GRASSHOPPER OPTIMIZED VISION TRANSFORMER WITH ADAPTIVE TOKEN CALIBRATION DROPOUT SCALING QKV ATTENTION REFINEMENT IN COTTON LEAF DISEASE IDENTIFICATION

Accurate identification of cotton leaf diseases remains a complex task due to high visual similarity among disease types, irregular lesion boundaries, background interference, and the early-stage subtlety of symptoms in field-captured images. Current literature lacks transformer-based disease identification frameworks that dynamically regulate attention, token relevance, and learning stability in response to lesion morphology. This study bridges that gap by introducing an optimization-driven attention control mechanism that generates new insights into lesion-focused transformer learning for precision agriculture. Traditional models and vanilla Vision Transformers struggle with fixed patch segmentation, redundant token processing, and diluted attention focus, leading to suboptimal classification. To address these challenges, this research proposes a novel model titled Grasshopper Optimized Vision Transformer (GO-ViT), engineered for lesion-aware feature refinement in cotton leaf disease identification. GO-ViT integrates swarm-based Grasshopper Optimization to dynamically configure patch embedding size, attention head scaling, dropout distribution, QKV matrix refinement, and entropy-based token pruning. This biologically inspired regulation emulates adaptive foraging patterns, ensuring high lesion focus and reduced attention waste. GO-ViT enhances inter-class separation, preserves spatial continuity, and minimizes false activation over non-symptomatic regions. Experimental evaluation across five performance metrics confirms GO-ViT’s robustness, with 94.491% Overall Detection Efficiency and 88.985% Balanced Class Correlation. The model demonstrates superior ability to isolate overlapping infections, suppress background noise, and deliver stable predictions across diverse cotton leaf conditions. GO-ViT stands as a reliable solution for scalable, high-precision disease classification in precision agriculture, combining visual attention learning with nature-driven optimization control.