Motif field combined with two-stream feature fusion network and double detection head for identification and...

Accurate and rapid detection of microalgae species and concentrations in water bodies is important for water quality management and microalgae culture. This study proposes a new method for quantitative analysis of microalgae by fluorescence and absorption dual-mode spectroscopy combined with sequence motif and deep learning. First, the motif field (MF) is proposed and utilized to encode the spectral data into two-dimensional images to discover potential connections between wavelengths. After that, a two-stream feature fusion (TSFF) network is designed based on convolutional network and attention mechanism for feature extraction and fusion of dual-mode spectra. Finally, a dual detection head (DDH) capable of simultaneous microalgae species identification and concentration regression was designed using convolution head (conv-head) and fully connected head (fc-head). The results show that the proposed method has excellent performance in the prediction of microalgae species and concentration, with a classification accuracy of up to 100 %, while the RMSE and R² of predicted concentration are 0.0181 cells/ml and 0.9941, respectively. This work demonstrates the feasibility and superiority of fluorescence and absorption dual-mode spectroscopy supplemented with image encoding and deep learning in microalgae detection, and lays the technical foundation for exploring similar methods in related fields.