2002 Electronic Monitoring of Transient Water Quality Events in the Wood-Pawcatuck Watershed
Principal
Investigator: Dr. Saul D. Saila,
Professor Emeritus, University of RI
Trustee,
Wood-Pawcatuck Watershed Association
Denise Burgess, WPWA Program Director,
Melanie Cheeseman and Katherine Fisher, WPWA Interns
Wood-Pawcatuck Watershed Association
203 Arcadia
Road
Hope
Valley, RI 02832
Introduction:
Conventional water quality studies undertaken to date by the Wood- Pawcatuck Watershed Association (WPWA) have examined conditions on a scale of season or years to look for changes or trends. In these cases water quality parameters have been estimated at discrete intervals (usually biweekly) by individuals during the summer season. Transient water quality trends have not been effectively addressed to date by the WPWA because they are logistically difficult to sample. We define transient events as significant deviations from usual conditions which exist for relatively short periods of time, ranging from minutes to days. Conventional sampling techniques are usually unable to effectively detect this type of event. However, electronic data acquisition techniques, coupled with sensor technology, provide a suitable system for estimating properties of transient events. This report describes our initial efforts in applying and demonstrating such technology and techniques to events in this watershed area.
It should be recognized that short duration fluctuations in water quality have relatively small impacts on average conditions. However, these events can be of considerable ecological importance. Some examples include rainfall driven pH depressions and stream chemistry changes due to spills of wastes into water courses. Additional causes of perturbations can include road salting, street cleaning, construction, and other human impacts. Water quality in small streams can be especially affected. A better understanding of transient events and their potential risks associated with various practices will lead to better environmental management in the watershed area.
Electronic data acquisition using data loggers and electronic probes overcome many of the logistical problems encountered in high frequency sampling. However, available sensors may sometimes limit our observations to indirect measures of the event and its consequences. In spite of this, availability of even circumstantial evidence can provide indications of the nature and potential causes of some transient events. We provide some indication of how we can deduce such information in material which follows.
Six examples of data logger deployments and preliminary analyses of results are provided in this report. These examples show how electronically acquired data provides direct or indirect evidence of anomalous conditions or events. Although transient events seem particularly related to rainfall, the unusually dry summer of 2002 precluded any consistent analysis of rainfall related events. Other possible causes of apparent perturbations are briefly considered and suggestions for improved future electronic monitoring and data analysis are made.
Methods:
Two electronic data acquisition and storage devices in the form of Hydrolab Model 4a water quality data loggers, (which included probes for temperature, dissolved oxygen, specific conductance, and pH) were kindly loaned to WPWA by the U.S. Environmental Protection Agency for the summer season of 2002. Sites for data logger deployments were chosen in an attempt to establish whether or not significant indications of perturbations could be detected for the above mentioned parameters. In general, one data logger was positioned above and one below a section of stream which was believed to contain some form of detectable condition.
The data loggers were calibrated prior to each deployment. Calibration procedures for each probe were taught to WPWA personnel by Steven Rego of the U.S. EPA Laboratory, Narragansett Rhode Island. In small streams the data loggers were utilized in a horizontal position facing about 45 degrees downstream from the main direction of flow. The data loggers were attached to a board by means of two stainless steel clamps. A patio block weight was attached to the board with duct tape to provide a stabilizing weight. Each instrument with its board and weight was then tethered to a steel rod driven into the sediment upstream of the instrument package during the deployment period.
Deployments of the data loggers in impoundments were made with each logger in a vertical position by attaching it to a painted steel rod with high-quality duct tape. The rod was then pushed into the sediment with the sensor probes located at a depth of about 1 m below the surface of the water.
An arbitrary decision was made to set each of the two data loggers to simultaneously record for 10 days at hourly intervals. For purposes of sophisticated data analysis it is now recognized that it would have been more appropriate to record for 12 days in order to obtain at least 256 data points. This would provide a more efficient sample size for Fourier types of data analysis. After downloading the data from each data logger from the six sites it was stored in Excel file format. Detailed analyses were performed for at least one variable at each site to demonstrate the statistical process control methodology utilized herein. This data is available to anyone wishing to reexamine the original data in digital format.
A similar protocol was established for the analysis of what was perceived to be the most interesting single data series from each of the six deployments of the two data loggers. It will be explained in detail only for the first deployment in order to avoid redundancy in the description of the methods of analysis. However, a brief description of the interpretation of each deployment will be provided along with charts of the actual data in this report.
In general, transients in water quality contain a range of frequencies, but notably they have multiple high frequencies, unlike diurnal or seasonal cycles. Frequency decomposition has been a basic tool for a recognition system. Although the Fourier transformation has been historically used to perform the frequency decomposition, it should be recalled that the Fourier transform requires an assumption of stationarity and time invariance which seems unlikely with this type of time series data.
The approach used in this report for detecting transient events involved statistical process control (SPC) methodology. Traditionally, control charts are developed on the assumption that the sequence of process observations to which statistical process control techniques apply are uncorrelated. Indeed, the traditional approach requires independent, identically distributed variables. The most important consideration is independence, which is clearly violated in data of the type gathered in this study. The process of autocorrelation impacts the performance of conventional control charts by causing a dramatic increase in false alarms, which indicate a perturbation when none exists. The approach used in this study is based on examining the autocorrelation and partial autocorrelation structure of the data, fitting a time series model to the data based on the autocorrelation and partial autocorrelation structure and analyzing the residuals of the fitted model to the observed data. These residuals are first tested for independence. If the independence assumption is satisfied then an exponentially weighted moving average (EWMA) model is utilized in this report. This creates control charts that are applied to the residuals which have derived from the model fitted to the raw data. These control charts are then used to indicate observations which deviate significantly from values which are more than three standard deviations removed from the average. Other control limits are possible, but these seem to be frequently used in various applications. Obviously, it is possible to utilize other standard control charts once the independence assumption is satisfied.
Figures DC FQ1-DC FQ5 illustrate the raw data obtained from this first deployment. All subsequent charts labeled DC indicate that they are raw data charts. Also, only one reliable record for pH is displayed for most charts because the pH probe in one data logger was not operating properly during the entire summer period. It is regrettable that no formal procedure was employed for testing for instrument drift during deployment in order to improve precision. This omission was due to the fact that no device for instantaneous measurement of the recorded parameters was available to WPWA on a continuing basis during the deployments.
The temperature chart (DC FQ1) clearly illustrates a diurnal fluctuation in temperature with the Queen River at TNC (The Nature Conservancy) site lagging the Fisherville site slightly. The lag is thought to be due to higher flow at the lower station which delays the response time to temperature differences. A reduction in amplitude of the diurnal change is evident in the mid portion of these data. Figures DC FQ2 and DC FQ3 illustrate dissolved oxygen and oxygen saturation for both locations. In the last two to three days a reduction in the dissolved oxygen is evident for the Queen River station. The Fisherville data show a similar diurnal fluctuations throughout, whereas this is obscured somewhat in the Queen River location in the last few days. No obvious explanation for this is apparent at this time. Conductance values (Figure DC FQ4) clearly indicate an increase in specific conductance for both sites with the Fisherville site showing slightly higher values. The data logger with the functioning pH probe was located at the Queen River site which is downstream from the other site. The pH values of Figure DC FQ5 seem to indicate some interesting changes, and it was decided to further analyze the pH data using the control chart methodology described above. In general the pH data as illustrated did not show clear diurnal fluctuations but there seemed to be some interesting apparent anomalies. For example, the initial data and the final data seemed to deviate from the average conditions. Therefore the pH time series of this dataset was examined in a more critical manner.
Figure FQ1 is a repeat of Figure DC FQ5 which is scaled differently, and includes a few more observations. The anomalies are evident at both ends of this time series. Protocol for the other five deployments was essentially similar to that for the pH analysis which will be described here in detail. In the interest of avoiding redundancy the detail applied to this pH study will not be repeated for the other deployments. However, each following control chart was developed in a manner identical to this description. The reason for this deployment was an effort to detect runoff from Bear Swamp in the vicinity of the former Ladd School sewage treatment plant.. Since runoff requires rain and none occurred during the 10 day deployment, no runoff related inferences were possible.
The first step in the analysis is to choose a parameter for examination. In this case a somewhat arbitrary choice of pH was made because it showed some interesting changes. Tables FQ1a and FQ1b illustrate autocorrelation and partial autocorrelation for the pH data of Figure FQ1. It is clear from Table FQ1a that there is significant autocorrelation as shown by the exponential decay of the autocorrelation coefficient and its confidence interval. The partial autocorrelation of Table 1b is used to help in the derivation of a specific form of an Auto Regressive Integrated Moving Average (ARIMA ) model to be used in fitting to the observed data. The specific algorithm used for this is a seasonal version of the ARIMA model termed SARIMA. Table FQ2 describes the derived ARIMA model with some indications of the goodness of fit. Table FQ3 illustrates that there is no autocorrelation in the deviations of the fitted model to the raw pH data. This table was obtained from a test of the standard residuals from each observation. An illustration of the fit of the ARIMA model to the pH data is shown in Figure FQ2. It is evident that the fit was excellent. Finally, an exponentionally weighted moving average (EWMA)was applied to the residuals in the form of a control chart, which is illustrated in Figure FQ3. It demonstrates that some statistically significant departures occur at both the beginning and end of this time series. This seems to be an indication of some transient events. However, without more detail in the form of chemical analyses of the water it is not possible to suggest causality. In our opinion the important point is that it seems quite possible to detect transient phenomena of relatively low magnitude with the approach described herein.
Figures DC RB1-DC RB5 illustrate the raw data obtained from this deployment. These two locations were chosen because the stream flowed through a golf course just below the Route 102 location where the first data logger was deployed. The second data logger was located about 500 yards below the golf course on the Brownell property. It was thought that this data logger might respond uniquely to a rainfall event. However, no significant rain occurred during the deployment. In spite of no rain some interesting data were recorded by each data logger.
The temperature data (Figure DC RB1) clearly showed diurnal temperature fluctuations. In this case they were nearly simultaneous. It is noted that the lower Brownell location temperatures were somewhat lower than those above the golf course. It is suggested that this apparent anomaly may be due to the warming effect of a small impoundment located just above the location of the upper data logger deployment location. The dissolved oxygen and oxygen saturation values (Figures DC RB2 and DC RB3) showed a declining concentration of dissolved oxygen as a function of time. The same is true for the oxygen saturation curves. As might be expected, the dissolved oxygen and saturation values decreased with increasing temperatures. However, for some unknown reason the lower station showed some rapid drops in dissolved oxygen and percent saturation during the last part of the time series. No explanation for this was evident from available information. The specific conductance values at the upper station in Figure DC RB4 are shown as the lower line and do not indicate much change over the period of observation. However, the lower site showed some interesting fluctuations in specific conductance values at about the middle of the time series. No explanation for this was found, but it was decided to use the specific conductance data to establish whether or not the fluctuations observed were real in terms of statistical significance. The pH data (Figure DC RB5) showed an increasing trend over time with some limited evidence of a diurnal fluctuation.
The results of the Statistical Process Control approach to detecting transients in the specific conductance at the lower station are summarized in Figures RB1 and RB2 Figure RB1 illustrates the time series data which was used for the analysis, and Figure RB2 shows the results of the EWMA control chart applied to the residuals. It is evident that a significant transient was present, and this coincides with the location of the spike in the raw data. Although no causality could be assigned from these limited observations and data, it seems that a careful future study is justified based on has been observed to date.
Figures DC LP1-DC LP5 illustrate the available raw data from Locustville Pond and Brushy Brook inlet data sets. In this case the objective of the deployment was an attempt to assess possible consequences of an herbicide application used for weed control in Locustville Pond. The data loggers were deployed one day after the herbicide application. It was thought that dissolved oxygen might be reduced as a result of the decay of dead aquatic vegetation after the herbicide application. Thus a comparison of inlet conditions outside of the treated area with conditions in the treated area was made. One data logger was placed about 50 feet upstream from the actual inlet to the pond. The other was placed in the pond near the middle but closer to the shore on the south side in about one and one-half meters of water. Both instruments were mounted in a vertical position attached to steel rods which were pushed into the sediment.
A roughly 7 degree centigrade difference in water temperatures is noted between the inlet and the pond as shown in Figure DC LP1 with Locustville Pond having consistently higher temperatures. In general it has been found that impoundments significantly raise stream inlet water temperatures during the summer season. The limited data at hand suggest that Brushy Brook has suitable temperatures for salmonid fishes but that critical levels of temperature for salmonids were exceeded in the pond at all times of the observation period.
Dissolved oxygen and oxygen saturation values in Figures DC LP2 and 3 show significant oscillations in the cases of both the inlet and the pond. However, the treated site had consistently higher dissolved oxygen values than the inlet. This seemed strange because it is expected that the lower temperature of the inlet would provide higher oxygen values. The consistently lower oxygen values in the inlet suggest that some factor(s) upstream of the instrument were adversely affecting oxygen at the inlet. It is known that there are developments in this upstream area which may adversely affect the oxygen concentrations at the inlet. Again, these preliminary results suggest that further careful investigation of the reason(s) for the oxygen deficits in the inlet should be conducted.
Note also the virtual loss of dissolved oxygen on 71802, which is the date when the rod holding the data logger apparently fell over and rested on the sediment. No explanation for the reasons for this displacement of the data logger were found. However, the record clearly indicating the displacement is very apparent in the raw data. The oxygen saturation values closely follow the dissolved oxygen. The chart showing specific conductance (Figure DC LP4) also clearly indicates the portion of the record in which the inlet instrument was displaced into the sediment from the water column. The pH values shown in (FigureDC LP5) have been utilized for construction of the control chart analysis. It does not appear, as indicated by the dissolved oxygen values in the pond and the pH data, that any obvious significant transients in the data were apparent with the exception of the displacement of one of the data loggers which was located at the inlet. Figure LP1 shows actual data used in the control chart development. The fitted ARIMA model is shown in Figure LP2. The residuals applied to an EWMA control chart are shown in Figure LP3. Although three points are found outside the control limits these three in data points did not occur in sequence. Therefore, these points were not considered significant departures from normal average pH values under the conventional assumption that three consecutive data points outside the limits are necessary for a significant transient.
The "bottom line" from this deployment was that no significant reductions in dissolved oxygen, specific conductivity nor pH were detected and that it can be concluded that the herbicide use in Locustville Pond for the 10 days following the application did not create adverse conditions in the pond based on the data logger data.. However, it appears that the dissolved oxygen at the inlet was lower than expected and had been reduced by some factors which are not known at present. Further, more detailed investigation of the Brushy Brook oxygen deficit seems justified on the basis of available information.
Deployment No. 4- Canonchet Brook above and below golf course
Figures DC CB1-DC CB5 illustrate the raw data obtained from this deployment. The data loggers were deployed in Canonchet Brook about 100 feet above and below the golf course through which the stream flowed. Figure DC CB1 indicates very little difference in the temperature regimes between the two sites. There is a slight increase in the temperature over time for these data. Diurnal changes were also evident. These diurnal fluctuations in dissolved oxygen temporarily reached critical values toward the end of deployment. Critical values are interpreted as those above 24 degrees centigrade. Diurnal fluctuations in dissolved oxygen occurred both above and below the golf course. See Figures DC CB2 and 3 for graphic displays. The general decline in dissolved oxygen and degree of saturation was quite a similar for the two sites. The specific conductance values seem to be the most interesting. Figure DC CB4 illustrates the specific conductance values for both locations. Those above the golf course, which are the lower line in the figure, are quite uniform. However, the data from below the golf course indicate several spikes occurring at irregular intervals. The pH values of Figure DC CB5 show marked diurnal fluctuations without any significant trend. Later conversations with Mr. Palmer, the landowner at the upper station above the golf course suggest that the spikes in the specific conductance may have been associated with the irrigation schedule of the golf course. Specific conductance was chosen for critical examination by statistical control chart methodology. Figure CB1 illustrates the actual raw data used for the critical analysis. Figure CB2 shows the fit of the ARIMA model to the raw data. Figure CB3 shows the exponentially weighted average (EWMA) control chart applied to the residuals from the model fit. It is evident that the observed spikes are transient events in specific conductance below the golf course. Our limited observations and analyses to date suggest that these may be associated with the golf course irrigation schedule. These results clearly indicate that further more critical studies of this type of phenomenon are justified in the future.
This deployment was designed to obtain information regarding the differences in water quality parameters in Meadow Brook proper versus the pond, which is Meadowbrook Pond. Figures DC MB1-DC MB5 illustrate the available data in graphic form. The temperature data shown in Figure DC MB1 clearly provide another example of the warming effects of an impoundment of the water. In this case the difference between stream and impoundment range from about 7 to 10 degrees centigrade over the period of deployment. This larger difference in temperature over prior pond results is probably due to the slow rate of stream water flow over the period of observation. Note that the temperature values for the pond, which are the upper line in the figure, exceed the critical value of 24 degrees over virtually the entire observation period, and values as high as 30 degrees were occasionally achieved. These values are indicative of unsuitable conditions for survival of salmonid fishes, over the observation period. Dissolved oxygen values in the pond were well below saturation as shown in Figures DC MB2 and 3. Oxygen saturation values were well below the physiological requirements of salmonid fishes in Meadowbrook Pond. The values in the stream were acceptable by virtue of the cooler temperatures found there. The specific conductance values in Figure DC MB4 were quite regular for the pond but showed a strange rise and fall in the middle section of the stream data. These are illustrated as the lower curve of this figure. The specific conductance data were chosen as a variable interest of interest for further analyses. The pH data, Figure DC MB5, showed a slightly anomalous behavior at the beginning but then followed a routine diurnal configuration . The specific conductance time series useful analysis is shown in Figure MB1. See model fit to the original data as shown in Figure MB2. The residuals from this model fit were used for the construction of the control chart which is shown as Figure MB3. Significant excursions of data from the middle portion of the record is evident. Again, no causality for this anomalous transient condition was established from the available information. Further critical study is required in order to obtain adequate explanatory information.
Figures DC K1- DC K5 depict the raw data obtained from the deployments. The purpose of this deployment was to ascertain if any transients could be detected in water quality between the upstream side and the downstream side the Kenyon mill effluent discharge. Figure DC K1 shows that the water temperatures from both data logger locations were virtually identical. Roughly the same statement can be made for dissolved oxygen and oxygen saturation. Figures DC K2 and K3 indicate no marked differences in both variables at the two locations. The time series chosen for analysis was the specific conductance below Kenyon Industries because it showed more interesting variations. Only one peak of short duration showed statistically higher values than the average. However the rest of the time series stayed within the confidence limits that were imposed.
Conclusions:
1. Six examples were included to demonstrate how in situ stream and impoundment electronic data acquisition provided useful data regarding transient events which could not have been readily obtained by other methods. The major contribution that electronic data acquisition devices make is to enhance our ability to sample and detect transient events.
2. The results presented from these six examples indicate the significance of some of the events that that were observed. The specific causes and ecological consequences of these transient events need further detailed study because at present these data only give an indication of an event Further study will require not only additional effort but also funding to permit chemical analyses of water samples during periods of instrument deployments in order to make the transition from an indication of a transient event to a careful expression of causality.
3. Electronically gathered data, such as that recorded by the data loggers, can be used to determine the statistical properties of these transient events. In our case this was illustrated by use of a statistical control chart approach , after satisfying the assumptions of the control chart method. It is recognized that other approaches, such as Fourier series and wavelets may also be utilized for the analysis of these data.
Acknowledgements
Rob Adler
Steve Rego
TNC (for access to the property at School House Road, Exeter)
Peter Brownelle
Mr. Palmer
Kenyon Industries