Retrieval of total suspended solids concentration from hyperspectral sensing using hierarchical Bayesian model aggregation for optimal multiple band ratio analysis

• Main Authors: Pak, Hui Ying, Law, Adrian Wing-Keung, Lin, Weisi
• Other Authors: Interdisciplinary Graduate School (IGS)
• Format: Article
• Language: English
• Published: 2023
• Subjects: Engineering::Environmental engineering; Remote Sensing; Water Quality Monitoring
• Online Access: https://hdl.handle.net/10356/164400
• Institution: Nanyang Technological University

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 by setting up gauging stations, which can be labour intensive and costly. Recently, the possibility of monitoring water quality through remote sensing with Unmanned Aerial Vehicles (UAVs) and hyperspectral sensors has shown great promise, with the key advantages of larger spatial coverage and possibly higher accuracy enabled by higher spectral resolution and more extensive data. Correspondingly, more advanced methods need to be established for hyperspectral analysis for water quality determination to capitalize on this wealth of information. In this study, a new method called Hierarchical Bayesian Model Aggregation for Optimal Multiple Band Ratio Analysis (HBMA-OMBRA) has been developed as a proof-of-concept for estimating Total Suspended Solids (TSS) concentrations from the hyperspectral data. The method leverages on the Bayesian ensembling of competing models because there is not a single best working model for all situations. It also encompasses a new approach called Ensemble Band Ratio Selection (ENBRAS) for the identification of best candidate band ratios (BBRs) via a set of ensembling and “bagging” procedures, followed by a modified Batchelor Wilkin's algorithm to cluster the candidate band ratios. A laboratory investigation was conducted in the present study to measure the hyperspectral reflectance in different experiments under various environmental conditions to verify the robustness of HBMA-OMBRA. From the experimental results, six distinct clusters of candidate BBRs were identified using ENBRAS. In particular, two clusters in the red, green, and near infrared spectrum showed the largest contribution. The significance of multi-clusters provides an explanation for previously contrasting results reported in the literature and some evidence for reconciling these findings.