Weighted feature fusion of dual attention convolutional neural network and transformer encoder module for ocean HABs classification

Abstract

Accurately classifying harmful algal blooms (HABs) is crucial in red tide warning. However, the current automatic recognition system for real-time HABs image classification has a low accuracy rate of only about 70% and a high computational complexity. To address these issues, we propose a Weighted Feature Fusion of Dual Attention Convolutional Neural Network and Transformer Encoder Module (WFTC) for Ocean HABs Image Classification. First, we design an image enhancement algorithm specifically for HABs images, which significantly improves the quality of real-time dynamic sampling data by enhancing contrast, edges, and texture features. Second, the transformer encoder module is utilized as the first branch of WFTC to generate a feature vector representation with global semantic information. Then, we combine the attention mechanism with the convolutional neural network as the second branch of WFTC, which is used to emphasize the main feature information of HABs. Finally, the extracted features are weighted fusion by combining the characteristics of the transformer encoder module and dual-attention convolutional neural network. The experimental results using real-world data from HABs in the Chinese nearshore waters in 2020 demonstrate that compared with other classification networks, the proposed WFTC method achieves the highest accuracy rate of 87% and competitive results in terms of stability and convergence.