Development of a Deep Learning Model for Harmful Algal Blooms Prediction

The risk of harmful algal blooms (HABs) is exacerbated by extreme climate and hydrologic events, as well as the increased non-point pollutant sources associated with agriculture and industrialization. The resulting deterioration in water quality due to the HABs poses significant threats to water management and aquatic ecosystems. HABs, in particular, emerge from intricate chemical interactions influenced by external conditions and diverse hydrologic and water quality factors.  Existing physical models encounter difficulties in predicting HAB occurrences and concentrations due to their limitations in addressing the intricate interactions of external environments and the characteristics of non-linear, non-stationary systems. In response to this challenge, we aim to develop a deep-learning algorithm based on the wavelet transform, with a focus on key hydrologic and water quality factors specific to the Nakdong River in South Korea. We identify water temperature and Chlorophyll-a as pivotal factors influencing HABs. Leveraging the wavelet transform, we extract denoised HAB data to enhance the robustness of our predictive model. Subsequently, we employ Long Short-Term Memory (LSTM) networks to construct a deep learning model, utilizing the identified key factors and denoised data as input features. Our preliminary results demonstrate a decent level of predictive accuracy showing a high Nash-Sutcliffe Efficiency (NSE) value of 0.88 and a low Root Mean Squared Error (RMSE) of approximately 9800 cells/ml, compared to the average HAB quantity of 14474 cells/ml. These outcomes indicate that the developed deep learning approach allows for accurate simulation of HAB. The implications of our research extend to the precise analysis of HABs, enabling the establishment of pre-emptive responses for effective water resources management.

 

Acknowledgment:

This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. NRF-2022R1A4A3032838).