AI-Powered Detection of Pumpkin Leaf Diseases Using DualFusion-CBAM-Stochastic for Yield Protection and Precision Agriculture

Ruchika Bhuria¹Rahul Singh¹Mudassir Khan^2*Mohamed Abbas²Jaibir Singh³Amel Ksibi⁴Nitin Kumar⁵Nitika Kapoor⁶Upinder Kaur^7*

• ¹Chitkara University, Rajpura, India
• ²King Khalid University, Abha, Saudi Arabia
• ³Galgotias University, Greater Noida, India
• ⁴Prince Sattam bin Abdulaziz University, Al Kharj, Saudi Arabia
• ⁵Graphic Era Deemed to be University, Dehradun, India
• ⁶Chandigarh University, Sahibzada Ajit Singh Nagar, India
• ⁷Lovely Professional University, Phagwara, India

Early and accurate detection of pumpkin leaf diseases is essential for precision agriculture, yet manual inspection remains slow, subjective, and difficult to scale in real field environments. To overcome these limitations, this study proposes DualFusion-CBAM-Stochastic, a hybrid deep-learning architecture designed to enhance discriminative feature learning and generalization for agricultural image classification. The framework integrates two complementary convolutional backbones: DenseNet121, which captures fine-grained texture variations through dense connectivity, and EfficientNetB3, which extracts multi-scale contextual features using compound scaling. Input images are preprocessed through resizing to 224×224 pixels, ImageNet-based normalization, and controlled augmentation (horizontal/vertical flips, rotation, and zoom) to improve feature diversity. The dual feature streams are refined using the Convolutional Block Attention Module (CBAM), which applies sequential channel and spatial attention to highlight disease-relevant regions before fusion. To further improve robustness, stochastic-depth regularization randomly drops deep layers during training, This is a provisional file, not the final typeset article mitigating overfitting while preserving essential semantic representations. The model was trained on a balanced dataset of 2,000 images comprising five disease categories and evaluated using ablation experiments and comparative analysis against state-of-the-art models. The proposed architecture achieved 96% accuracy, outperforming existing CNN-based methods. The experimental results demonstrate that the synergistic combination of dual-backbone fusion, attention-guided refinement, and stochastic-depth regularization substantially enhances classification performance, feature interpretability, and model stability under diverse visual conditions. These contributions collectively advance automated pumpkin leaf disease diagnosis and offer a rigorous methodological foundation for future research in agricultural image analysis.