A UV-DOAS-Based Multi-Scale Interaction Attention Network for Simultaneous Retrieval of NO and NO2 in Gas Mixtures

Accurate monitoring of nitrogen oxides (NOx) is essential due to their adverse effects on environmental quality and public health. To address the spectral overlap between NO and NO2 in ultraviolet differential optical absorption spectroscopy (UV-DOAS), we propose a multi-scale dual-branch interaction attention network (MDIAN) for simultaneous concentration retrieval in gas mixtures. The model employs a dual-branch multi-scale convolutional architecture to extract local narrow-band absorption details and broad spectral profile features. A cross-attention mechanism is introduced to enable feature interaction between the NO and NO2 branches. A bidirectional long short-term memory (Bi-LSTM) network is further incorporated to model contextual dependencies along the wavelength dimension, enabling joint regression of both target gases. Experimental results show that the proposed model achieves mean absolute errors (MAE) of 0.076 ppm for NO and 0.062 ppm for NO2, with coefficients of determination (R2) of 0.9998 for both gases, outperforming traditional regression methods and baseline deep learning models. The uncertainties are 0.69% and 0.76%, respectively, and the inference time per sample ranges from 48.9 to 74.5 ms. These results indicate that MDIAN achieves a favorable balance among accuracy, stability, and real-time performance, offering a promising approach for intelligent monitoring of complex gas mixtures using UV-DOAS.