An Explainable Deep Learning Network for Environmental Microorganism Classification Using Attention-Enhanced Semi-Local Features

Abstract

Environmental micro-organisms (EM) play a crucial role in ecosystems, contributing to nutrient cycling, bioremediation and, supporting overall biodiversity. The identification and classification of these microbes are of significant importance for environmental monitoring and management. However, traditional detection methods of EM detection are labour-intensive, time-consuming and, require extensive material resources. Deep learning (DL) offers a more efficient and accurate solution to this problem by automating the detection and classification of EMs. This research introduces a novel network for EM classification using Dense network, Efficient Hierarchical Channel Refinement (EHCR) block, Dilated Adaptive Shuffled-Attention (DASA) block and Neighbourhood Attention (NA). The DenseNet-169 architecture is modified to integrate the EHCR and DASA blocks, allowing the model to focus on semi-local feature extraction for improved feature representation. NA efficiently targets relevant regions within the extracted features, enhancing EM classification performance. Additionally, the model incorporates Grad-CAM-based Explainable Artificial Intelligence (XAI) to improve interpretability, providing visual insights into the decision-making of the model. The performance of the proposed network surpassed state-of-the-art architectures and achieved an accuracy of 82.74% on the EMDS-6 dataset.

Keywords

• CNN
• deep learning
• DenseNet
• EM
• neighbourhood attention