Author(s): Hui Ying Pak; Adrian Wing-Keung Law; Weisi Lin

Linked Author(s): Hui Ying Pak

Keywords: Remote sensing; Water quality monitoring; Bayesian model averaging; Clustering; Hyperspectral

Abstract: Water quality monitoring plays an essential role in water resource management and water governance. At present, the monitoring is commonly conducted via in-situ sampling and/or through the setting up of gauge stations, which can be labour intensive and costly. Recently, the possibility of monitoring the water quality through remote sensing with Unmanned Aerial Vehicles (UAVs) and hyperspectral sensors show great promises, with the key advantages of larger spatial coverage and possibly higher accuracy with the fine frequency band resolutions and more extensive data. Correspondingly, more advanced methods need to be established to capitalize on the richer information. Current machine learning methods for water quality retrieval from remote sensing data, such as Optimal Band Ratio Analysis (OBRA) and XGBoost, entail some limitations such as statistical inconsistencies in identifying the “best” band ratio and poor inference/interpretability. In this study, a new method called Hierarchical Bayesian Model Aggregation for Optimal Multiple Band Ratio Analysis (HBMA-OMBRA) was developed to overcome these limitations as a proof-of-concept for the retrieval of Total Suspended Sediments (TSS) concentration. It leverages on Bayesian ensembling of competing models since there is not a single best working method in all situations, instead of the traditional paradigm where a “best” model is selected. HBMA-OMBRA also encompasses a new method called the Ensemble Band Ratio Selection (ENBRAS) for the identification of candidate “best” band ratios (BBRs) via ensembling and “bagging” procedures, and a modified Batchelor Wilkin’s algorithm was employed to cluster the candidate band ratios. The clustering procedure and hierarchical structure in HBMA-OMBRA allows the quantification of uncertainties for each band ratio cluster and sidesteps the limitation of Bayesian Model Averaging (BMA), which does not permit high dimensional covariates due to the exponential expansion in the models’ computation. HBMA-OMBRA extends OBRA and BMA as a more robust framework by capitalizing on the rich high-dimensional data, offering interpretability using ensemble methods as the backbone. A laboratory study was conducted to measure remote sensing reflectance under various simulated environmental conditions which include combinations of varying illumination-sensor geometry, TSS concentrations (0 – 300 mg/l) and wave effect. This seeks to verify the robustness of HBMA-OMBRA in predicting a large range of TSS concentrations under various environmental conditions. Six distinct groups of candidate BBRs were identified under ENBRAS, and results have shown that HBMA-OMBRA performs significantly better than individual OBRA models in the analysis of the experimental results, with an average RMSE of 32 and 42 respectively. In addition, two clusters of candidate BBRs in the red and near infrared spectrum showed the largest contribution in HBMA-OMBRA, which provides the reconciliation of previously differing results in the literature.

DOI: https://doi.org/10.3850/IAHR-39WC2521711920221294

Year: 2022