Monitoring multi-water quality of internationally important karst wetland through deep learning, multi-sensor and multi-platform remote sensing images: A case study of Guilin, China

• Published in: Ecological indicators 2023-10, Vol.154, p.110755, Article 110755
• Main Authors: Yang, Wenlan, Fu, Bolin, Li, Sunzhe, Lao, Zhinan, Deng, Tengfang, He, Wen, He, Hongchang, Chen, Zhikun
• Format: Article
• Language: English
• Subjects: Deep learning and machine learning; Karst wetlands; Model inversion; Multispectral and hyperspectral images; UAV and satellite platform; Water quality parameters
• ISSN: 1470-160X
• DOI: 10.1016/j.ecolind.2023.110755

•Remote sensing estimating multi-water quality parameters in the largest karst wetland of China.•Tansformer-based retrieval model presents a better performance than other ML models.•UAV images outperform satellite remote sensing images in inversion of water quality.•Huixian Karst Wetland of International Importance has displayed a trend of eutrophication.•Spectral range from 660 nm to 840 nm is sensitive to capture water quality changes of karst wetland. Karst wetlands are widely distributed throughout the southwest China, and play an important role in enhancing carbon sequestration and improving water quality in karst areas. The internationally important karst wetland of Huixian is the largest karst wetland in China, but its water quality has continued to deteriorate as a result of human influences in recent years. Remote sensing technology has become an important approach to estimate water quality parameters (WQPs). However, the feasibility of combining multi-sensor remote sensing images with deep learning to estimate different WQPs in karst wetlands has not been demonstrated yet. To resolve this issue, this study constructed multiple retrieval models of WQPs (Chlorophyll-a (Chla), Phycocyanin (PC), Turbidity (Turb), Dissolved Oxygen (DO)) in karst wetlands using deep learning (Transformer and Mixture Density Network (MDN)) and optimized shallow machine learning (Random Forest (RF), XGBoost (XGB) and Gradient Boosting (GB)) based on multi-sensor images from satellite and UAV platforms. The performance of deep learning in the inversion of WQPs demonstrated to compare with shallow machine learning using multispectral and hyperspectral images. We further quantitatively evaluated the retrieval performance of UAV and satellite, multispectral and hyperspectral images, and presented predictive mapping of the gradient distribution of WQPs. Finally, this study adopted the SHapley Additive exPlanations (SHAP) to tackle the local and global interpretability of the input features contribution to the output of retrieval models. The results showed that (1) Transformer model presented a good prediction of PC and DO (R2 = 0.649 ∼ 0.844), XGB and GB models achieved the highest accuracy estimation of Chla and Turb (R2 = 0.75). (2) The estimation results of WQPs based on UAV platform (R2 = 0.419 ∼ 0.695) was higher than that of satellite-based images. The estimation accuracy of multispectral images (R2 = 0.338 ∼ 0.718) was slightly higher than that of Zhuhai-1 Orbita hyper