Interpreting the Effect of Generative Adversarial Network Application on Deep Learning Model Performance for...

Area of Science:

• Environmental Science
• Artificial Intelligence
• Water Quality Management

Background:

• Algal bloom prediction is vital for water quality management.
• Data-driven models, especially deep learning, show promise but require extensive, high-quality data.
• Collecting real-world environmental data is often expensive and time-consuming.

Purpose of the Study:

• To investigate the impact of synthetic data generated by a time-series Generative Adversarial Network (GAN) on the performance of a Long Short-Term Memory (LSTM) network for algal bloom prediction.
• To compare a model trained solely on real data (LSTM_REAL) with one incorporating GAN-generated data (LSTM_GAN) via knowledge distillation.
• To assess the influence of different input sequence lengths on model performance.

Main Methods:

• Employed a time-series GAN to generate synthetic data.
• Utilized a Long Short-Term Memory (LSTM) network for time-series prediction.
• Compared two scenarios: LSTM_REAL (real data only) and LSTM_GAN (real + GAN data).
• Analyzed input sequence lengths from 3 to 18.
• Applied Shapley value analysis to quantify the importance of GAN-generated data.

Main Results:

• The LSTM_GAN model with a sequence length of 6 achieved the best performance (NSE of 0.802).
• Performance varied with sequence length; LSTM_GAN showed improvement over LSTM_REAL for lengths 3-12 but degradation for lengths 15-18.
• Shapley value analysis indicated GAN-generated data contributed 14.5%-24.3% to variable importance, demonstrating its influence on model inference.
• The overall performance impact of GAN data was modest, but its contribution to model understanding was significant.

Conclusions:

• Generative Adversarial Network (GAN) algorithms hold potential for enhancing algal bloom prediction models by providing valuable synthetic data.
• GAN-generated data can meaningfully influence the internal inference processes of deep learning models, even if overall performance gains are modest.
• Further research into optimizing GAN data integration and sequence length selection is warranted for improved water quality management through advanced AI.