Summer Stream Water Temperature Models for Great Lakes Streams: New York

Abstract

Temperature is one of the most important environmental influences on aquatic organisms. It is a primary driver of physiological rates and many abiotic processes. However, despite extensive research and measurements, synoptic estimates of water temperature are not available for most regions, limiting our ability to make systemwide and large‐scale assessments of aquatic resources or estimates of aquatic species abundance and biodiversity. We used subwatershed averaging of point temperature measurements and associated multiscale landscape habitat conditions from over 3,300 lotic sites throughout New York State to develop and train artificial neural network models. Separate models predicting water temperature (in cold, cool, and warm temperature classes) within small catchment–stream order groups were developed for four modeling units, which together encompassed the entire state. Water temperature predictions were then made for each stream segment in the state. All models explained more than 90% of data variation. Elevation, riparian forest cover, landscape slope, and growing degree‐days were among the most important model predictors of water temperature classes. Geological influences varied among regions. Predicted temperature distributions within stream networks displayed patterns of generally increasing temperature downstream but were patchy due to the averaging of water temperatures within stream size‐classes of small drainages. Models predicted coldwater streams to be most numerous and warmwater streams to be generally associated with the largest rivers and relatively flat agricultural areas and urban areas. Model predictions provide a complete, georeferenced map of summer daytime mean stream temperature potential throughout New York State that can be used for planning and assessment at spatial scales from the stream segment class to the entire state.

Comparing various artificial neural network types for water temperature prediction in rivers

“…Streamwater temperature prediction approaches proposed in the past mainly included physically-based, temperature equilibrium concept-based or simple statistical models (Webb et al, 2008;Wehrly et al, 2009;Bustillo et al, 2014). In recent years various kinds of deterministic models (Caissie et al, 2007), data-driven approaches (St-Hilaire et al, 2012;Grbic et al, 2013;Cole et al, 2014) or artificial neural networks (ANNs) (Sahoo et al, 2006;Sivri et al, 2007;Chenard and Caissie, 2008;Sahoo et al, 2009;Daigle et al, 2009;Faruk, 2010;McKenna et al, 2010;Jeong et al, 2013;Napiorkowski et al, 2014;Piotrowski et al, 2014;Hadzima-Nyarko et al, 2014;Rabi et al, in press) have been applied to this task. In some studies (Sahoo et al, 2009;Bustillo et al, 2014) regression and ANN models are claimed to perform not worse than the more sophisticated empirical or heat budget-based models.…”

“…For example Sahoo et al (2006) compared regression, chaotic and multi-layer perceptron ANN models, Wehrly et al (2009) compared various statistical models, Cole et al (2014) compared three data-driven approaches with heat flux model and Bustillo et al (2014) verified the performance of various regression and temperature equilibrium-based models in the context of streamwater temperature prediction. However, although a large number of different types of neural networks have been developed so far, for the prediction of streamwater temperatures almost always the ''classical'' multi-layer perceptron ANNs (MLP) have been used (Sahoo et al, 2006;Sivri et al, 2007;Chenard and Caissie, 2008;Daigle et al, 2009;McKenna et al, 2010;Jeong et al, 2013;Piotrowski et al, 2014;Hadzima-Nyarko et al, 2014;Cole et al, 2014;Rabi et al, in press). Similar MLP networks were also applied for somehow related problem, the prediction of temperatures of stormwater runoff in urban watershed (He et al, 2011;Sabouri et al, 2013).…”

Thermal and hydrologic suitability of Lake Erie and its major tributaries for spawning of Asian carps

“…Using predicted velocity during flood events and temperatures, all or nearly all flood events in the Maumee, Huron, and Grand Rivers were high-quality events, i.e., those for which the length of passable river was greater than the length of river required for hatching of eggs given temperature and velocity. All flood events following attainment of thermal thresholds in all rivers in all years occurred during June-September, which is the period for which the McKenna et al (2010) model predicts temperature. The Maumee, Sandusky, and Grand Rivers all had high quality events in at least half of the years examined.…”

“…From these dates we calculated duration in consecutive days that velocity exceeded 0.7 m/s. We used a simple backpropagation neural network temperature model with one hidden layer of neurons (see McKenna, 2005 for description of neural networks) for Ohio streams developed with the general methods of McKenna et al (2010) for New York streams to estimate summer (June-September) river temperatures. The model predicted four broad classes of mean summer daytime water temperature (b18°C, 18-b21°C, 21-b24°C, and ≥24°C) based on 18 stream and landscape variables (those used by McKenna et al, 2010, plus Shreve link number;Shreve, 1967) for 40 to 125 reaches of each river.…”

Predicting Thermally Stressful Events in Rivers with a Strategy to Evaluate Management Alternatives

“…McKenna et al . () used these classification schemes to build ANN models to predict temperature classes at 16 °C (cold water), 19 °C (cool‐transition), 22 °C (warm transition), and 24 °C (warm) for some 52 000 stream segments across the US State of New York. Other modelling efforts have focused on evaluating the relationship between land use and climate change and river temperatures (e.g.…”