Research on Chlorophyll-a Water Quality Parameter Inversion Based on Multi-Scale Attention Fusion Network Model
SUN Bang-yong1, 2, GONG Kai-jie1, YU Tao2*, BIE Qian-wen3
1. Faculty of Printing, Packaging Engineering and Digital Media Technology, Xi'an University of Technology, Xi'an 710048, China
2. Key Laboratory of Spectral Imaging Technology, Xi'an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences, Xi'an 710119, China
3. TÜV Rheinland (Guangdong) Co., Ltd., Guangzhou 510700, China
Abstract：Water resources are one of the core elements of the ecological environment, and there is currently a large number of water bodies being polluted by industrialization or nutrient enrichment, making real-time monitoring of water quality parameters crucial for maintaining the health of water bodies. Traditional water quality monitoring methods often use on-site sampling measurement or linear regression prediction methods. Still, due to the significant non-linear characteristics of water quality parameters, traditional water quality monitoring methods are time-consuming and inaccurate in their predictions. In recent years, deep learning methods have shown good performance in dealing with complex non-linear problems and have been applied by many scholars to the inverse estimation of water quality parameters. However, water quality inversion models based on deep learning still suffer from inaccurate analysis of remote sensing spectral images and poor model generalization capabilities. Therefore, this paper proposes a water quality inversion network model based on a multiscale attention fusion mechanism, which can accurately predict water quality parameters such as chlorophyll-a, providing a basis for assessing the health of water bodies. The network integrates advanced attention mechanisms and feature fusion strategies, combining the advantages of local feature learning from CNN and global feature extraction capabilities from Transformer to construct a Dense ASPP module for obtaining multiscale features of remote sensing images. It uses a channel attention decoder and pooling fusion modules to extract detailed features. Then, it estimates the concentration of chlorophyll-a by fusing different scales and levels of feature information, achieving higher inversion accuracy and generalization performance. To validate the performance of the proposed inversion model, experiments were implemented in Python 3.7 and the PyTorch framework, using ocean remote sensing spectral images and chlorophyll-a concentration data from January 2021 to December 2022 for network training. The experiment compares the proposed method with seven other water quality inversion methods, achieving the best performance in all objective indicators. It improves the R2 index by 0.09 compared to the best method in the comparison, and reduces the RMSE, MAE and MAD indices by 11.99, 0.089, and 0.029, respectively, and improves the Evar index by 0.098, and the NSE index by 0.041. Meanwhile, in subjective evaluation, the proposed method obtains more precise chlorophyll-a concentrations, smaller errors, and higher generalization ability in different waters.
孙帮勇，巩凯杰，于 涛，别倩雯. 基于多尺度注意力融合的叶绿素a水质参数反演研究[J]. 光谱学与光谱分析, 2025, 45(04): 1190-1200.
SUN Bang-yong, GONG Kai-jie, YU Tao, BIE Qian-wen. Research on Chlorophyll-a Water Quality Parameter Inversion Based on Multi-Scale Attention Fusion Network Model. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2025, 45(04): 1190-1200.
