Improved predictive performance of cyanobacterial blooms using a hybrid statistical and deep-learning method

Cyanobacterial harmful algal blooms (CyanoHABs) threaten ecosystem functioning and human health at both regional and global levels, and this threat is likely to become more frequent and severe under climate change. Predictive information can help local water managers to alleviate or manage the adverse effects posed by CyanoHABs. Previous works have led to various approaches for predicting cyanobacteria abundance by feeding various environmental variables into statistical models or neural networks. However, these models alone may have limited predictive performance owing to their inability to capture extreme situations. In this paper, we consider the possibility of a hybrid approach that leverages the merits of these methods by integrating a statistical model with a deep-learning model. In particular, the autoregressive integrated moving average (ARIMA) and long short-term memory (LSTM) were used in tandem to better capture temporal patterns of highly dynamic observations. Results show that the proposed ARIMA-LSTM model exhibited the promising potential to outperform the state-of-the-art baseline models for CyanoHAB prediction in highly variable time-series observations, characterized by nonstationarity and imbalance. The predictive error of the mean absolute error and root mean square error, compared with the best baseline model, were largely reduced by 12.4% and 15.5%, respectively. This study demonstrates the potential for the hybrid model to assist in cyanobacterial risk assessment and management, especially in shallow and eutrophic waters.

Automated summary

The study considers combining statistical models with deep-learning models to better capture temporal patterns of highly dynamic observations. The proposed ARIMA-LSTM model shows promising potential to outperform current baseline models for predicting CyanoHABs in highly variable time-series observations. The predictive errors of mean absolute error and root mean square error were reduced significantly compared to the best baseline model.