Insect pests pose a significant threat to crop yields, leading to considerable economic losses in agriculture. This study presents two innovative deep learning-based methods for intelligent pest detection. The first approach integrates the Fused-MBConv architecture with a Gradient-based Multinomial Logistic Regression (GMLR) activation function to enhance classification performance through efficient feature extraction and improved non-linear decision boundaries. The second method leverages hyperspectral correlation features derived from normalized cross-correlation matrices, exploiting unique spectral signatures of insects for precise identification. A dataset of 3,516 real-world insect images spanning 12 species was used for training and evaluation. Experimental results demonstrated that the proposed FMBC-GMLR model achieved a classification accuracy of 96%, outperforming conventional models such as ResNet50, YOLOv3, and MobileNet in precision, recall, and F1 score. These findings underscore the model’s robustness and potential for deployment in smart farming systems to enable early pest detection and reduce crop damage.