HASPNet: a hierarchically attentive signal-preserving network for papaya leaf disease classification with explainable deep learning.

The accuracy of papaya leaf disease classification is of highest priority in early-stage plant health surveillance and green farming. This paper presents HASPNet, a hierarchically attentive signal-preserving network specially designed for fine-grained papaya leaf disease classification from the newly proposed BDPapayaLeaf Dataset of 2,159 high-resolution images of five pathological classes. The network introduces a coordinated hierarchical attention framework; by integrating residual feature fusion with sequential SE and CBAM modules, HASPNet synchronizes multi-scale signal preservation with dual-stage recalibration, allowing the model to isolate subtle pathological signatures while maintaining global structural integrity. The architecture is additionally optimized using Swish activation, depthwise separable convolutions, and a cosine warm-up learning rate schedule to produce efficient gradient flow and convergence stability. Exhaustive ablation experiments validate the critical contribution of each architectural block, and the complete HASPNet obtains an accuracy of 93.87% (corresponding to a 6.13% error rate), an F1-score of 94%, and a reduced inference time of 21.33 ms, by a large margin surpassing top state-of-the-art backbones like MobileNetV2, DenseNet121, Inception-V3, Xception, and ResNet50 in both performance and computational efficiency. Additionally, activation function experiments validate Swish as the optimal non-linearity for this task. Interpretability is enhanced using Grad-CAM visualizations, which validate the model's attention on disease-specific leaf regions. Given the lack of existing benchmarks for the BDPapayaLeaf Dataset, HASPNet is evaluated against standard CNN backbones (MobileNetV2, ResNet50, etc.) to establish a performance-complexity baseline, justifying its selection for resource-constrained agricultural environments. The results validate the model's domain adaptability, and it is a strong candidate for real-world agricultural diagnostic systems and a valuable addition to vision-based plant pathology.