Effective control of agricultural pests determines both food security and the prevention of significant crop losses. Depending on entomological expertise and hand inspections, traditional methods of pest control take time and are prone to human error. Recent advances in deep learning and computer vision have improved pest identification; yet, existing models sometimes focus simply on visual input, therefore excluding critical environmental aspects such as temperature and humidity, which influence insect behavior. This work introduces Consistency Regularization (CR), Cross-modal Autoencoders (CMA), and Multi-modal Pseudo-information Guidance (MPIG) as a fresh approach to improve agricultural pest categorization. MPIG improves pest identification by combining environmental and visual cues using multi-modal data. Common latent representations produced by CMA across modalities help to ensure enhanced feature extraction and data fusion. By ensuring prediction consistency across several input perturbations, such as image noise, and environmental changes, CR increases model robustness. Our trials on the APTV-99 dataset, which includes 94 pest species, indicate that the proposed model surpasses traditional deep learning models such as CNN and ResNet50, exhibiting enhanced classification accuracy, precision, recall, and F1-score. The hybrid methodology (MPIG + CMA + CR) attained a testing accuracy of 91.8%, markedly enhancing pest identification by integrating supplementary data and reinforcing resilience by regularization. This research emphasizes the potential of multi-modal learning and consistency regularization to improve agricultural pest management tactics, hence enhancing crop protection and promoting sustainable agriculture.