TriAttnNet based deep learning model for automated cotton pest detection and disease classification

This paper presents a deep learning model to detect cotton plant pests and classify diseases, which must overcome limited datasets, class imbalance, and feature redundancy. At the preprocessing phase, the Gaussian blur filtering and Contrast Limited Adaptive Histogram Equalization (CLAHE) are used to sharpen images by improving their clarity and contrast. To increase and diversify the data, Spa-GAN-based data augmentation is used to produce realistic synthetic samples. To obtain an accurate Region of Interest (RoI), an Attention-Guided Multi-Scale Residual U-Net (AGMS-U-Net) is considered to segment local and global structural information. The proposed framework makes three major contributions: (i) a new attention-based feature extractor, TriAttnNet, that incorporates spatial, channel, and contextual attention to represent diseases on a fine-grained level (ii) a new optimization strategy, Hybrid Mongoose Ray Chaotic Optimization (HMRCO), which includes chaotic strategies to better tune the parameters and explore the feature space and (iii) classification layer with focal loss for final decision. Experimental analyses prove that the suggested method is much more effective than the current state-of-the-art models, providing a powerful and understandable solution to precision agriculture and sustainable cotton crop health monitoring. Experimental results show that TriAttnNet achieves 98.66% accuracy, 98.71% recall, and 98.81% F1-score, which is better than the state-of-the-art algorithms, such as EfficientNetB1-CBAM (96.38%) and BERT-ResNet-PSO (95.69%). The proposed system is computationally feasible and interpretable, and it is interpretable to provide a practical solution to precision agriculture and sustainable monitoring of the health of cotton crops.