Using sensors, satellites and artificial intelligence to improve catchment-scale modelling and nowcasting of microbial risks to water quality

Overview

Background. Scotland’s rivers, lakes and public bathing waters provide substantial benefits to local economies and the broader wellbeing of its citizens, but these benefits can be significantly impaired where episodes of poor microbial water quality impose restrictions on recreational use of water bodies. In Scotland, and elsewhere in the UK, significant improvements in water quality have been achieved since the introduction of the EU Water Framework and Bathing Water Directives. Nevertheless, microbial pollution of water bodies is still an issue in many areas due to runoff from urban and agricultural catchments and discharges from sewer outflows during periods of intense rainfall. While these events are generally short-lived, they can pose significant risks to public health and are predicted to occur more often in the future as the frequency of high rainfall events during summer increases under climate change. Hence there is an urgent need to improve our ability not only to monitor but also to forecast the occurrence of pollution events. The Scottish Environment Protection Agency (SEPA), for example, has developed a two-pronged strategy to mitigate rainfall-driven pollution events focusing on: (a) reducing land-water transfer of faecal indicator organisms (FIOs) at the catchment scale; and (b) the provision of daily ‘nowcasts’ for bathing waters to advise the public against bathing during periods of poor water quality. These approaches rely heavily on the use of predictive models, supported by regular water quality sampling and analysis. In regard to (a), models such as the Sensitive Catchment Integrated Mapping Analysis Platform (SCIMAP) provide a framework for risk-based modelling of diffuse pollution at the catchment-scale enabling the identification of critical source areas as a means to inform more efficient and targeted management interventions In regard to (b), as it is infeasible to undertake compliance sampling on a daily basis across a large number of sites, models have been developed to facilitate real-time (RT) predictions of water quality at sites of interest. SEPA currently use these models to provide nowcasts at a number of bathing waters throughout Scotland with the information relayed to the public through on-site electronic signage.

In spite of significant advances in the use of models for predicting microbial risks to water quality over recent years, current approaches are often limited by incomplete representations of the key processes influencing the fate of FIOs in the environment and/or by the lack of appropriate data for model parameterisation, calibration and validation. For example, while the RT models currently used in Scotland provide satisfactory predictions of rainfall-driven pollution events in some catchments, they perform poorly in more complex settings where the interacting effects of other factors such as land cover, meteorology (e.g., temperature, ultraviolet radiation), physico-chemical water quality (e.g., turbidity, salinity) and tidal state also influence FIO dynamics. Furthermore, the models are currently calibrated using historical compliance data generated from weekly sampling, but it is now recognised that FIOs can exhibit pronounced within-day variability sufficient to influence bathing water classifications. In this context, the recent emergence of novel coliform sensors has the potential to transform early-warning monitoring of microbial water quality not only through the direct delivery of RT information but also by improving the provision of data for assimilation in predictive models. In addition, the increased availability of open and free satellite data (e.g. Copernicus), which can be used to derive a suite of catchment-relevant variables, has the potential to improve spatial representations of these parameters within modelling frameworks. These new data streams not only offer opportunities to improve existing model parameterisations, but also to enable the evaluation of more innovative, data-driven, modelling approaches based on artificial intelligence that have very recently been shown to improve short-term predictions of microbial pollution elsewhere.

Methodology

The project will focus primarily on the Firth of Forth, its sub-catchments and associated bathing waters capitalising on recent investments in state-of-the-art sensor and satellite technology through the Forth-ERA programme (see below). The project will be approached through three discrete, but closely interlinked, research phases (RP) each one mapped against a key project objective.

RP1: In-field evaluation of fluorescence-based coliform sensors (months 1-12). The utility of a new tryptophan-like fluorescence sensor (Proteus instruments) for RT monitoring of E. coli and other coliforms will be evaluated through targeted sampling in the River Leven and River Almond catchments. The sensor will be installed near Levenmouth in November 2021 as part of Forth-ERA and will provide RT data via LoRaWAN-enabled telemetry with a further sensor due to be installed in the River Almond in 2022. This will be one of the first studies to evaluate this new sensor technology as part of a long-term field deployment. Sensor-derived estimates of E. coli concentration will be compared to concurrent measurements obtained via standard membrane filtration techniques. The sampling strategy will combine high frequency (~30 minute) sampling of microbial water quality over the course of several days (repeated seasonally) with regular weekly samples obtained during the bathing water season (01 June to 15 September) to capture both within-day and seasonal variability in water quality providing a robust dataset for sensor calibration and validation. Data from additional sensors simultaneously measuring other water quality parameters (i.e., temperature, salinity, dissolved oxygen, turbidity, chlorophyll, fluorescent dissolved organic matter, biological oxygen demand) will be used to assess potential interferences and sources of error in sensor-derived E. coli estimates.

RP2: Using sensors and satellites to improve risk-based models of microbial pollution (months 12-24). Data provided by satellite observations and IoT sensors and will be used to develop new approaches to the parameterisation of the SCIMAP-FIO model for catchment-scale assessment of FIO risks. One of the main factors limiting the accuracy of this and similar modelling frameworks is a reliance on relatively static data sources describing key input variables such as land cover, hydrology and livestock densities[3]. Here we will exploit open and free satellite data from Copernicus, including annual land cover products from Proba-V and high frequency surface soil moisture data from Sentinel-1, to enable more dynamic, spatially-explicit modelling of microbial pollution risks at the catchment scale. Additionally, we will examine the potential to derive estimates of livestock densities from very high-resolution satellite data (an application being developed in tandem by our Forth-ERA partners 3DEO). Model calibration will be supported by standard FIOs assays and data from the Proteus coliform sensor.

RP3: Improving real-time predictions of bathing water quality using data-driven methods (months 24-36). The project will evaluate how data streams from IoT sensor networks providing information on meteorology, river flow, water quality and tidal conditions can be used in conjunction with other data sources (e.g. numerical weather prediction, satellite observations) to improve real-time predictions of bathing water quality techniques from statistical and machine learning (e.g., feature engineering, support vector machines, random forests). Improvements will include greater accuracy and resolution (15 min vs daily), and quantification of forecast uncertainty, which will enable risk-based decision-making. The models will be initially developed and piloted for the River Leven catchment using historical compliance data to support model development and validation with further comparisons made against sensor-derived estimates of E. coli. The model will be developed and tested by hindcasting, but the framework will be readily adaptable to enable ingestion of live data streams. The approach will also potentially allow comparisons to be made against the existing models used by SEPA. Finally, the model will be extended to a limited number of other bathing waters within the Firth of Forth (simplifying its parameterization as necessary) to test the transferability and scalability of the modelling framework.

Project Timeline

Year 1

Development of literature review and project proposal
Training in laboratory and field sampling techniques
Collection and analysis of sensor data; comparison against data from field and compliance sampling
Drafting of first paper on in-field evaluation of fluorescence-based coliform sensors

Year 2

Review of progress in Year 1
Training in SCIMAP modelling and Earth observation data
Refinement of SCIMAP model using sensor and satellite data streams
Modelling of selected sub-catchments in the Firth of Forth
Drafting of second paper on improved risk-based models of microbial pollution

Year 3

Review of progress in Year 2
Training in statistical and machine learning approaches to forecasting
Development, calibration and validation of models through hindcasting
Evaluation of models against new, real-time data streams
Drafting of third paper on forecasting microbial water quality

Year 3.5

Draft additional thesis chapters
Submit thesis for examination
Develop career development plan
Undertake viva examination

Training
& Skills

In addition to the training provided through the IAPETUS DTP, the student will receive in-house and external training in all key components of the project including:
• Deployment, operation and maintenance of sensors
• Scientific programming in Python and R
• Field sampling and laboratory analysis of water samples
• Risk-based modelling of catchments using SCIMAP
• Statistical and machine learning approaches to forecasting

References & further reading

Richardson J., Feuchtmayr H., Miller C., Hunter P.D., Maberly S.C., and Carvalho L. (2019). Response of cyanobacteria to warming, extreme rainfall events and nutrient enrichment. Global Change Biology. 25, 3365-3380.

Quilliam R.S., Taylor J., and Oliver D.M. (2019). The disparity between regulatory measurements of E. coli in public bathing waters and the public expectation of bathing water quality. Journal of Environmental Management 232, 868-874

Porter K.D.H., Reaney S.M., Quilliam R.S., Burgess C. and Oliver D.M. (2017). Predicting diffuse microbial pollution risk across catchments: the performance of SCIMAP and recommendations for future development, Science of the Total Environment, 609, 456-465

Oliver D.M., Porter K.D.H., Pachepsky Y.A., Muirhead R.W., Reaney S.M., Coffey R., Kay D., Milledge D.G., Hong E., Anthony S.G., Page T., Bloodworth J.W., Mellander P-E., Carbonneau P.E., McGrane S.J. and Quilliam R.S. (2016). Predicting microbial water quality with models: over-arching questions for managing risk in agricultural catchments, Science of the Total Environment, 544, 39-47

Further Information

Apply Now