Abstract

A growing body of literature demonstrates the feasibility of developing robust predictive models for microorganisms by integrating Raman micro-spectroscopy with chemometric techniques, frequently resulting in high-precision outcomes. While advancements in machine learning and broader software accessibility have simplified the generation of predictive models from intricate datasets, the underlying mechanisms driving such high accuracy remain insufficiently explored. This research utilizes Raman spectroscopic data from fungal spores and carotenoid-bearing microorganisms to reveal that precise classification often does not depend on specific peak positions or subtle variations in band ratios related to chemical composition. Instead, characteristic baseline effects in biochemically analogous microorganisms, which may be intensified by particular data pretreatment methods, or even seemingly neutral spectral regions, can hold significant weight for convolutional neural networks (CNNs). By employing Gradient-weighted Class Activation Mapping (Grad-CAM), this study seeks to demystify the "black box" of CNNs in microbiological contexts, identifying the specific Raman spectral regions fundamentally responsible for successful classification.