HECA-MobileNet: an explainable lightweight attention-based deep learning framework for custard apple disease detection

The cultivation of custard apple is challenging as it is prone to several fungal and insect pest diseases which cause substantial loss to yield and quality. Early and accurate disease detection is essential to facilitate timely actions to reduce the level of damage. In this paper, we propose HECA-MobileNet, a lightweight hybrid deep architecture that combines MobileNetV3 and Convolutional Block Attention Module (CBAM) to automatically classify custard apple disease. The proposed framework utilizes depth wise separable convolutions, attention-guided feature refinement, and transfer learning with selective fine-tuning to improve the discriminative power while retaining the efficiency of computation. The model detects efficiently the following six prominent disease types: Anthracnose, Blank Canker, Diplodia Rot, Leaf Spot on Fruit, Leaf Spot on Leaf and Mealy Bug. Extensive evaluations on a challenging dataset of 8,226 images collected from field show the excellent performance in terms of 99.87% overall accuracy, 99.46% mean precision, 99.49% recall, and 99.42% F1-score. All disease classes achieved ROC–AUC in the range of 0.99–1.00, implying a good separability among classes. In addition, Grad-CAM++ visualizations illustrate that the suggested model, in general, focuses on biologically meaningful disease-specific areas, making it more interpretable and trustworthy. Due to its small size, low computational demands, and high accuracy, HECA-MobileNet model provides a reliable and scalable platform for real-time custard apple disease diagnosis, making it particularly well suited for mobile and edge-computing-based solutions for agricultural use.