Content area
The increase in economic impacts due to the rapid growth of harmful algal blooms (HABs) in freshwater lakes has prompted the development of monitoring and prediction systems. However, limited target data hinders reliable predictions on HAB formations. To address this limitation, this dissertation focuses on providing solutions through data-driven, interpretable, and explainable machine learning models. In Chapter 3, an interpretable multivariate regression model such as Vector Autoregressive (VAR) was used to analyze the important factors that contribute to the formation of cyanobacteria in lakes. The causality tests performed on multiple water quality data helped to determine Alkalinity, Chlorophyll-a (Chl-a), and Water Temperature as the influential factors. The use of remote sensing satellite Sentinel-2 data for detection of one of the influential factors, Chl-a, using segmentation models such as Otsu and Random Forest was applied in chapter 4. Since microcystins are the most common cyanobacterial toxins found in freshwater bodies, further study in chapters 5 and 6 focused on applying multiple supervised machine learning classification models on microcystin toxicity levels in lakes. The proposal of a threshold detection framework using clustering and explainable AI was used in Chapter 7 to detect HAB hotspots in lakes. A new composite spectral index called AANI was proposed in Chapter 8 that proved to be effective for HAB hotspot detection from satellites. The dissertation also contribute the design of a real-time monitoring dashboard that can be used for HAB hotspot detection. Compared to the existing method of hotspot analysis, which is single-image-based analysis, the current approach provides a robust, explainable, and data-driven solution that can be applied to environmental prediction models.