Agricultural insect pests are a significant cause of crop damage, often leading to reduced crop yield. Early detection of pests is crucial for effective pest management, but this remains challenging due to the large biodiversity of insects encountered outdoors. Traditional methods for pest detection are labor-intensive and prone to human error. Recent advancements in computer vision and deep learning have shown promising results in automating pest detection. However, the collection of large datasets for training remains a bottleneck, particularly for rare insect species. This work explores the use of Few-Shot Learning (FSL), specifically Prototypical Networks, for insect pest classification in scenarios where only a few labeled images are available for training. We combine three distinct pest datasets to test the generalization capabilities of Prototypical Networks in diverse outdoor settings. Additionally, we investigate the use of various distance metrics, including Cosine Similarity, Euclidean Distance, Mahalanobis Distance, and Wasserstein Distance, among others, to enhance the performance of Prototypical Networks. Our results show that Mahalanobis Distance provides superior classification performance, particularly for classes unseen during training. We also explore the impact of Squeeze-and-Excitation (SE) attention layers on improving feature extraction, demonstrating their effectiveness in enhancing the robustness of Prototypical Networks.