Effect of Data Gaps on Harmful Algal Bloom Prediction Models for Inland Lakes

Harmful algal bloom (HAB) occurrences continue to increase globally in water bodies, with detrimental impacts on ecosystems and socioeconomics of surrounding populations. The production of toxins from HABs can also be harmful to animals and humans. The ability to accurately predict HAB abundance is crucial for implementing timely mitigation and intervention strategies to protect aquatic ecosystems and public health. Extensive studies are being conducted to acquire a better understanding of HAB initiation and growth. In this regard, the application of data-driven models for forecasting HAB occurrences is becoming increasingly prevalent. However, data-driven models rely heavily on data availability, and building a predictive model for water bodies with limited or no data impairs the performance and accuracy of models. This study investigated the potential impacts of data gaps on (1) reconstructing the missing data and (2) the prediction of cyanobacteria abundance over 15 inland lakes in the United States using these reconstructed sequences. The data utilized in this study cover short durations ranging from 3 to 7 years for the different study lakes depending on the availability of records. Three data imputation methods were examined to reconstruct missing values in these data sets: random forest regression (RFR), k-nearest neighbors (kNN), and multiple imputation by chained equations (MICE). A random forest model was then trained to predict cyanobacteria abundance under different data-availability scenarios. To improve predictions in data-limited settings, lakes were clustered based on environmental similarity to identify potentially transferable relationships between donor and target lakes. The ability of imputation methods for reconstructing missing values depended on the manner in which gaps existed in the data and also on the lake-specific conditions. For prediction of HAB abundance, clustering lakes by environmental similarity showed potential to enhance model transferability under certain conditions; however, the improvements were not consistent across all test cases. These findings highlight the importance of both high-quality data and careful consideration of lake characteristics when applying data-driven models to predict HABs in unmonitored or data-limited lakes.