A Lightweight Dual-Attention Neural Network for In-Situ Hyperspectral Classification of Microalgae

Abstract

Accurate monitoring of microalgae is essential for assessing marine ecological health and preventing harmful algal blooms in ocean engineering. Current in situ identification methods often suffer from limited discriminative feature extraction and inadequate adaptation to complex underwater imaging conditions. This study introduces a lightweight dual-attention neural network, termed ANMM, designed for real-time, in situ hyperspectral classification of microalgae within integrated underwater monitoring systems. The model strengthens a deepened AlexNet backbone with multi-head latent attention (MLA) and multi-head self-attention (MSA) mechanisms, which jointly enhance local feature refinement and global spectral dependency modeling. An early-stopping strategy is further incorporated to prevent overfitting and ensure robust generalization. Evaluated on a custom dataset of field-collected fluorescence spectra, the model achieves a classification accuracy of 98.91%, outperforming several state-of-the-art deep-learning counterparts. With a compact parameter size of 16.34 M and low-latency inference on edge hardware, the system demonstrates strong potential for deployment on embedded underwater sensing platforms. This work provides a practical and efficient AI-driven solution for continuous marine microalgae monitoring, supporting advances in ocean observation technology and ecological engineering.