POSTER Session 4

LONG-TERM MONITORING OF CHLOROPHYLL-A IN SMALL INLAND WATERS USING LANDSAT HERITAGE AND MACHINE LEARNING MODELS

Despite the well-established utility of Landsat heritage in land-based applications, its effectiveness in water-quality assessment has been limited by inconsistent results owing to insufficient spectral and radiometric resolution for aquatic contexts. This study addresses this gap by leveraging a paired dataset spanning 38 years (1984-2021) comprising chlorophyll-a (Chl-a) measurements and near-coincident satellite-derived reflectance from Landsat 5, 7, and 8. We demonstrate the capability of Landsat in generating accurate time-series of Chl-a across seven eutrophic prairie lakes in Saskatchewan, Canada. Subsequently, we propose a workflow to refine Landsat Chl-a retrieval, validated against high-frequency in-situ measurements. Our proposed methodology enhances the accuracy of Chl-a retrieval using Landsat imagery by ~12%. The proposed approach utilizes machine learning (ML) models trained on in-situ Chl-a data acquired through diverse methods such as high-performance liquid chromatography (HPLC), spectrophotometry, and field fluorometry. Notably, we find that a local mixture density network (LMDN) model consistently outperforms other ML techniques in Chl-a estimation. Our study reveals the importance of achieving a critical threshold of 250-300 lake-satellite matchups for robust training of local ML models, irrespective of sensor type. Furthermore, the workflow is adaptable to all three sensors within the Landsat heritage, improving accuracy in Chl-a time-series while enhancing data frequency and consistency across sensor platforms. These processes pave the way for the development of locally-trained ML models for inland waterbodies in diverse regions, facilitating the quantification of regional variations in lake productivity.

Amirmasoud Chegoonian, University of Waterloo, University of Regina, Canada, [email protected], https://orcid.org/0000-0003-0862-2763

Claude Duguay, University of Waterloo, Canada, [email protected]

Peter Leavitt, University of Regina, Canada, [email protected]

Code of Conduct
Privacy Policy