Deep learning for the early detection of harmful algal blooms and improving water quality monitoring

Dagtekin, Onatkut

Abstract

Climate change will affect how water sources are managed and monitored. The frequency of algal blooms will increase with climate change as it presents favourable conditions for the reproduction of phytoplankton. During monitoring, possible sensory failures in monitoring systems result in partially filled data which may affect critical systems. Therefore, imputation becomes necessary to decrease error and increase data quality. This work investigates two issues in water quality data analysis: improving data quality and anomaly detection. It consists of three main topics: data imputation, early algal bloom detection using in-situ data and early algal bloom detection using multiple modalities.
The data imputation problem is addressed by experimenting with various methods with a water quality dataset that includes four locations around the North Sea and the Irish Sea with different characteristics and high miss rates, testing model generalisability. A novel neural network architecture with self-attention is proposed in which imputation is done in a single pass, reducing execution time. The self-attention components increase the interpretability of the imputation process at each stage of the network, providing knowledge to domain experts.
After data curation, algal activity is predicted using transformer networks, between 1 to 7 days ahead, and the importance of the input with regard to the output of the prediction model is explained using SHAP, aiming to explain model behaviour to domain experts which is overlooked in previous approaches. The prediction model improves bloom detection performance by 5% on average and the explanation summarizes the complex structure of the model to input-output relationships.
Performance improvements on the initial unimodal bloom detection model are made by incorporating multiple modalities into the detection process which were only used for validation purposes previously. The problem of missing data is also tackled by using coordinated representations, replacing low quality in-situ data with satellite data and vice versa, instead of imputation which may result in biased results.

Citation

Dagtekin, O. Deep learning for the early detection of harmful algal blooms and improving water quality monitoring. (Thesis). University of Hull. https://hull-repository.worktribe.com/output/4220252

• Thesis Type: Thesis
• Deposit Date: Feb 23, 2023
• Publicly Available Date: Feb 23, 2023
• Keywords: Computer Science
• Public URL: https://hull-repository.worktribe.com/output/4220252
• Award Date: Sep 1, 2022