Automated imaging and machine learning for soil bacteria classification: Challenges and insights

Soil bacteria play a vital role in biofertilizer production, making accurate classification of bacteria genera essential for optimizing industrial processes. While previous studies report high classification accuracy, they often overlook the challenges of deploying such systems in real-world production environments, where image acquisition must be automated rather than conducted by the microbiologists manually. This study addresses these practical issues by investigating the automated classification of bacteria images under varying conditions of lighting and glass type. Both classical machine learning approaches and advanced neural networks, including modifications of Extreme Learning Machines - Radial Basis Function and deep convolutional architectures of Residual Neural Networks, were employed. Unforeseen challenges impacting the reliability of automated classification were identified, highlighting novel factors that could significantly affect classification accuracy. This work not only emphasizes the limitations of existing approaches in production settings but also proposes possible directions for future research in automating bacteria classification for industrial biofertilizer development. Importantly, it brings attention to a broader issue: the critical need for rigorous dataset preparation and validation when applying artificial intelligence to microscopic image analysis across biological domains.