Accurate and timely identification of insect species is a foundational requirement for precision agriculture, ecological monitoring, and integrated pest management. This paper presents a novel deep learning architecture for fine-grained acoustic classification of insect species using multichannel wingbeat recordings. Targeting critical agricultural use cases, the model is designed to differentiate between two morphologically and acoustically similar pest species, Halyomorpha halys and Nezara viridula, and a beneficial pollinator, Episyrphus balteatus. The proposed approach utilizes log-Mel and Per-Channel Energy Normalization (PCEN) spectrograms as input features and integrates a residual convolutional backbone with Squeeze and Excite blocks and Transformer-inspired self-attention modules to enhance spectro-temporal representation learning. Evaluation on a curated subset of the InsectSound1000 dataset demonstrates a classification accuracy of 88.73% and a macro-averaged F1 score of 88.78%. Class-wise analysis reveals that the model effectively captures subtle differences in wingbeat harmonics, achieving an AUC of 1.00 and AP of 0.99 for the pollinator class. The results suggest strong discriminative power across ecologically distinct classes and highlight the model’s potential for deployment in automated, non-invasive insect monitoring systems for real-time agricultural decision-making.