Application of airborne hyperspectral imagery to retrieve spatiotemporal CDOM distribution using machine learning in a reservoir

Abstract

Colored dissolved organic matter (CDOM) in inland waters is used as a proxy to estimate dissolved organic carbon (DOC) and may be a key indicator of water quality and nutrient enrichment. CDOM is optically active fraction of DOC so that remote sensing techniques can remotely monitor CDOM with wide spatial coverage. However, to effectively retrieve CDOM using optical algorithms, it may be critical to select the absorption coefficient at an appropriate wavelength as an output variable and to optimize input reflectance wavelengths. In this study, we constructed a CDOM retrieval model using airborne hyperspectral reflectance data and a machine learning model such as random forest. We evaluated the best combination of input wavelength bands and the CDOM absorption coefficient at various wavelengths. Seven sampling events for airborne hyperspectral imagery and CDOM absorption coefficient data from 350 nm to 440 nm over two years (2016–2017) were used, and the collected data helped train and validate the random forest model in a freshwater reservoir. An absorption coefficient of 355 nm was selected to best represent the CDOM concentration. The random forest exhibited the best performance for CDOM estimation with an R2 of 0.85, Nash-Sutcliffe efficiency of 0.77, and percent bias of 3.88, by using a combination of three reflectance bands: 475, 497, and 660 nm. The results show that our model can be utilized to construct a CDOM retrieving algorithm and evaluate its spatiotemporal variation across a reservoir.

Keywords

• CDOM
• Hyperspectral imagery
• Reflectance band selection
• Absorption coefficient
• Machine learning
• Spatiotemporal distribution