PestReKNet-X: Integrating Explainable AI to enhance pest disease detection and combat crop senescence

Crop pests and diseases remain a significant obstacle to sustainable
agriculture, necessitating innovative and eco-friendly detection
solutions. This study introduces PestReKNet-X, a state-of-the-art
explainable deep learning framework that combines a ResNet18 backbone with
a custom Kolmogorov Arnold Network (KAN) Linear layer (KANLinear), which
captures complex non-linear patterns, surpassing the limits of traditional
fully connected layers. The framework is evaluated on the benchmark CCMT
crop pest and disease dataset, containing 102,097 images across 22
classes. To tackle class imbalance, the Mean Intersection over Union (IoU)
metric is used alongside accuracy for robust performance evaluation. The
model achieves a testing accuracy of 95.09% and a Mean IoU of 0.9133,
reflecting strong generalization across diverse categories. Moreover,
PestReKNet-X surpasses advanced architectures like Swin Transformer and
MobileNetV3Large concerning performance and reliability. Monte Carlo
Dropout for uncertainty estimation and model calibration is used to
achieve reliable predictions with well-calibrated probabilities. A strong
focus on explainable AI (XAI) ensures transparency and interpretability,
addressing gaps in recent studies. The inclusion of Grad-CAM and LIME
provides intuitive visualizations and localized insights, enhancing
understanding of predictions. With high accuracy, efficiency, and
interpretability, PestReKNet-X provides a scalable solution for precise
pest and disease monitoring.