Previous Studies

Scientists have estimated that the northeastern United States will experience the largest increases in temperatures in the contiguous United States. These temperature increases are predicted to occur by 2035, as much as two decades before global average temperatures should achieve similar increases (Reidmiller and others, 2018). Thus, the prediction of hydrologic vulnerability caused by future climate variation has increasingly become a major focus of research in the northeastern United States.

Although many climatologic and hydrologic studies have been done to date, two previous studies are foundational to this investigation. The first study is the development of the SWG for Massachusetts. The SWG was developed at Cornell University (Steinschneider and others, 2019; Steinschneider and Najibi, 2022) and was designed to generate synthetic weather data while allowing perturbations related to climate change. The second study foundational to our investigation is the development of the SWM (Shabestanipour and others, 2023). The SWM was designed to address systematic biases of extreme events often exhibited by deterministic watershed models.

Purpose and Scope

The purpose of this study is to develop tools for projecting 21st-century hydrologic characteristics in Massachusetts that give more-accurate estimates of future streamflows at the high- and low-flow extremes. This report presents the data, methods, and results of a cooperative study of the Squannacook River watershed (fig. 1) of Massachusetts to build a SWM that uses output from a calibrated deterministic model. The calibrated deterministic model simulates future hydrology on the basis of projected climate estimates from a SWG. Multiple future scenarios of rising temperatures are simulated with the models, and the percent change in streamflow characteristics are provided. All data used for this study are available in an associated USGS data release (Olson and others, 2024).

The Squannacook River Watershed Study Area

For this study, a watershed was needed that met several criteria related to data availability and physical setting. First, the watershed needed a streamgage with a long-term continuous streamflow record. A long-term record allows for a means to calibrate a deterministic model and provides enough data for the SWM to correct for bias. Second, the watershed needed to be mostly free of substantial hydrologic manipulations, such as regulation and flood control. Although this criterion was not strictly adhered to, the intent was to avoid complexities in developing the pilot application. Third, the watershed needed to have basin characteristics typical of watersheds in Massachusetts. The intent of this third criterion was to select a watershed that would streamline the transfer of the methods developed in this study to other watersheds in Massachusetts. To meet the third criterion, the selection of watersheds was limited to those with hydrology and climate that were not substantially affected by urban, coastal, or orographic effects. The last criterion, the location of a streamgage in the watershed, where streamflow was determined, needed to have a corresponding computational node in the National Hydrologic Model (NHM; Viger and Bock, 2014). Meeting this last criterion allowed the extraction and use of a local watershed model from the NHM without significant model adjustment or recalibration.

From the 28 Massachusetts watersheds determined to be potential study areas, the Squannacook River watershed (fig. 1) was selected. The streamgage within the watershed (Squannacook River Near West Groton, Massachusetts [USGS streamgage 01096000; USGS, 2023]) has 74 years of continuous streamflow record from October 1949 to date (2023). The basin and climatic characteristics of the Squannacook River watershed (table 1) are typical of Massachusetts, and the streamgage location has a corresponding location in the NHM where streamflow was calibrated and computed. Some anthropogenic effects exist in the form of groundwater withdrawals, occasional regulation at low streamflows, and streamflow diversion from 2.2 square miles (mi²) of drainage by the Fitchburg Reservoir (USGS, 2023), but these effects were considered insignificant relative to anthropogenic effects of many watersheds with streamgages in Massachusetts.

The Squannacook River watershed is in north-central Massachusetts (fig. 1) and crosses the State border into New Hampshire. The watershed is primarily forested and contains more than 7,000 acres of State- and town-owned forest lands. Areas of development are along transportation corridors and in the center of Townsend, Mass. The Squannacook River originates from the confluence of four brooks draining the western end of the watershed: Mason Brook, Walker Brook, Locke Brook, and Willard Brook. The brooks drain hilly, forested lands with elevations as high as 1,500 feet (ft). The elevation at the confluence of the brooks is about 300 ft. Downstream from the confluence, the Squannacook River flows southeasterly along a winding, 11-mile path to the location of the USGS streamgage (01096000; fig. 1). The drainage area and elevation of the streamgage are 65.9 mi² (Falcone, 2011) and 250 ft (USGS, undated a), respectively. Past the streamgage, the river continues another winding 6 miles where it flows into the Nashua River. At the mouth of the Squannacook River, the drainage area and elevation are 72.5 mi² (USGS, undated b) and 200 ft (USGS, undated a), respectively.

In the headwaters of the Squannacook River watershed, at the upstream end of Willard Brook, are two reservoirs: Fitchburg Reservoir and Ashby Reservoir. These reservoirs are owned and operated by the City of Fitchburg (fig. 1). Fitchburg Reservoir is on the Squannacook River watershed boundary and was constructed in 1915 as an emergency water supply for the City of Fitchburg. The reservoir covers about 155 acres and has a north spillway and gated outlet that drains to Willard Brook and the Squannacook River watershed; the reservoir has a south spillway and gated outlet that drains to a neighboring watershed and other reservoirs supplying water to Fitchburg. The south spillway, as reconstructed in 1963, is 0.56 ft higher than the north spillway and rarely overflows. The gate valve on the north outlet leading to Willard Brook has malfunctioned and is stuck in a partially open position. The south reservoir outlet has two gate valves that are partially functional and are used to divert additional water as needed for water supply to Fitchburg. The streamflows are not metered, and no formal records are maintained for the reservoir (City of Fitchburg, 2013b).

Ashby Reservoir is entirely within the Squannacook River watershed and is downstream from Fitchburg Reservoir. It covers about 39 acres and was constructed in 1917 as a compensatory water supply for mill owners downstream. The reservoir now serves as a site for recreation and is not used to regulate streamflow. No formal operating or streamflow records are kept for the dam at the outlet of Ashby Reservoir (City of Fitchburg, 2013a).

The climate in the Squannacook River watershed is temperate (mild summers and cold winters). The mean annual air temperature during 1981–2010 was about 7.8 °C and mean monthly air temperatures ranged from about −5.6 °C in January to 20.6 °C in July (National Oceanic and Atmospheric Administration, 2021a). For the same period, the mean annual precipitation was 48 inches (in.), and the mean annual potential evapotranspiration for the area was 23 in. (National Oceanic and Atmospheric Administration, 2021a, b25).

The Stochastic Weather Generator

The SWG was developed for Massachusetts to estimate the future climatologic data (daily precipitation and air temperature) needed to run the deterministic model. The SWG was developed at Cornell University (Steinschneider and others, 2019; Steinschneider and Najibi, 2022) and was designed to generate synthetic weather data while allowing perturbations related to climate change. The SWG synthesizes weather data using a resampling algorithm on observed weather records within selected weather regimes (the weather regimes selected reflect recurring large-scale, persistent atmospheric-flow patterns). The SWG can synthesize future scenarios of weather data by (1) perturbing the frequency of weather regimes, (2) adjusting air temperature by simple addition, and (3) adjusting precipitation using a Clausius-Clapeyron scaling method (Held and Soden, 2006) that relates the effects of warming temperatures on precipitation because of increases in the moisture holding capacity of the atmosphere.

The observed weather records used as the basis for the SWG was the Livneh and others (2015) climate dataset, referred to in this report as the “Livneh dataset”, gridded to 1/16-degree resolution. The period of record used from the Livneh dataset was 1950 to 2013. This study used 17 warming scenarios, covering 0 °C to 8 °C, in 0.5-degree increments. For each scenario, the SWG generated an ensemble of 100 members. Each ensemble member contained a 64-year record of daily precipitation, minimum air temperature, and maximum air temperature gridded at the same resolution as the Livneh dataset covering Massachusetts. These data were input to the PRMS deterministic model.

The Deterministic Model

The deterministic watershed model used for this investigation was the USGS PRMS version 5.1.0 (Leavesley and others, 1983; Markstrom and others, 2015). The PRMS is a distributed-parameter, physically based watershed-modeling system. Responses to climate and land-cover changes are simulated in terms of water and energy balances, streamflow regimes, flood peaks and volumes, soil-water relations, and groundwater recharge. In PRMS, the components of streamflow include contributions from surface runoff, subsurface (soils) interflow, and groundwater. The watershed’s water budget consists of storage in snowpack; soil moisture; groundwater; inputs from precipitation and snowmelt; losses to evapotranspiration; recharge to the deeper aquifer system; and outflows to streams from surface, subsurface, and shallow-groundwater reservoirs.

The PRMS is a modular modeling system, whereby modules that represent watershed-process algorithms, selected from a library of subroutine modules, are combined into a customized model. The customized PRMS can then simulate components of a particular hydrologic system, including water, energy processes, and stream temperature. The PRMS operates on a fixed daily time step and is driven by daily inputs of total precipitation and maximum and minimum temperatures. The data release for the deterministic model (Olson and others, 2024) includes a list of modules used in this configuration. Supplemental information regarding PRMS modules and source code specific to the PRMS configuration presented in this report is available in the PRMS users’ manual (Markstrom and others, 2015) and the USGS PRMS developers’ website (https://www.usgs.gov/software/precipitation-runoff-modeling-system-prms).

The PRMS model for the Squannacook River watershed was extracted from the NHM (Regan and others, 2018). The NHM is a model platform that uses the PRMS hydrologic simulation code and was developed as a comprehensive, consistent hydrologic model for the contiguous United States. The NHM uses parameters derived from characteristics of topography, land cover, soils, geology, and hydrography; these characteristics are obtained using traditional geographic information system (GIS) methods. Some parameters are set to long-established default values, whereas others are determined through calibration. The Daymet climate dataset (Thornton and others, 2016) was used to configure and calibrate the NHM (Hay and LaFontaine, 2020).

In a PRMS model, and thus in the NHM, a modeled region is divided into polygon-shaped subwatersheds called hydrologic response units (HRUs). The daily water balance is simulated in each HRU on the basis of precipitation and temperature input data. The HRUs in the NHM have a wide range of sizes and have boundaries that coincide with land use or catchments that drain to confluences. The mean area of HRUs in the NHM is 29 mi². The total runoff simulated by the model is aggregated from the HRUs draining to and routed through stream segments that represent the stream-channel network. Each stream segment represents a channel reach bounded by an HRU or, in most cases, bounded on each side of the channel by an HRU. Each channel, except for the most upstream segment, is fed by a segment upstream from it and by the bounding HRUs. The accumulated discharge in the segment is routed to the next downstream segment by a Muskingum routing scheme (Markstrom and others, 2015).

The PRMS model extracted from the NHM for the Squannacook River watershed upstream from the mouth of the river consisted of three HRU catchments and two stream segments. Only one HRU and one stream segment were upstream from the location of interest for this study, streamgage 01096000 (fig. 2). Two HRUs and one stream segment were downstream from the streamgage location to the mouth of the river. The upstream HRU and the stream segment corresponding to the streamgage location was used for additional model calibration, and the streamflow computed at this location was used to develop the SWM.

Evaluation of the Squannacook River Watershed PRMS Model Performance

To evaluate model performance, the Daymet climate data—used in the NHM calibration—was input to the PRMS model. The model output was compared with daily streamflow data from streamgage 01096000 (USGS, 2023). Although the volume of water diverted from the Squannacook River watershed at the Fitchburg Reservoir is unknown, it is considered to have an insignificant effect on the streamflow records because the reservoir is in the headwaters of the watershed and its primary outlet drains into the Squannacook River watershed. The Fitchburg Reservoir has a drainage area of 2.2 mi² or only about 3.4 percent of the gaged drainage area. Hence for this study, the observed streamflows at the streamgage were considered unaffected by storage, regulation, and diversion.

The concurrent Daymet and streamflow record data used to evaluate model performance extended from January 1, 1981, to December 31, 2013. The 33-year period provides time-series data of sufficient length to yield mean values and variability representative enough to ensure that inferential statistics would be meaningful. Additional evaluation of model performance was done using the Livneh dataset and concurrent streamflow records from January 1, 1981, to December 31, 2013.

The model’s performance was evaluated by matching general water-balance information and by goodness-of-fit statistics for simulated streamflow compared with USGS measured streamflow. The goodness-of-fit statistics included the Nash-Sutcliffe Efficiency (NSE) statistic (Nash and Sutcliffe, 1970). Following Moriasi and others (2007), an NSE of 0.5 or greater is considered a satisfactory or good fit between the simulated and measured hydrographs, and an NSE of 0.4 to 0.5 is considered marginally satisfactory. Values of NSE less than 0.4 are considered unsatisfactory. Additional goodness-of-fit statistics included the normalized root-mean-square error (NRMSE; normalized on the difference between the maximum and minimum values in the record), and the percent bias (PBIAS). A comparison of streamflow-quantiles was also used for evaluating model performance.

The PRMS-simulated long-term water balance (greater than 30 years) for the watershed upstream from the streamgage had values for annual precipitation, evapotranspiration, and runoff within the expected ranges. The mean annual precipitation input to the model from the Daymet dataset was 49.0 in.; the expected value for the Squannacook River watershed is about 48 in. The mean annual evapotranspiration computed by the model for the watershed was 20.9 in.; the expected value is about 23 in. The mean annual runoff simulated by PRMS was 22.5 in., slightly larger than the concurrently measured 21.3 in.

In addition to the general water balance, the calibrated model simulations of streamflow were compared with the USGS measured streamflows using the NSE statistic, the PBIAS, the NRMSE expressed as a percent, and the log residual of the simulated streamflow minus the measured streamflow. The mean and the standard deviation of the log residual reduce the weight (effect) of low- and- high-streamflow errors and thus provide a better picture of the typical residual error of a time series in which the daily and monthly residual errors are averaged. These statistics of the calibrated model runs were computed using the period of 1981–2013.

Using the Daymet climate data, which calibrated the original HNM model, the Squannacook River PRMS model showed good simulation of the mean streamflow (fig. 3; table 2). The error for the mean-annual simulated daily streamflows was −6.61 percent, which indicates that the annual water balance, on average, was well simulated. The NSE, log residual, the PBIAS, and the NRMSE together characterize the overall error and bias in the simulated hydrograph compared with the measured hydrograph. The NSE statistic is a measure of how well a simulated time series (streamflow hydrograph) fits the measured time series during the period of record relative to the mean, accounting for timing and quantity. A value of 1 for the NSE indicates that the simulated and measured time series are identical, and a value of 0 indicates that the simulated time series provides as much predictive information as the mean of the measured time series alone. An NSE statistic of 0.5 or higher indicates a satisfactory or “good” calibration; 0.4 to less than 0.5, a marginal or “fair” calibration; and less than 0.4, a relatively “poor” calibration (Moriasi and others, 2007). The NSE for the model was 0.647 and 0.524 using the Daymet and Livneh climate datasets, respectively. For this study, the mean log residual of 0.0062 indicates a satisfactory calibration; an NRMSE of less than 10 percent is considered satisfactory; and a PBIAS greater than 20 percent is considered to represent a fair or poor simulation (Moriasi and others, 2007).

Table 2.    

Statistics calculated for streamflows measured at the U.S. Geological Survey streamgage 01096000 (Squannacook River Near West Groton, Massachusetts) and Precipitation-Runoff Modeling System simulated streamflows using Daymet and Livneh and others (2015) climate input data.

[Location of streamgage is shown in figure 1. Data are from Olson and others (2024). Daymet climate dataset from Thornton and others (2016). PRMS, Precipitation-Runoff Modeling System; ft³/s, cubic foot per second; —, not applicable; NRMSE, normalized root-mean-square error; percentile is the percentage of streamflows in the record that exceed the listed streamflow value]

Table 2.    Statistics calculated for streamflows measured at the U.S. Geological Survey streamgage 01096000 (Squannacook River Near West Groton, Massachusetts) and Precipitation-Runoff Modeling System simulated streamflows using Daymet and Livneh and others (2015) climate input data.

Statistic1	Measured streamflow at calibration streamgage 01096000 (Squannacook River Near West Groton, Massachusetts)	PRMS simulated streamflow with selected climate input data
		Daymet climate data	Livneh and others (2015) climate data
Mean of streamflow (ft3/s)	121	113	107
Standard deviation of streamflow (ft3/s)	168	126	119
Percent error of the mean	—	−6.61	−11.6
Percent error of the standard deviation	—	−25.0	−29.2
Percent bias	—	−6.88	−11.3
NRMSE,2 in percent	—	2.75	3.19
Nash-Sutcliffe Efficiency coefficient	—	0.647	0.524
Mean of the log residuals	—	0.0062	0.019
Standard deviation of the log residuals	—	0.245	0.265
99th-percentile streamflow (ft3/s)	7.98	5.32	3.89
90th-percentile streamflow (ft3/s)	17.0	18.3	17.7
80th-percentile streamflow (ft3/s)	26.3	32.3	29.9
20th-percentile streamflow (ft3/s)	172	163	157
10th-percentile streamflow (ft3/s)	263	238	230
1st-percentile streamflow (ft3/s)	744	619	558

¹

For the period 1981—2013.

²

Normalized on the maximum range of flows in the record.

Whereas the goodness-of-fit statistics listed in table 2 show that the simulation has a good calibration on the basis of the NSE, PBIAS, NRMSE, and mean log-residual, the error of the standard deviation of the log residuals and the relatively large percent error (compared to the mean statistics) of the standard deviations (more than 10 percent) indicates that the simulation is not representing the range of flows as well as it is representing the mean and timing of the streamflow. The large percent error of the standard deviation is the result of the large relative errors at the high and low ends of the streamflow distribution (fig. 3; table 2). The larger errors on the high and low ends of the discharge range are relatively common with deterministic rainfall-runoff models (Farmer and Vogel, 2016) and is the reason this study uses stochastic watershed modeling.

As stated in “The Stochastic Weather Generator” section, the SWG was used to generate climate input data for warming scenarios that could potentially occur in future climates. Whereas the PRMS model was calibrated using the Daymet climate data, the Livneh climate dataset was used to develop the SWG because of its record length (1950–2013). The evaluation of the PRMS model using the Livneh climate dataset was deemed satisfactory with the goodness-of-fit statistics shown in table 2.

Stochastic Watershed Model

Because of the performance inadequacies involved with simulating the extreme high- and low-flow events commonly exhibited by deterministic models, simulated daily mean streamflows often have lower variance than the observed record the model was based upon. Relying on the output of deterministic models can lead to a systematic bias of extreme events. To address these performance issues (low variance and bias), a SWM was developed at Tufts University (Shabestanipour and others, 2023). The SWM applies model residuals to the simulated streamflow output producing a stochastic output of streamflow ensembles to better represent observed extreme streamflows.

The SWM is an autoregressive model of the logarithm of the ratios of observed and simulated streamflows combined with a bootstrap residual resampling approach to generate streamflow ensembles. Details of the SWM and its validation and verification are in Shabestanipour and others (2023). For validation purposes, the SWM was used to generate 10,000 realizations of daily streamflow records. Shabestanipour and others (2023) compared the percent error of the mode of the 10,000 low-flow and flood-flow statistics of the SWM output with the percent error of the PRMS results. The SWM output showed improvements over the deterministic PRMS model output (table 3).

The SWG and SWM were developed concurrently using different meteorological datasets with different periods of record, as described in “Evaluation of the Squannacook River Watershed PRMS Model Performance” section. The SWG was developed using the longer Livneh dataset (1950–2013), whereas the SWM was developed using residuals from the PRMS model with the Daymet dataset (1980–2017) used as input. The change in meteorological product altered the performance of the PRMS model and the characteristics of the residuals. Because the SWM operates by replicating the characteristics of the residuals, the change of meteorological product was conveyed to the operation of the SWM (by means of the altered characteristics of the residuals), thereby degrading the performance of the SWM. Though the SWM could be altered to account for the different residual structure, or the PRMS model recalibrated with the Livneh climate dataset, both options were deemed beyond the scope of this pilot project, except for a small change to the bootstrapping approach. To account for any emergent biases introduced by the different meteorological dataset, changes to flood- and low-flow statistics resulting from this investigation are reported in terms of percent changes from the baseline 0 °C warming scenario.

The SWM was applied to PRMS output for each of the 17 temperature change scenarios: 0 °C to 8 °C, in 0.5-degree increments. Using output from the SWG, PRMS simulated 100 ensembles of 64 years of daily streamflow estimates for each of the 17 temperature scenarios. The SWM was used to generate 10,000 realizations from each ensemble, for a total of 1 million realizations for each temperature scenario. Because of the degraded performance of the SWM discussed above, the SWM procedure was altered slightly from Shabestanipour and others (2023) by using separate, simple bootstrapping for simulated streamflows above and below 10 cubic feet per second for the autoregressive model. The residuals used in the bootstrapping technique were resampled from the run of PRMS that used Livneh dataset inputs. For each realization, the annual maximum flood flow at the 2-, 5-, 10-, 25-, 50-, 100-, and 500-year recurrence intervals; and the 7-day low flows at the 2- and 10-year recurrence intervals were determined. Also, the percent changes from the 0 °C warming scenario was computed for each realization.

Baseline Streamflow Characteristics

The output from the SWM of 0 °C temperature change scenario was used to compute the baseline streamflow characteristics. Each temperature scenario covered the 64-year modeled period and had 1 million realizations from the SWM. For each realization, the 2-, 5-, 10-, 25-, 50-, 100-, and 500-year recurrence intervals of the annual maximum daily discharge; and the 7-day low flow at the 2- and 10-year recurrence intervals were determined. The streamflow statistics were determined on a calendar year basis and were computed using a log-Pearson type 3 distribution. The median of each flow statistic from the realizations is considered the final SWM result.

For verification of the baseline streamflow characteristics (0 °C temperature scenario) from SWM, the streamflow characteristics were compared to streamflow characteristics computed from the observed record (1950–2013) of streamgage 01096000 using table 4. Table 4 also provides streamflow statistics from the 64 years of PRMS deterministic model output with the Livneh dataset (1950–2013) as input. Although the SWM did not have extreme streamflow results that were more comparable to the observed streamflow than the PRMS results, the goal of the study was limited to determining the percent change in streamflow characteristics for future warming scenarios to account for biases in results.

Streamflow Characteristics for Warming Scenarios

Streamflow statistics for the 2-, 5-, 10-, 25-, 50-, 100-, and 500-year recurrence intervals of the annual maximum daily discharge and the 7-day low flow at the 2- and 10-year recurrence intervals were determined from the output of the SWM for the warming scenarios in the same manner as done for the baseline conditions. Then, for a given temperature scenario, the median and the 5th and 95th percentile of the percent change in streamflow statistics from the baseline conditions were determined (tables 5 and 6) for the flood-flow and low-flow characteristics, respectively.

Table 5 indicates that floods of all recurrence intervals increase with warming. For example, the 500-year flood has an 82.9 percent increase for the 8 °C warming scenario compared to the 0 °C warming scenario. Figure 4 also illustrates that the increase caused by warming is greater with each recurrence interval. That is, for a fixed warming scenario, the streamflows at the greater recurrence intervals are expected to increase more than the lower recurrence intervals. Figure 5 illustrates a similar dynamic for low flows. Under warming conditions, low streamflows are expected to decrease. In addition, for a fixed warming scenario, the 7-day low streamflow at a 10-year recurrence interval is expected to decrease more than the 7-day low streamflow at a 2-year recurrence interval. That is, for the 8 °C warming scenario, the median 7-day low streamflow at a 10-year recurrence interval is projected to decrease by 36.0 percent compared with the 0 °C warming scenario, whereas the median 7-day low streamflow at a 2-year recurrence interval decreases by 21.3 percent for the same warming scenario (8 °C; fig. 5).

Future Streamflow Characteristics

To evaluate the streamflow characteristics that may occur in response to climate change in the future—for instance, the decades of 2030, 2050, 2070, and 2090—the warming scenario applicable to these target decades is needed. Estimating the localized warming scenario for future projections requires data from downscaled Global Circulation Models (GCMs); such data were obtained from the Multivariate Adaptive Constructed Analogs (MACA) archive (Abatzoglou and Brown, 2012; Abatzoglou and Brown, undated). The MACA archive contains downscaled output from 20 GCMs of the Coupled Model Inter-Comparison Project Phase 5 for the contiguous United States and for future Representative Concentration Pathways (RCPs) with a historical baseline of 1950–2005. For this investigation, the RCP8.5 emissions scenario was used. RCP8.5 is an emissions trajectory that assumes a future, in 2100, in which no actions are taken to reduce emissions, equivalent to radiative forcing of more than 8.5 watts per square meter (IPCC Core Writing Team and others, 2014, p. 127).

From the MACA-downscaled GCM projection data, the 30-year average annual temperature changes around each of the target decades was determined for the Nashua River watershed area (the Squannacook River watershed is within the Nashua River watershed). From these ensembles of data, the medians of the ensembles were determined and are shown in table 7.

Using table 7, one can determine which warming scenario to use for a target decade. For example, to determine the percent change of the annual daily-maximum streamflow at the 50-year recurrence interval for the 2070s, table 7 could be used to find the temperature change for 2070 of 4.64 °C. The user would then select the closest warming scenario to 4.64 °C, 4.5 °C, to find in table 5 that the 50-year recurrence interval streamflow estimate would increase by about 23 percent above the baseline conditions (0 °C scenario) on the Squannacook River.