A novel quad-modality deep neural network for estimating chlorophyll-a concentrations in Lianyungang's lakes and...

Eutrophication has become a growing threat to lake ecosystems, driven by climate change and intensified human activity. Chlorophyll-a (Chl-a), as a proxy for trophic status and algal biomass, plays a central role in water quality assessment and lake management. To estimate Chl-a concentrations, this study integrates in situ measurements with Sentinel-2 satellite imagery collected from lakes and reservoirs in Lianyungang City, China, between 2017 and 2024. Based on this dataset, we developed a novel Quad-Modality Deep Neural Network (QM-DNN) grounded in the concept of multimodal learning. The model incorporates dual-band, tri-band, and quad-band spectral features combinations, along with auxiliary temporal and environmental variables. Experimental results on the test set (N = 197) show that the QM-DNN model outperforms the traditional Deep Neural Network (DNN) and machine learning methods, achieving an R² of 0.73, MAE of 6.76 mg/m³, and RMSE of 12.65 mg/m³. Spatiotemporal analysis from 2017 to 2024 revealed significant variability in Chl-a concentrations. Spatially, lower concentrations were observed in the northern water bodies (mesotrophic, 6-10 mg/m³), while the central and southern regions exhibited higher levels (mildly eutrophic, 10-26 mg/m³). Temporally, mesotrophic waters increased from 7.06 % to 11.14 %, while mildly eutrophic waters slightly declined, and moderately eutrophic (26-30 mg/m³) areas remained low and stable. Correlation analysis based on ERA5 climate data identified air temperature and precipitation as key climatic drivers of Chl-a variability, both significantly positively correlated with Chl-a concentrations (p < 0.05), whereas wind speed exhibited a weaker, site-dependent effect. To enhance model interpretability, the integrated gradients method was employed to quantify the relative contribution of each input feature. These findings highlight the effectiveness of multimodal deep learning in remote sensing-based Chl-a monitoring, providing robust support for regional water quality management and eutrophication mitigation strategies.