Discipline: Mathematics and Statistics
Subcategory: Water
Cassia Smith - University of the Virgin Islands
Dissolved oxygen (DO) is necessary to many forms of life in aquatic ecosystems as most living organisms require oxygen for their basic metabolic processes. Healthy marine environments maintain a delicate balance between anoxia & hypoxia; DO concentration is one of the main indicators in assessing water quality. As such, the ability to forecast oxygen (O2) levels will be invaluable in monitoring the health of local marine areas and analyzing the effect of bio phenomenon and human interference. This project seeks to construct an accurate model for predicting DO concentrations in a lagoon area within Brewers Bay, St. Thomas VI and to explore what may influence DO fluctuations. A machine learning approach via MATLAB software will enable the development of several regression models (neural networks, decision trees, generalized linear models, linear regression). Accordingly, a miniDOT logger was installed in the lagoon at a depth of eight feet. The sensor collects hourly data on water temperature, day/time category, DO, and O2 saturation. Data is also obtained from the St. Thomas airport database for the following categories: lunar phase/ illumination, air temperature, dew point, humidity, sea level pressure, visibility, wind direction/speed, precipitation, and conditions. The collected raw data is then cleaned and inputted into an analyzed dataset comprised of hourly averages. The analyzed dataset is bifurcated into ‘training’ and ‘test’ sets to begin creation of the model. Model variable inputs are air temp, humidity, condition, water temp, precipitation, sea level pressure, and wind speed. The output variable is DO. The adequacy of the model is reviewed through performance and fit tests. The model is then evaluated for accuracy by comparing its predictions to data not used for calibrating the model.
Though research is still ongoing, preliminary results indicate an inverse relationship between DO levels and temperature. There also appears to be a strong day/night cycling of DO, with levels of O2 higher in the day than in the night. This is likely due to photosynthetic activity during the day and higher respiration at night. It has been possible to develop a meaningful neural network model from the dataset that has a Mean Squared Error of 46.9 and a Regression Value of 0.967. Future work will focus on refining the neural network model and constructing the other model types. Once a highly accurate model is developed, we can apply the working model to other regions in Brewer’s Bay in order to contribute to a better understanding of the effects of natural processes and human factors.
References: Bishop, C. M. (2013). Model-based machine learning.’Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences,’ 371(1984)
Palani, S., Liong, S. Y., Tkalich, P., & Palanichamy, J. (2009). Development of a neural network model for dissolved oxygen in seawater. ‘Indian Journal of Marine Sciences,’38(2), 151.
Funder Acknowledgement(s): Funding provided by NSF/HBCU-UP grant #1137472.
Faculty Advisor: Robert Stolz and Jonathan Jossart, rstolz@uvi.edu
Role: I developed the prediction models through the usage of MATLAB software for machine learning and statistics.