A changing climate for lake monitoring - harnessing satellite imagery and machine learning for tracking and forecasting algal blooms

Abstract

Bloom-forming algae, particularly toxin-producing cyanobacteria, pose significant risks to inland water quality. Consequently, developing algae monitoring and forecasting capabilities is critical for understanding drivers, tracking long-term trends, and delivering early warnings. Satellite remote sensing and machine learning (ML) show considerable promise for addressing these challenges, but at present, operational algal bloom monitoring and forecasting programs are rare, and have only been implemented for large, economically-significant lakes. Therefore, extending these capabilities to smaller, less well-studied lakes is critical. Consequently, it is important to understand data requirements for ML forecasting and explore ways to monitor lakes which have, historically, been too small to study using satellites. Furthermore, climate change may exacerbate algal blooms and diminish the predictability of freshwater ecosystems, and so it is critical to understand and prepare for this. This thesis addresses these challenges through three novel studies: Chapter two explores data requirements for short-term ML algae forecasts. Subsequently, Chapter three evaluates the impact of several climate change scenarios on ML water temperature forecast performance. Finally, Chapter four evaluates the capabilities of new, high-resolution satellite imagery for monitoring small lakes. It is found that ML forecasts require approximately five or more years of fortnightly training data to be effective, but that performance improvements show diminishing returns as more training data are used. Furthermore, more extreme climate change scenarios are likely to diminish forecast performance. However, high-resolution satellite imagery may help to address these challenges with increased sampling frequency that can potentially offset climate-induced performance losses; and potential for widespread lake monitoring, which may lead to a step-change in the size of training datasets readily available for forecasting. Collectively, this work highlights potential future directions for algal monitoring and forecasting practices, and ultimately underlines the importance of harnessing synergies across diverse data sources for addressing present and future water quality challenges.