| Title |
Performance Characteristics of Deep Learning Models for Algal Bloom Prediction in Rivers Using Transfer Learning |
| DOI |
https://doi.org/10.15681/KSWE.2025.41.5.365 |
| Keywords |
Algal bloom; Deep learning; Transfer learning; Water quality management |
| Abstract |
Acquiring sufficient high-quality data is essential for developing machine learning models. However, collecting water quality data in river environments can be both costly and time-consuming, and the availability of adequate data is not always guaranteed. Transfer learning provides a promising solution by enabling the application of pre-trained models, developed using data from different locations, to a target site. In this study, we employed two deep learning models commonly used for time series forecasting: long short-term memory (LSTM) and one-dimensional convolutional neural network (1D CNN), to predict chlorophyll-a concentrations, a key indicator of algal blooms. For the model input data, we tested various sequence lengths (Seq): 1, 3, 5, 7, and 9 for the LSTM model, and 3, 5, 7, and 9 for the 1D CNN model. The results indicated that the LSTM model utilizing transfer learning achieved the best performance at sequence lengths of 7 and 9, with a Nash?Sutcliffe efficiency (NSE) of 0.843. In contrast, the corresponding models without transfer learning yielded significantly lower NSE values of 0.349 and ?0.014, respectively. For the 1D CNN model, the highest performance using transfer learning was observed with an NSE of 0.814 at Seq 5, while the model without transfer learning had an NSE of 0.608. Although the degree of improvement varied by model type and sequence length, the results clearly demonstrate that transfer learning has the potential to enhance the performance of algal bloom predictions. |