STRATEGIC DIRECTIONS FOR THE APPLICATION OF ARTIFICIAL INTELLIGENCE IN MICROBIOLOGICAL MONITORING OF WATER BODIES: EXPLANABLE MODELS AND MULTI-MODE INTEGRATION

Abstract

This article provides a critical analysis of current artificial intelligence approaches used in the microbiological monitoring of water bodies. It demonstrates that the high accuracy reported in many deep learning studies—particularly those based on datasets such as DIBaS—is frequently the result of methodologically improper data splitting, which leads to overestimating model performance. When more rigorous validation strategies that simulate real-world conditions (e.g., strain-wise splitting) are applied, accuracy drops significantly, revealing a generalization crisis. The study substantiates the need to transition from traditional convolutional neural networks toward explainable and multimodal AI architectures that integrate microscopic images with genomic data (k-mers). This integration improves interpretability, resolves phenotypic ambiguity, and enhances the ability of models to classify previously unseen bacterial strains. Two strategic directions are proposed: the development of Self-Explainable AI for verifying diagnostic features and the creation of multimodal systems capable of predicting functional microbial characteristics. These findings lay the groundwork for next-generation ecological monitoring tools