Sensor data analytics and web applications to improve monitoring and understanding of lake processes

Lakes are complex systems that involve numerous physical, chemical and biological processes. With modern sensor technology, large amounts of sensor data on lake water chemistry are being generated to help researchers understand the spatial and temporal patterns of these lake processes. Each sensor generates different datasets and effectively utilizing the resulting large and diverse datasets to improve understanding of lake processes and optimize sampling strategies is essential to protect and improve lake resources. For example, in the Great Lakes, the case study in this thesis, the US Environmental Protection Agency (USEPA) conducts several monitoring programs with various sensors, including the TRIAXUS undulating vehicle, the Sea-Bird CTD (Conductivity, Temperature, Depth) depth profiler, and a dissolved oxygen (DO) logger network that are the focus of this study. In this work, we develop three data analysis frameworks to support limnologists in more effectively collecting and analyzing these types of datasets, providing a lake system perspective. The frameworks have been made available to the research community as open-source code, including three prototype interactive Web applications. For towed undulating vehicles such as TRIAXUS, we propose a geospatial analysis framework and software to interpret water-quality sampling data in near-real time. The framework includes data quality assurance and quality control processes, automated kriging interpolation along undulating paths, and local hotspot and cluster analyses. The approach is demonstrated using historical sampling data from an undulating vehicle deployed at three rivermouth sites in Lake Michigan during 2011. The normalized root-mean-square error (NRMSE) of the interpolation averages approximately 10% in 3-fold cross validation. The results show that the framework can be used to track river plume dynamics and provide insights on mixing, which could be related to wind and seiche events. Next, we develop and test algorithms for rapid and consistent analysis of depth profiling data sampled from CTD profilers to identify lake stratifications and deep chlorophyll layers (DCL). We develop a segmentation method to approximate vertical temperature profiles with linear segments using Piecewise Linear Representation (PLR) algorithm, from which stratification patterns can be extracted. We also propose an automated peak detection algorithm to identify the fluorescence peak where the DCL lies. Testing the algorithms with data from the Great Lakes, we obtained similar results to human judgments from historical surveys. The algorithms are able to reveal spatial and temporal trends of the thermocline and DCL, as well as analyzing the shape of temperature and fluorescence profiles to detect unusual patterns such as a double thermocline. Finally, we develop a spatio-temporal interpolation framework that identifies the spatially varying temporal trend and estimates hourly hypoxia extent (dissolved oxygen [DO] concentration lower than 2mg/L) with estimation uncertainty. The framework is used to analyze spatio-temporal datasets of dissolved oxygen in Lake Erie, which were sampled from a logger network placed at the lake bottom in 2014, 2015, and 2016. The results show that hypoxia developed differently in these years. The locations with longest total hypoxic duration and longest continuous hypoxic duration are also different. Based on cross-validation results and DO time series patterns, some implications for optimizing logger locations are discussed