Amount awarded: £385,159
Led by: Dŵr Cymru Welsh Water
Partners: Aigües Vic, Anglian Water, Bristol University, Cardiff University, United Utilities, Wessex Water
Ensuring drinking water is safe to drink requires constant monitoring and prediction of risk. This is true for the water quality risks associated with algae and cyanobacteria e.g., taste and odour compounds which are predicted to increase with frequency and intensity with climate change. Traditional algal monitoring is time consuming, resource intensive and does not provide sufficient data for predictive modelling of algal risks. Our project will use artificial intelligence (AI) to transform algal monitoring into a high-throughput, high-accuracy laboratory or field-based process for a fraction of the cost, allowing better risk prediction enabling water companies to take earlier, more cost effective and targeted actions.
“This funding will enable a significant leap forward in algal analysis and accelerate the use of algal data to predict water quality risks. This will provide benefits to customers and wider society, as it will better equip the water industry to tackle current and future algal related water quality challenges.” – Matthew Jones, Public Health Manager, Wessex Water
—
We are actively receiving monthly water samples from four distinct water companies, each contributing samples from different reservoirs. Occasionally, we received samples from two other companies. To date, we have a gathered a total of 682 samples. These samples are being utilized to construct a dataset aimed to train our model. Currently, this dataset comprises 11,937 image files, predominantly from winter samples. Since starting the project, I have produced a full draft of chapter 1 for my thesis. I have also been on a lab visit to Wessex Water, along with Sonia (the project research technician), to understand how samples are processed, how algal data is stored, and how the project produced model could be effectively utilised. Additionally, I have had meetings with Anglian Water about the model as well as discussions about notable community changes within their catchment. This led to discussions about the benefits of predictive modelling from the dataset produced by the project. I am currently testing different CNNs in order to learn more about model architectures/structures and the subsequent influences on the accuracy for image classification. This work will also go towards a review of current CNNs and applications for microalgal classification and enumeration.
Over the past quarter, I have been testing different model components with an open source algal dataset containing 6 named classes/cell types. This includes testing each model with and without pretrained weights, the amount of epochs used for training (50, 100, and 200 were compared) and four different optimizers (used to accelerate model training). The results of optimizer testing varied greatly depending on the individual model being tested, as well as varying depending on weights (e.g. with or without pretrained weights). With these tests some models showed classification accuracies of 95%. For the next quarter, I will be testing another open source dataset with 88 classes to compare previous test results as well as testing project produced data. Then data augmentation techniques will be tested and the accuracy will then be compared to results where these techniques were not applied. (Holly Liken, Bristol University)
Automated identification and counting of algal cells from raw water. Credit: Dr. Chris Williamson, University of Bristol
