CMEAP: Cluster Modeling With Explainability for Threshold Identification of Spectral Indices in Algal Bloom Prediction

Cyanotoxins in freshwater lakes are creating huge economic impacts, prompting state agencies to monitor and predict their formation. Satellite-based remote sensing is often used for early prediction of Harmful Algal Blooms (HABs), but limited literature exists on identifying and generalizing the threshold for toxic bloom formation. Most existing work focused on using a predefined and empirical threshold for spectral indices to identify hotspots of HABs in lakes. However, in this work, a new framework called Cluster Modeling with Explainability for Algal bloom Prediction (CMEAP) is proposed that uses unsupervised clustering with explainable AI to dynamically determine the cluster boundaries for identifying the HAB regions in lakes. The study attempts to find the thresholds of multiple spectral indices, such as Normalized Difference Chlorophyll Index (NDCI), Normalized Difference Vegetation Index (NDVI), surface algal bloom index (SABI), algal bloom detection index (ABDI), and adjusted floating algae index (AFAI) for high algal blooms. The spectral indices from four North Dakota lakes were used in the clustering process and validated with ground microcystins concentration across multiple external lakes. The interpretable rule-based explanations for threshold derivation of spectral features for HAB hotspot detection in lakes help to notify the environmental agencies of the areas in lakes where sampling needs to be done. The results suggest that NDCI (0.04-0.35 with a threshold mean of 0.16) plays a major role in cyanoHAB formation in both large and shallow lakes, while NDVI (-0.28) negatively contributes to the formation of cyanoHAB. SABI <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\leq$</tex-math></inline-formula> 0.82 causes a positive impact to high microcystin formation in North Dakota lakes.