<p>In agriculture, preserving crop health, increasing yield, and minimizing financial losses all depend on the quick and accurate identification of cotton diseases. Traditional manual inspection and conventional methods are often inaccurate, inefficient, and impractical for large-scale implementation. Existing disease detection techniques struggle with poor generalization, high false positive rates, and computational inefficiency. The lack of robust, scalable, and real-time solutions limits precision agriculture, making automated disease identification essential for sustainable farming. This study proposes the Slimmable Pruned Graph Sample and Aggregate Neural Network optimized with Billiards-Inspired Optimization Algorithm (SPGSANN-BOA) for accurate and efficient cotton disease classification. The framework integrates Bilateral Texture Filtering (BTF) for preprocessing, Dual Attention Multi-Head Generative Adversarial Network (DAMHGAN) for segmentation, and SPGSANN for feature extraction and classification, with weight optimization using BOA to enhance accuracy and generalization. When tested on the Cotton Leaf Disease Dataset, SPGSANN-BOA successfully identifies healthy leaves, bacterial blight, curl virus, growth damage caused by herbicides, leaf hopper jassids, reddening, and variegation. 99.72% accuracy, 98.65% precision, 98.7% recall, 99.22% specificity, 97.68% F1-score, and 0.28% error rate are all attained by the suggested method. The suggested model exhibits better classification performance than current techniques, guaranteeing greater dependability for real-time cotton disease diagnosis. The combination of bio-inspired optimization, GAN-based segmentation, and graph-based deep learning (DL) yields a real-time and scalable cotton disease detection system. This work contributes to precision agriculture by enhancing automated disease diagnosis, reducing economic losses, and promoting sustainable farming practices.</p>