<p>Butterfly species recognition is essential for biodiversity monitoring and pest management. However, traditional approaches based on RGB images are limited by their reliance on color and shape features, making them sensitive to visual ambiguities such as similar color patterns across species and variations in wing shape or appearance during flight. While hyperspectral imaging (HSI) provides rich spectral information, its conventional use relies on the reconstruction and analysis of an HSI datacube, which is computationally intensive and unsuitable for real-time applications. We address this limitation by proposing a framework that bypasses the need for datacube reconstruction. Instead, it operates directly on raw spatio-spectral images captured by a compact HSI camera before reconstruction, resulting in limited spectral data for the detected butterfly. Assuming that the spectral features of butterfly species follow Gaussian distributions, we observe overlaps between some distributions due to spectral similarities among some species. To overcome this, we introduce a classification approach based on a convex combination of Gaussian Naive Bayes and Z-score methods. This approach proves more effective, yielding better results when butterflies are detected in the near-infrared wavelengths. We achieve an accuracy of 88.75% using near-infrared information from a single spatio-spectral image, reaching up to 97.5% when aggregating spectral information from a sequence of four spatio-spectral images.</p>