Deep learning for water quality multivariate assessment in inland water across China

• Published in: International journal of applied earth observation and geoinformation 2024-09, Vol.133, p.104078, Article 104078
• Main Authors: Ali, Aamir, Zhou, Guanhua, Pablo Antezana Lopez, Franz, Xu, Chongbin, Jing, Guifei, Tan, Yumin
• Format: Article
• Language: English
• Subjects: Deep neural network; GLORIA; Inland water body; Machine learning; Sentinel-2; Water quality parameters
• ISSN: 1569-8432
• DOI: 10.1016/j.jag.2024.104078

•Demonstration of limited but representative training dataset for efficient modeling.•Robust DNN models for independent and simultaneous retrieval of Chl-a, TSS and SDD.•Better performance of DNN over XGBoost, RF, and SVM.•Applicability of models on heterogeneous lakes.•Challenges of significant water quality degradation trends in Chinese lakes. Remote sensing of optically complex inland waterbodies is challenging due to the complex nonlinear correlation between water quality parameters and water optical properties. However, integration of deep learning techniques and representative datasets offers the potential to address these challenges effectively. This study aims to develop robust deep learning models, utilizing limited but highly representative dataset of in-situ water quality and radiometrically corrected hyperspectral remote sensing reflectance (Rrs) measurements collected from optically diverse lakes of China, for independent and simultaneous retrieval of Chlorophyll-a (Chl-a), Secchi Disk Depth (SDD), and Total Suspended Solids (TSS) using Sentinel-2 analysis ready products. The GLObal Reflectance community dataset for Imaging and optical sensing of Aquatic environments (GLORIA) provides over 400 such measurements for Chinese lakes, which are simulated to Sentinel-2 Rrs with its spectral response function to build a representative dataset. Using this dataset, Multilayer Perceptron (MLP) based Deep Neural Network (DNN) models are developed and compared with eXtreme Gradient Boosting (XGB), Random Forest (RF), and Support Vector Machine (SVM) algorithms. The DNN models outperformed in effective evaluation of Chl-a (Root Mean Squared Error (RMSE) = 14.18 mg/m3), TSS (RMSE=7.23 g/m3) and SDD (RMSE=0.12 m) on test datasets and Chl-a (RMSE=14.42 mg/m3) and SDD (RMSE=0.07 m) against Sentinel-2A validation dataset of Liangzi lake. Mixed Density Network (MDN) model showed less accuracy for Chl-a (RMSE=16.76 mg/m3) on same validation dataset. Impact of different atmospheric correction processors is also assessed and DNN models achieved their accuracy on Sentinel-2 Atmospheric Correction (Sen2Cor) processor. Finally, water quality maps for various lakes in China are produced showing realistic ranges. These results show the potential of DNN models trained with limited but representative dataset in practical applications for spatial and temporal analysis of water quality.