray spot) and adapted to varying lighting and camera angles. This breakthrough enables farmers to swiftly implement targeted treatments via mobile apps, reducing pesticide misuse and economic losses—all without specialized expertise.
DOI: 10.15302/J-FASE-2024583