Peak flow regime changes following forest harvesting in a snow-dominated basin: Effects of harvest area, elevation, and channel connectivity

Authors


Corresponding author: Y. Alila, Department of Forest Resources Management, University of British Columbia, 2045-2424 Main Mall, Vancouver, BC V6T 1Z4, Canada. (younes.alila@ubc.ca)

Abstract

[1] Numerical modeling in conjunction with a stochastic weather generator was used to investigate the immediate impact of forest harvesting upon the annual-maximum peak discharge regime of 240 Creek, a snow-dominated headwater basin of low relief located in south-central British Columbia (BC), Canada. Harvesting effects were simulated using 11 hypothetical harvest scenarios that specifically assess the impact of clear-cut harvesting in the absence of roads. Distribution statistics show that the annual-maximum peak flow frequency curve for hourly discharge is affected both by harvest area, AH, and by harvest elevation, ZH. Annual peak discharge magnitude tends to increase with increasing AH, but decreasing ZH. Forest harvesting does not have a statistically significant (α = 0.05) impact on the peak flow distribution until AH ≥ 20% to 30%, depending on the statistical comparison used. There is no substantial elevation gradient in snow accumulation within 240 Creek, and the sensitivity of the hourly peak discharge regime to ZH is instead related to the spatial variation of channel density with elevation. Flood frequency analysis was used to directly compare control and treatment events of equal frequency of occurrence, using return periods, T, of 1.003–100 years. This analysis indicates that, contrary to the prevalent hydrological wisdom and regardless of AH and ZH, the relative increase in peak discharge quantiles (ΔQT) increases with increasing event magnitude (increasing T) for events equal to or larger than the median (T = 2 years) event. Overall, ΔQT ranges from 7% for AH = 30% to 84% for AH = 100%. The magnitude of peak flow change is found to be a function of changes in both net 48-h input fluxes (rainfall plus snowmelt minus evapotranspiration) and changes in snowmelt runoff synchronization, which directly affect treatment peak discharge magnitude and timing.

1. Introduction

[2] In the Okanagan region of south-central British Columbia (BC), the majority of streamflow originates from snowpacks in high-elevation forested basins, many of which are managed, in varying priority with other activities, for timber extraction. Commercial forestry activities affect the hydrological processes generating streamflow, potentially altering the natural flow regime. Shifts in the streamflow regime can have substantial consequences arising from changes to the distribution of water, energy, sediment, and nutrients within the channel and its associated riparian area [Poff et al., 1997; Ziemer and Lisle, 1998], potentially jeopardizing the availability and quality of stream-associated resources. Annual peak flow events are considered a particularly important component of the streamflow regime; in mountainous terrain, geomorphic and ecologically significant discharge is often associated with peak flow events with recurrence intervals of anywhere from 1 to >100 years [Naiman and Bilby, 1998; Beschta et al., 2000; Church, 2002]. Additionally, changes in peak discharge magnitude can adversely affect forest-road stream crossing and conveyance structures, which must, by law in BC, be designed to withstand the 100-year flood [BC Ministry of Forests, 1995]. Therefore, with respect to a range of issues, forest planners require knowledge of how forest operations may affect the peak discharge regime for a wide range of recurrence intervals.

[3] In high-elevation watersheds, radiation fluxes dominate the energy budget during the snowmelt period [Male and Granger, 1981] due to the suppression of turbulent fluxes from high atmospheric stability over snow surfaces [Adams et al., 1998]. The spatial and temporal variability of net radiation fluxes and the subsequent snowmelt and runoff rates is governed by the spatial variability of topography and vegetation [Tarboton et al., 2000; Sicart et al., 2004; Hardy et al., 2004]. By removing the forest canopy, energy inputs at the snow surface are increased, mostly in the form of increased net radiation, resulting in increased snowmelt and runoff fluxes [Troendle and Reuss, 1997]. The degree to which snowmelt runoff and the subsequent discharge is affected depends upon the density and spatial distribution of the original forest cover [Winkler et al., 2005]. It is expected then that the relative impact of forest harvesting upon annual-maximum peak flow is sensitive to basin topography, physiography and the spatial distribution, size, and type of forest management practices [Alila and Beckers, 2001].

[4] Empirical evidence regarding the impact of forest harvesting on peak flows gathered throughout the interior west of North America derives from a variety of climatic, physiographic, and topographic settings under differing harvesting treatments over varying periods of observation and, as such, offers mixed results [Troendle and King, 1985, 1987; Cheng, 1989; Gottfried, 1991; Burton, 1997]. No consistent relationships between mean peak flow change, harvest treatment, and catchment properties are evident from these studies. This has often led to the conclusion that the effect of forest harvesting on peak flows is highly variable and unpredictable [e.g., Leopold, 1972; Hewlett, 1982; Eaton and Church, 2001; Scherer and Pike, 2003; Guillemette et al., 2005; Eisenbies et al., 2007; Grant et al., 2008]. Those few studies that have explicitly attempted to relate peak flow change to return period [e.g., Jones and Grant, 1996; Jones, 2000; Thomas and Megahan, 1998; Beschta et al., 2000; Moore and Wondzell, 2005] have been recently challenged on methodological grounds by Alila et al. [2009] because they were based on the premise that the effects of harvesting is a difference in the magnitude of peak flows when control and harvested basins are subject to the same storm input (pluvial regime) or freshet season (nival regime). Such chronological type of peak flow event pairing, a hallmark of traditional paired-basin data analyses in decades of forest hydrology literature, leads to erroneous outcomes because it does not explicitly consider the all-important highly nonlinear relationship between peak flow magnitude and frequency [Alila et al., 2010; Schultz, 2012; Green and Alila, 2012].

[5] Alila et al. [2009, 2010] advocate that changes in both the magnitude and frequency of peak discharge events due to forest disturbance should only be assessed via techniques that directly compare peak flow events on the basis of equal event frequency and not equal chronology, which can only be conducted via a direct comparison of the preharvest and postharvest peak flow frequency distributions. However, for the range of events under consideration in this study (e.g., up to the 100-year event), such an approach requires large sample sizes deriving from stationary distributions. A challenging issue is that data derived from paired-basin studies to date are often of insufficient duration with which to quantify the preharvest and postharvest peak flow regime; either due to short preharvest and postharvest observation periods, or due to the confounding effect of nonstationarity deriving from vegetation recovery in longer paired-basin studies. Hence, there are usually an inadequate number of observations with which to draw any meaningful conclusions about forest harvest effects on large, infrequent events [Thomas and Megahan, 1998; Beschta et al., 2000]. Consequently, numerical simulation is often the only recourse to generate sufficiently large preharvest and postharvest peak flow samples with which to effectively assess harvest impacts to the peak flow regime. In this regard, paired-basin studies offer tremendous value, particularly with respect to affording a real-world study prototype, in some cases providing examples of a range of disturbance types, with the necessary data and experimental information for testing, calibration, and validation of deterministic models.

[6] To derive a greater understanding of how land-use activities may affect watershed processes in the south-central interior of BC, a long-term paired-basin study, the Upper Penticton Creek (UPC) watershed experiment, was launched in 1984 [Winkler et al., 2003]. In addition to observations of streamflow and extensive climate monitoring, field data collected to date include direct observations of forest and open radiation balance, snow accumulation, and melt [Spittlehouse and Winkler, 2002], rainfall interception [Spittlehouse, 1998], soil water content [Spittlehouse, 2000], tree transpiration [Spittlehouse, 2002], forest and clear-cut stand water balance [Spittlehouse, 2006], and runoff generation and shallow groundwater table dynamics [Kuraś et al., 2008]. The various experiments conducted within UPC provide insight into the climatic and physiographic controls that affect small-scale hydrologic processes, affording a greater understanding of the means by which forest removal affects plot-scale runoff generation. However, extrapolation of these results to the basin scale is affected by spatial variability in controlling factors (such as vegetation, soil, topography, and meteorology). Therefore, to integrate stand-level experimental information into a physically reasonable description of basin response to forest harvesting, a modified version of the distributed hydrology soil vegetation model (DHSVM) [Wigmosta et al., 1994] has been developed for the study area, initially tested by Thyer et al. [2004] and further validated by Kuraś et al. [2011].

[7] This paper describes the use of the UPC DHSVM application to conduct an exploratory analysis of the impact of forest harvesting upon peak annual discharge regime in the UPC study area. The specific objective of this study is to explore the effect of clear-cut harvest size and elevation upon postharvest peak discharge magnitude and frequency, using methodology that explicitly addresses changes in the distribution of peak flow events. Resultant peak flow regime impacts are examined within the hydrological context provided by the specific topography, morphology, physiography, and spatial scale of a single headwater catchment. The current study forms part of an ongoing project to examine the impacts of forest harvesting in BC using numerical modeling to supplement field experiments [Alila and Beckers, 2001]. It utilizes an approach identical to that employed by Schnorbus and Alila [2004a] for Redfish Creek in southeastern BC. Compared with the 25 km2 Redfish Creek basin, the 5 km2 UPC watershed affords an opportunity to assess harvest impacts in a smaller headwater basin located in the much gentler topography, more open forests, and much drier climate characteristic of the south-central interior of BC. The current study also expands on the work of Kuraś et al. [2012], who investigated the effects of harvesting and, in particular, roads on the peak flow regime of a neighboring headwater basin in the UPC study area, by using a larger number of scenarios to more explicitly address the effects of harvest rate and elevation.

[8] This paper is organized into the following sections: Section 2 provides a brief description of the study area and collected hydro-meteorological data. Section 3 describes the research method and includes a brief overview of the DHSVM, a description of the harvest scenarios, the use of flood frequency distributions to compare preharvest and postharvest peak flow response, and a summary of the generation of synthetic meteorological input time series. Results and discussion of the relevant hydrological processes thought to govern the results are presented in section 4. Concluding remarks are provided in section 5.

2. Study Area and Data

[9] The study focuses on 240 Creek, which forms the control basin in the UPC paired-basin experiment for two treatment basins, the adjacent 241 Creek (Figure 1) and Dennis Creek (not shown). The study area is located in the Okanagan Highlands roughly 25 km northeast of Penticton in south-central BC. The physiographic characteristics of 240 Creek are summarized in Table 1.

Figure 1.

Location of the Upper Penticton Creek study area in south-central BC and detailed map of 240 Creek and neighboring 241 Creek.

Table 1. Physiographic Characteristics of 240 Creeka
Descriptor240 Creek
  1. a

    Derived from 30 m DEM.

Drainage area, Ad (ha)470
Forested area, Af (ha)437
Minimum elevation, Zn (m)1609
Maximum elevation, Zm (m)2036
Relief, ZmZn (m)427
Mean elevation (m)1782
Mean aspect (degrees clockwise from North)153
Mean slope (%)24.3
Total channel length, Lc (km)6.2
Drainage density, Lc/Ad (km/km2)1.3

[10] The study area occupies the dry, cold subzone of the Engelmann Spruce-Subalpine Fir biogeoclimatic zone [BC Ministry of Forests, 2003]. Mean summer (June–August) and winter (November-March) air temperatures are 11 and −5°C, respectively, and average 2°C over the year. Mean annual precipitation is 750 mm, of which about half falls as snow. Seasonal snow cover persists from late October until May (or early June in some years). Due to the absence of any significant precipitation gradient (see section 3.5.4) and winter temperatures that are uniformly below freezing across the study area (i.e., winter precipitation falls as snow at all elevations), there is no appreciable gradient of snow accumulation with elevation (unpublished data, BC Ministry of Forests, Lands and Natural Resources Operations; http://www.for.gov.bc.ca/rsi/research/penticton/data.htm). The melting of this snowpack dominates the annual hydrograph, which peaks in late spring to early summer. Rainfall is a rare event during the snow accumulation season and winter rain-on-snow events are not known to occur in the UPC study area (D. Spittlehouse, personal communication). Late season streamflows are sustained by baseflow and rain. 240 Creek has a relatively open forest canopy containing predominantly mature lodgepole pine (Pinus contorta Dougl.) with small amounts of Engelmann spruce (Picea Engelmannii Parry) and subalpine fir (Abies lasiocarpa (Hook.) Nutt). The bedrock of the Okanagan Highlands consists of unfractured intrusive granodiorite and metamorphic granitic orthogneiss [Hudson and Golding, 1997]. The soil, derived from glacial till and local bedrock, is predominantly a sandy loam and loamy sand with a high-coarse fragment content. The soil mantle ranges in depth from 0.1 to 2.0 m and is generally well to rapidly drained [Hope, 2001]. The geology of 240 Creek is assumed to give a tight water balance, ensuring that the majority of net precipitation (precipitation less abstractions) exits each basin as streamflow.

[11] Air temperature, relative humidity, shortwave radiation, wind speed, and precipitation (rain and snow) have been measured on an hourly basis in large forest openings at the lower elevation P1 site (1620 m; Figure 1) since August 1997 and at the upper elevation PB site (1900 m; Figure 1) since September 1999. Details regarding the collection of meteorological data are fully described by Thyer et al. [2004] and Kuraś et al. [2011]. Hourly streamflow has been measured by a Water Survey of Canada gauging stations since 1986 (Figure 1).

3. Methods

3.1. Overview

[12] In general, our study approach adopts a paired-watershed experimental design but deviates from the approach using a numerical model to simulate both the control and treatment watersheds. Past approaches using numerical simulation have mostly focused on the retrospective detection of nonstationarity in streamflow time series in an attempt to analyze the impact of actual (and often gradual) changes in forest cover over a given study period [e.g., Bowling et al., 2000; Andréassian et al., 2003]. This retrospective approach relies on the hydrologic model to represent a control watershed (i.e., describing some past forest cover state) that is subject to the same meteorological forcings observed during the study period; the simulated streamflow (stationary with respect to land use) and observed streamflow (potentially nonstationary with respect to land use) are compared, and the residuals are analyzed for trend. In the current study, however, we are not concerned with the detection of streamflow trend in response to changing forest cover over time, but with the immediate impact of forest harvesting upon the magnitude and frequency (i.e., regime) of annual peak discharge. As such, our approach is to use a hydrologic model to simulate the control (current forest cover) and treatment (hypothesized harvesting) watersheds subject to an ensemble of identical meteorological forcing, with both watersheds based on the same real-world prototype of 240 Creek. This then allows us to compare two stationary peak flow frequency distributions, one the natural (i.e., unmanaged) flow regime and one that arises immediately following human disturbance. Such a study design filters out any effects due to temporal and spatial variation in climate and physiography and isolates the effect of forest harvesting upon the annual-maximum peak flow regime.

[13] The experiment utilizes numerical simulation to generate a sample of size N = 100 of annual maximum streamflow events for 12 harvest scenarios (including a control scenario). Each scenario was conducted as a continuous 100-year simulation and run at an hourly resolution for the period 1 January, simulation year 00 (SY00) to 30 September, SY100. The period from 1 January to 30 September, SY00 was used to “warm-up” each run. Analysis of output results was limited to the period 1 October, SY00 to 30 September, SY100. Between scenarios forest cover parameters were adjusted to reflect different proportional areas and elevations of clear-cut harvesting activities, while all remaining basin parameters were held constant. Each scenario was run with an identical 100-year time series of synthetic meteorological input (see section 3.5), and forest parameters were kept constant during each scenario (i.e., no regrowth or vegetation succession). Initial conditions (soil moisture, interception storage, snow water equivalent, and channel network storage) for 1 January SY00 were identical between scenarios. The continuous simulation approach allowed initial conditions for subsequent years to vary between scenarios as a function of forest cover. The 100-member ensemble of annual streamflow simulations for each scenario was used to construct the annual maximum series (AMS) by selecting the largest discharge events, Qp, from every year. Each AMS was then used to estimate the frequency distribution of annual maximum peak flows.

3.2. DHSVM UPC Application

[14] The hydrologic response of the study basin was simulated with the DHSVM of Wigmosta et al. [1994]. This is a physically based distributed parameter model that explicitly estimates the spatial distribution of water and energy fluxes by subdividing the model domain into small computational grid elements based on a 30 m digital elevation model (DEM). Vegetation and soil properties are assigned to each grid cell and are allowed to vary spatially throughout the watershed. A complete description of model structure and governing equations is given by Wigmosta et al. [2002], with recent modifications for the UPC study area described by Thyer et al. [2004]. One-dimensional energy and water balances are solved individually for every grid cell in the model domain at an hourly time step, considered the best representation of the diurnal meteorological fluctuations and rapid streamflow response that occurs within the study area. The derivation and assignment of land-cover and soil properties to each grid cell is conducted by the use of geographic information system (GIS)-based analysis. The model stream network was extracted from the DEM and constrained by ground-based Global Positioning System (GPS) survey of the actual channel system. Energy balance and vegetation parameters were calibrated and validated using split samples of internal catchment data collected at both 240 and 241 Creeks. The calibration and validation of soil parameters was done using split sample discharge data from the fully forested 240 Creek only. Discharge data from 241 Creek, which was 20% clear-cut harvested, was used to further validate and test the transferability of model parameters. The preparation of input data, calibration, and validation of the UPC DHSVM application is fully described by Thyer et al. [2004]. In addition, this DHSVM application at UPC has more recently proven to be successful in realistically simulating spatiotemporally variable runoff generation processes in further performance evaluations [Kuraś et al., 2011] and is believed to be a reliable enough tool for assessing the effects of forest harvesting on the peak flow regimes of snow-dominated watersheds.

3.3. Harvest Scenarios

[15] Hourly AMS samples were derived for the control and 11 hypothetical scenarios representing clear-cuts of different proportions (AH, % of forested basin area) in two elevation bands (ZH, bottom or top). The control vegetation cover (representing current conditions) for 240 Creek, given as vegetation classes, is shown in Figure S1 in the supporting information, with the overstory properties of each class described in Table S1. The two elevation bands are partitioned such that both bands contain 50% of the total forested area (i.e., areas exclusive of bare rock; vegetation class 1). Due to the larger proportion of bare rock at higher elevations, the area of the top elevation band is 18% larger (254.0 ha) than the total area in the bottom elevation band (215.8 ha). As such, the bands are divided by the H54 elevation (that elevation above which lies 54% of the basin area; 1750 m above sea level). Clear-cuts were introduced by converting mature forest vegetation classes 2–9 to vegetation class 10; areas composed of vegetation class 1 (rock) were not changed. The clear-cuts were designed as one contiguous block connecting opposite divides of each drainage basin to minimize any potential bias from variations in slope and aspect. Harvesting in the bottom elevation band occupies an area centered on the H77 elevation, whereas harvesting in the top elevation band occupies an area centered on the H27 elevation. The harvest scenarios are summarized in Figure 2, which shows the spatial distribution of all open areas (clear-cut and rock). Forest roads and the effect of soil compaction were not incorporated into the harvest scenarios. The harvest scenarios will henceforth be referred to in the text in Figure 2.

Figure 2.

Spatial distribution of clear-cut (class 10) and rock areas (class 1) for hypothetical harvest scenarios. Figure captions of the form “XXY” indicate proportion of forest area harvested (XX) and elevation band (Y; where B = bottom, T = top); except for 100% forest area harvested, which is indicated simply as “100.” The H54 elevation is indicated by a solid line.

3.4. Analysis

[16] The generated sample peak discharge distributions were compared using the Wilcoxon rank sum (WRS) test, which tests the null hypothesis that two samples (control and a treatment) are identically distributed versus the alternate hypothesis that the treatment sample tends to generate larger values than the control sample [Hirsch et al., 1993]. The two samples are combined, ranked, and then separated and the WRS statistic is the smaller of the two rank sums. The WRS test gives an indication of overall differences between two sample distributions, but changes in location and shape between the control and a treatment sample distributions are not specified. These changes were assessed by comparing the median peak discharge, Med[Qp], and the L-moment coefficients of variation (L-CV) and skew (L-CS) [Hosking, 1990]. A final comparison of the control and treatment distributions is based on Kendall's rank correlation (KRC), measured using the correlation coefficient, τ [Hirsch et al., 1993]; which tests the strength of the relationship between the two sample distributions of Qp. The fact that τ is rank based is of direct interest to us, as our motivation is to assess changes in the rankings of Qp events before and after harvesting [Alila et al., 2009]. The KRC is used to test the null hypothesis that two variables, x and y, are independent (i.e., the distribution of y does not change as a function of x).

[17] The generated peak flow regime for each scenario was quantified by flood-frequency analysis of the simulated AMS using the generalized extreme value (GEV) distribution with parameters estimated using the sample L-moment statistics [Stedinger et al., 1993]. Peak flow quantiles, QT, were estimated for return periods, T, of 1.003, 1.25, 1.5, 2, 5, 10, 20, 50, and 100 years (which correspond to cumulative probabilities of 0.003, 0.02, 0.2, 0.5, 0.8, 0.9, 0.95, 0.98 and 0.99, respectively). Change to the peak flow regime for each harvest scenario was assessed by determining the relative change in the estimated peak discharge quantiles, ΔQT, calculated as inline image, where superscripts C and H refer to the control and treatment values, respectively [Alila et al., 2009]. The ΔQT values, which are a more specific manifestation of differences in the location and shape between the control and treatment distributions, were tested for statistical significance. Assuming that the GEV parameters estimated from the sample Qp (either observed or simulated by DHSVM) are the “true” parameters, the distribution of the quantile estimator for each T was approximated using Monte Carlo simulation, as follows: (1) a Qp sample of size N was randomly drawn from the true GEV distribution, (2) the GEV parameters were re-estimated using the resampled Qp, (3) quantiles, QT, were recalculated using the reestimated GEV parameters, and (4) steps 1–3 were repeated 10,000 times. The 10,000 estimates of QT for a given T were ranked and used to directly approximate the 95th, 99th, and 99.9th percentiles of the quantile estimator. The percentiles of inline image were used to test the significance of inline image, i.e., when values of inline image exceeded the 95th, 99th and 99.9th percentile of inline image, increases in peak flow were considered indicative, significant, or highly significant, respectively.

[18] It is presumed that in any particular year changes in peak flow due to forest removal can be explained by changes in volume of moisture available for runoff, for example via increased snowmelt or decreased evapotranspiration. Hence, we attempt to explain the effect of forest removal on peak flow generation using the 48 h net moisture input flux, Fn48. This is defined as the total basin-wide net moisture input flux that accumulates over the 48 h period prior to occurrence of peak discharge, calculated as

display math(1)

where M, R, and ET are basinwide snowmelt, rainfall and evapotranspiration (including soil evaporation, evaporation of intercepted snow and water and transpiration), respectively at some hour, t, and tp is the hour (from 1 April of the given year) of peak hourly discharge. Fn48 is normalized by basin area and presented as an equivalent depth. A 48 h integration period in (1) was chosen as it tends to encompass the period of highest correlation between snowmelt and discharge during the spring freshet period (section 5.2). Snowmelt is defined strictly as the conversion of ice, stored in the snowpack (including any snowfall that may occur during the 48 h period), to water, whereas rainfall is strictly liquid precipitation. Evapotranspiration includes soil evaporation, evaporation of intercepted snow or water and transpiration. Note that Fn48 is evaluated based on the date/time of peak flow occurrence, tp, which can change due to harvesting, as opposed to some arbitrary fixed date. Hence, forest cover removal can have both a direct effect on Fn48, predominantly through changes in the magnitude of M and ET, as well as an indirect effect via changes in tp, which changes the 48 h period over which M, ET, and R are integrated.

3.5. Synthetic Meteorology Generation

[19] The generation of a 100-year streamflow time series via continuous simulation requires an input meteorological time series of the same duration. This required the extension of the observed meteorological record collected at UPC by stochastic means. Due to the relatively short record length at the PB weather station, a synthetic meteorology series was generated only for the P1 station based on observations collected during the period August 1997 to December 2001. This generated time series can be considered a stationary reproduction of only this short climate period, and the severe limitation of using such a short calibration period to recreate the climatology of the study area is recognized. However, the intent is to use this proxy meteorology to generate weather phenomena (and peak discharge events) representative of the study area for use in ensemble analysis. The derived synthetic meteorology is considered sufficient for this purpose. To give an indication of how the climate of August 1997 to December 2001 compares with respect to the 1971–2000 climatology of the region daily average air temperature and total precipitation by month for Penticton Airport (Environment Canada station 1126150), and first-of-the-month SWE observed at the Greyback Reservoir (located 10 km from the study area at an elevation of 1550 m; BC Snow Survey Network station 2F08) were plotted as percentiles of the 1971–2000 climatology (Figure 3). The period of August 1997–2001 provides a range of conditions with respect to temperature and precipitation, with 1998 experiencing one of the warmest springs and the hottest summer of record and 1999 experiencing a very cool spring and summer; precipitation was highly variable, with large ranges observed during the late summer and fall. Snow accumulation (1 January through 1 April) for the years 1998–2001 tended from above normal to well below normal whereas the spring melt rate (1 April through 1 May) was above normal (1 May percentile < 1 April percentile) for 1998, 1999, and 2000 but well below normal for 2001.

Figure 3.

Climate of August 1997 to December 2001 compared to 1971–2000 climatology using (a) daily average temperature and (b) total precipitation by month for Penticton Airport, and (c) first -of-the-month SWE at Greyback Reservoir. All graphs plot 1997–2001 values as percentiles of 1971–2000 climatology for respective stations.

[20] The generation of the synthetic meteorological data follows the procedure described by Schnorbus and Alila [2004b] and is only summarized herein. The meteorological generation technique is initiated by the direct generation of hourly precipitation using the random pulse based stochastic model of Rodriguez-Iturbe et al. [1987, 1988]. This step is followed by the stochastic generation of daily weather variables (maximum and minimum air temperature, average wind speed, average dewpoint temperature, and global solar radiation) using a multivariate first-order autoregressive (MAR(1)) procedure conditioned upon generated daily precipitation occurrence [Parlange and Katz, 2000]. The daily meteorological data are disaggregated to derive meteorological data at an hourly resolution that, in conjunction with the generated hourly precipitation series, constitute the required hourly DHSVM input time series (precipitation, air temperature, relative humidity, wind speed, solar beam and diffuse radiation, longwave radiation, temperature lapse rate, and precipitation gradient). The generation of hourly air temperature, temperature lapse rate, and solar irradiance has since been modified from that described by Schnorbus and Alila [2004b] and is discussed below.

3.5.1. Hourly Air Temperature

[21] The diurnal fluctuation in hourly air temperature (Ta) was estimated by first modeling normalized air temperature using the first two harmonics of a Fourier series [Campbell and Norman, 1998]. Dimensional values of Ta were obtained using the generated values of daily maximum and minimum air temperature, assuming that the daily minimum air temperature normally occurs just prior to sunrise and daily maximum temperature occurs 2 h after solar noon. This process is assumed stationary throughout the year.

3.5.2. Temperature Lapse Rate

[22] Under ideal circumstances the temperature lapse rate (Tlapse) would be estimated based on the temperature gradient between two (or more) stations at different elevations [i.e., Schnorbus and Alila, 2004b]. However, in the current application meteorological data are only being generated at a single station, therefore, an alternative approach was adopted. Concurrent temperature observations between P1 and PB stations during September 1999 to December 2001 indicate that Tlapse tends to be higher (more negative) during the daytime and lower (less negative) at night. Temperature inversions occur frequently in all months, usually between sunset and sunrise. The minimum lapse rate generally occurs just prior to sunrise and inversions, when they occur, break up almost immediately after sunrise. This process suggests that the magnitude and sign of Tlapse between P1 and PB is closely related to the diurnal process of cumulative heating and cooling of the atmospheric boundary layer by surface heat fluxes and the resultant occurrence and degree of atmospheric stability [Stull, 2000]. It has been assumed that the total daytime heating of the boundary layer is proportional to the average daytime temperature. The degree of cumulative heating or cooling of the boundary layer for a given hour, t, on any given day, n, is taken as the difference between the current air temperature and the running daytime average air temperature as ΔT(t, n) = Ta (t, n) – Tavg (t, n); where Tavg (t, n) for hour, j, between sunrise (t = tr) and sunset (t = ts) on day n is inline image and Tavg for t > ts on day n and t < tr on day n + 1 is inline image. In general we have found that Tlapse is inversely dependent upon ΔT and the functional relation is well described by a second-order polynomial. Examples of Tlapse versus ΔT for the months of February and July are given in Figure S2. Similar relations (i.e., second-order polynomial) have been fit for each month and are used to predict Tlapse from P1 ΔT. In general, temperature inversions occur/breakup when ΔT approaches −5°C, i.e., decreases (increases) with increasing elevation for ΔT less than (greater than) −5°C. Tlapse asymptotically approaches the dry adiabatic rate (−9.8°C (1000 m)−1) as ΔT increases. Prediction of Tlapse in this fashion is reasonably accurate for the months of February through October, with coefficients of determination (R2) ranging from 0.64 to 0.76. Prediction is less accurate for November (R2 = 0.48) and January (R2 = 0.42) and poor for December (R2 = 0.22). For December, it was found that Tlapse is better described as an inversely linear function of ΔT for ΔT < 4.0°C, and equal to the dry adiabatic lapse rate, i.e., −9.8°C (1000 m)−1, for ΔT ≥ 4.0°C.

3.5.3. Hourly Solar Irradiance

[23] It is assumed that atmospheric transmissivity, H, is uniform throughout the day and can be estimated from H = S/So, where S is daily global solar radiation and So is the potential (i.e., above-atmosphere) daily solar radiation calculated from solar geometry [i.e., Gates, 1980]. In the current application S is generated stochastically as part of the MAR(1) used to generate daily weather variables (vice using an empirical relationship between S and diurnal temperature fluctuations as was done in Schnorbus and Alila [2004b]). To satisfy assumptions regarding normality of the residual series in the MAR(1) model, a Box-Cox transformation of S using monthly parameters was required. The hourly values of direct and diffuse solar radiation are derived as per Schnorbus and Alila [2004b]. Solar radiation at P1 is calculated assuming a horizontal surface.

3.5.4. Local Meteorological Data

[24] Local hourly meteorological data for each model pixel are extrapolated during runtime from the time series provided for the P1 station. Air temperature is extrapolated by adjusting for local elevation using the estimated lapse rate. Wind speed, relative humidity, and downward longwave radiation are assumed to be uniform across the basin. Atmospheric transmissivity is assumed uniform across the basin for each time step, but direct solar radiation is adjusted locally to account for the effects of slope, aspect and topographic shading, and diffuse radiation is adjusted based on the local topographic skyview. Based on concurrent measurements of precipitation at the P1 and PB stations, no precipitation gradient is evident in 240 Creek. Note that winter precipitation at both stations has been corrected for gauge undercatch by scaling to match local measurements of snow accumulation (BC Ministry of Forests, unpublished data). No scaling factor was introduced to correct for summer precipitation (which is mostly convective rainfall and less prone to undercatch). Given the gentle relief and rolling topography of the study area, this lack of evidence for any substantial orographic enhancement of precipitation is assumed to apply to the basin as a whole. Consequently, no precipitation gradient is applied and precipitation is assumed to be spatially uniform [Thyer et al., 2004].

3.5.5. Performance of DHSVM Coupled With Synthetic Meteorology

[25] The performance of DHSVM using the synthetic meteorology data was assessed by comparing simulated point melt dynamics at three paired locations within both 240 and 241 Creek (Figure 1). Values were simulated using observed meteorology data over a four-year period and generated met data over a 15-year period (SY01 to SY15). The pixel locations were selected to sample melt in three forest types, classes 8, 4, and 6 for pixel-pairs S1, S2, and S3, respectively, and adjacent open areas (classes 1 and 10). Results are presented in Figure 4, which shows median values (±one quartile) of 1 April SWE and melt averaged over 1 April to 15 May plus forest-to-open (F/O) ratios of SWE and melt for each pixel pair. Point SWE and melt rates simulated using synthetic meteorology give a reasonable reproduction of the values simulated using observed meteorology (Figure 4), with the range of simulated values showing at least 50% overlap. The exception is the melt rate for S3-O, where the median values have been increased by 63% when using synthetic met data. Regardless, the ranges of simulated melt F/O ratios between the observed and generated meteorology, including pixel-pair S3, show acceptable overlap. Therefore, despite some bias in open melt simulation, the relative impact of forest removal on snowmelt is considered to have acceptable accuracy for an exploratory analysis.

Figure 4.

Median (±one quartile) of simulated point (a) 1 April SWE, (b) forest-to-open (F/O) ratio of 1 April SWE, (c) melt, and d) forest-to-open ratio of melt using observed and generated met data. Melt calculated as average rate over period 1 April to 15 May. Pixel locations are as shown in Figure 1.

[26] The combined DHSVM-meteorology generator model was also assessed for its ability to recreate the general characteristics of the flood frequency curves based on observed streamflow data. The hourly discharge flood frequencies for 240 Creek using observed streamflow (17 observations) and streamflow generated using the control scenario (100 observations) is shown in Figure 5. The two flood frequency curves are in general agreement throughout the range of T from 1.003 to 100 years, although simulated quantiles tend to be slightly overestimated at T ≈ 1.25 years. The generated curve is not statistically dissimilar from the observed curve; the two-sided WRS test is not significant (p = 0.32), and the control curve falls within the 5–95 percentile range of the observed curve. The interpercentile ranges for both curves shown in Figure 5 were determined by using Monte Carlo simulation, as described in section 3.4. The slope of the two flood frequency curves is also consistent; the ratio of 100- to 2-year peak discharge, Q100/Q2, is 2.1 and 1.7 for observed and simulated AMS, respectively. Some discrepancy between the two curves is expected; the observed peak flows, which derive from the period 1984 to 2001, are being compared to simulated peak flows that have been generated using meteorology representative of the climate during a limited portion of this record (1997–2001).

Figure 5.

Comparison of GEV fit to observed (N = 17) and simulated (N = 100) hourly annual peak flow frequency for 240 Creek, also showing 5–95 interpercentile range for observed (dashed lines) and control (bars).

4. Results and Discussion

4.1. Frequency Analysis

[27] The suitability of the GEV distribution for describing the synthetic AMS series can be examined using the probability plots given in Figure 6, which compares the empirical and fitted peak flow frequency for scenario 100 to that of the control. From these plots it is apparent that the synthetic data are suitably described by the GEV distribution. However, there is a tendency in some scenarios for events of T < 1.05 and T > 50 years to be slightly under- and over-estimated, respectively by the fitted GEV functions. As there is no obvious break in curvature with the control or treatment probability plots, the “sample” peak flows can be considered to derive from a homogeneous parent peak flow population.

Figure 6.

Comparison of control with scenarios 30B, 50B, and 100 sample and fitted (GEV) hourly flood frequency. Dashed lines indicated 1 to 99 interpercentile range for control scenario flood frequency.

[28] Comparison of the Med[Qp], L-CV, and L-CS statistics of the synthetic flood frequency curves shows the effect of both harvest area, AH, and harvest elevation, ZH (Table 2). The Med[Qp] increases consistently with increasing harvest area with very little difference between ZH. The largest Med[Qp] occurs for AH = 100%. The response of L-CV shows a slightly more complex relationship. L-CV peaks at intermediate harvest levels of AH = 20% and AH = 40% when harvesting occurs in the top and bottom elevation bands, respectively. The relative skewness (L-CS) of the derived flood frequency curves increases with increasing AH, with little difference between elevation bands.

Table 2. Distribution Statistics for Generated Peak Discharge Samples
StatisticSample Statistics by Scenario
C10B10T20B20T30B30T40B40T50B50T100
  1. a

    One-sided WRS test inline image is significant at p < 0.05.

  2. b

    One-sided WRS test inline image is significant at p < 0.001.

  3. c

    One-sided WRS test inline image is significant at p < 0.01.

  4. d

    Null hypothesis of independent samples rejected with p < 0.001.

Med[Qp] (m3/s)1.221.251.251.271.281.321.321.411.391.481.441.70
L-CV0.1550.1560.1590.1630.1680.1710.1600.1760.1580.1750.1570.154
L-CS0.0300.0570.0580.0730.0740.0700.0670.0660.0620.0680.0690.092
WRS statistic 88708821855685228138a8318a7734b8074c7358b7816b6056b
KRC τ 0.82d0.90d0.70d0.84d0.65d0.80d0.60d0.77d0.57d0.74d0.56d

[29] The results of the WRS test suggest that for AH ≤ 20% the treatment and control samples are identically distributed (Table 2). For AH = 30%, changes in the distribution of Qp are significant (treatment values tend to be higher than control values; α = 0.05) for both ZH. The distribution differences between the control and treatment scenarios become increasingly more significant with increasing AH. Differences between the control and treatment distributions are also higher when harvesting occurs in the bottom elevation band. Rank correlation decreases with increasing harvest area, although the correlations remain significant (p < 0.001) for all scenarios (Table 2). Regardless, this trend suggests that following forest harvesting there is a restructuring of the AMS such that control and treatment events in the same SY need not have the same discharge rank. Rank correlation is also sensitive to ZH and is lower when harvesting is in the bottom elevation band. The rank correlation statistic, τ, ranges from a high of 0.90 to a low of 0.56 for scenarios 10T through 100, respectively.

[30] Using the fitted GEV distributions, control flood quantile values and associated relative increases in hourly quantile discharge (ΔQT) are listed in Table 3. For all scenarios and quantiles ΔQT is positive following harvesting, although the magnitude of the impact varies with AH, ZH, and T, reflecting changes in the distribution statistics (Table 2). In general, ΔQT increases with increasing AH (for a given T and ZH). Also, ΔQT tends to be larger (for a given T and AH) when harvesting occurs in the bottom elevation band, with differences becoming more pronounced with increasing harvest area. For peak flow events equal to or larger than the median event (T ≥ 2 years), ΔQT tends to increase with increasing T for all scenarios, although this trend is more pronounced when harvesting occurs in the bottom elevation band. The range of quantiles that are significantly affected increases with increasing AH, and all quantiles are (highly) significantly increased for AH = 100%.

Table 3. The 240 Creek Control Scenario Quantile Magnitude and Relative Change in Flood Quantile Magnitude by Treatment Scenario (Harvest Area and Elevation) and Return Period
T (years)Control QTa (m3/s)ΔQT (%) by Treatment Scenario
10B10T20B20T30B30T40B40T50B50T100
  1. a

    Standard deviation in brackets.

  2. b

    ΔQT significant with p < 0.001 (highly significant).

  3. c

    ΔQT significant with p < 0.05 (indicative).

  4. d

    ΔQT significant with p < 0.01 (significant).

1.0030.39 (0.09)1291311718323123184b
1.050.68 (0.05)43657101014c17c19d55b
1.250.93 (0.04)22559c8c14b12d20b16b47b
21.21 (0.04)2256c11b8d17b11b23b15b44b
51.51 (0.04)237d8d14b10b21b12b28b16b45b
101.66 (0.05)349d10b16b11b24b13b30b17b47b
201.79 (0.064510b11b18b12b26b14b32b18b49b
501.93 (0.08)5612d13d20b14d28b16b35b20b53b
1002.01 (0.10)6714d15d22b15d30b17d37b21b55b

[31] Strictly looking only at changes in the magnitude of peak flow events, however, only considers one dimension of the peak flow regime. Looking at changes in event frequency can be just as revealing of the potential impacts of forest harvesting. For example, changes in event frequency (for a given magnitude) were estimated by inverting the fitted GEV functions for given event magnitudes (Table 4). In the case of scenario 30B the simulated 10-year event becomes twice as frequent (4 year event), the 20-year event becomes three times as frequent (6-year event), the 50-year becomes five times as frequent (10-year event), and the 100-year becomes seven times more frequent (14-year event). The change in event frequency becomes larger with increasing event magnitude; the larger the original event, the greater its increase in frequency following harvesting (Table 4).

Table 4. The 240 Creek Return Periods by Treatment Scenario (Harvest Area and Elevation) and Event Magnitude
QT (m3/s)Control T (years)T (years) by Treatment Scenario
10B10T20B20T30B30T40B40T50B50T100
0.391.0031.0021.0021.0011.0021.0021.0011.0031.0011.0021.0011.000
0.681.051.041.041.041.041.041.031.031.021.021.021.00
0.931.251.221.221.201.201.171.161.141.141.111.111.02
1.21222222212111
1.51544433323231
1.661087664535342
1.79201413996857462
1.9350292516151014712593
2.011004740242213219176133

[32] The relationship of ΔQT with T is generally consistent with the findings of Schnorbus and Alila [2004a], who conducted a study with a similar type of analysis in the 26 km2 Redfish Creek watershed in southeastern BC. However, the ΔQT values estimated for 240 Creek are two- to three-times larger than those estimated for Redfish Creek for comparable AH and identical T. A large proportion of Redfish Creek is composed of sub-alpine parkland (that is not subject to forest harvesting) above H40; this region develops a very deep snowpack that contributes the majority of melt runoff during peak annual discharge, thus moderating any peak flow changes due to forest harvesting at lower elevations. Additionally, the larger peak flow events in Redfish Creek were generated during periods of rain-on-melting snow, which are less sensitive to forest removal than the radiation melt events that dominate the UPC peak flow regime.

[33] As anticipated, the effect of forest harvesting on ΔQT for 240 Creek is also consistent to that assessed for neighboring 241 Creek (based on forest harvesting without roads) by Kuraś et al. [2012]. Nevertheless, despite strong similarities in climate, forest cover, drainage area and hypsometry between 240 Creek and 241 Creek, some differences in peak flow response are still evident at low harvest rates. For AH ≤ 30%, ΔQT for 240 Creek is higher for all T (occasionally by a factor of two or more), whereas results are similar for AH = 50%. Differences in peak flow response are presumably attributed to differences in drainage density (241 Creek has higher drainage density), predominant aspect (i.e., 240 Creek is predominantly oriented east-west, whereas 241 Creek is predominantly south facing; Figure 1) and the spatial distribution of cutblocks (Kuras et al. [2012] use historical as opposed to hypothetical cutblock locations.)

4.2. Snowpack Energy Balance

[34] Removal of the forest canopy alters the snowpack energy balance and subsequent melt rates. This effect is demonstrated by comparing the simulated energy balance in two adjacent model pixels conforming to site S2 in 241 Creek (Figure 1). The forest site represents vegetation class 4 (lodgepole pine), which is the predominant vegetation class in both the 240 and 241 basins (57% of combined area). In this example the energy balance is simulated in DHSVM using meteorological conditions observed at the P1 climate station from May 10 through May 28, 1999. This period includes the occurrence of the annual maximum peak discharge on 26 May. Results are given as Figure 7. The period was characterized by generally clear skies, punctuated with occasional periods of precipitation (including snowfall on 16 May). Daytime air temperatures were above freezing, but nighttime lows frequently dropped below freezing (Figure 7a). The presence of a forest canopy has a clear attenuating effect on all components of the energy balance. Net solar radiation is lower under the forest canopy than in the open. Longwave radiation tends to represent a heat loss to the snowpack in the open, particularly in the nighttime hours, whereas under a forest canopy longwave radiation tends to act as a source of heat during the day and a much smaller loss of heat during the night. The presence of a forest canopy acts to dampen near-surface wind speeds (Figure 7b), such that both sensible and latent heat fluxes in the forest snowpack are reduced as compared to the open site. Sensible heat generally acts as a source of energy, whereas latent heat acts as an energy sink in both the open and forest sites. Energy advected by rainfall is insignificant. Net solar radiation is generally the dominant source of melt energy in both the forest and open environments, although the occurrence of peak melt on May 24 coincides with the occurrence of high sensible heat (with a daytime maximum temperature of 20°C and peak wind speeds of 2.5 m/s in the open). The overall effect is higher net energy and higher melt rates in the open compared to the forest site (Figures 7c–7e). In both sites the total energy balance is generally lower during periods of precipitation (and overcast) than during clear-sky conditions.

Figure 7.

Observed meteorological conditions and simulated snowpack energy balance for neighboring and open and forested pixels (site S2, see Figure 1) for the period 10–28 May, 1999. (a) Observed hourly temperature, rain and snow(extrapolated from P1), (b) hourly open and forest 2-m wind speed (extrapolated from P1), (c) total hourly snow surface energy balance in the open (stacked bars), (d) total hourly snow surface energy balance under forest (stacked bars), and (e) hourly open and forest melt.

4.3. Antecedent Water Fluxes

[35] The components of Fn48 by SY for the control scenario show that antecedent rainfall occurs for most peak discharge events and the contribution of rainfall to Fn48 can on occasion exceed 50%; on four occasions rainfall constitutes the only source of 48-hour antecedent moisture input (Figure S3). Nevertheless, when inputs are considered separately Qp is more strongly correlated with snowmelt (KRC of 0.48 for the control) than rainfall (KRC of 0.33 for the control), and melt flux is on average three to four times higher than rainfall (for all scenarios). Therefore, it is the snowmelt process that generally governs the magnitude and timing of annual peak discharge in nival regimes [i.e., Kattelmann, 1991; MacDonald and Hoffman, 1995; Loukas et al., 2002; Troendle et al., 2001].

[36] Considering that annual peak discharge for any particular treatment scenario for any given SY need not occur on the same date or time as that of the control scenario, the change in Fn48 following harvesting can include changes not only in antecedent snowmelt, but also rainfall. In certain cases, rainfall-only events in the control scenario (which occur in the fall) can change to pre-dominantly snowmelt events following harvesting (by shifting to an earlier time in the year; (i.e., SY08); and vice-versa (i.e., SY84) (Figure S3). Generally, forest harvesting tends to shift the occurrence of Qp to earlier in the year (not shown). For instance, for scenario 50B the date of peak flow occurrence is shifted 3.5 days earlier, on average, but ranges from 31 days earlier to 11 days later, although there are only five occurrences in which Qp occurs later following harvesting for this scenario. Scenario 50B results in, on average, a 16% (approximately 6 mm) increase in 48-hour antecedent snowmelt and a 32% (approximately 3 mm) decrease in 48-hour antecedent rainfall. Hence, following harvesting, peak flow events which occur earlier in the year are increasingly dominated by snowmelt. Losses due to ET during the 48-hour period prior to Qp for both the control and treatment scenarios are negligible, although ET losses are reduced following harvesting. In the case of scenario 50B, Fn48 is increased by 10% (5 mm) on average.

[37] The relationship between Qp and Fn48 is illustrated in Figure 8, which compares the control to scenarios 50B, 50T and 100. Local polynomial regression [Cleveland and Loader, 1996] is used to identify trends in the data. The increase in Qp that occurs for scenarios 50B and 100 is not only related to increased Fn48, but the steeper trend line following harvesting suggests increased runoff efficiency, where runoff efficiency is defined as the ratio of Qp to Fn48 (Figures 8a and 8e). Greater runoff efficiency for these two scenarios appears related to the fact that the entire channel network is confined to the bottom elevation band such that subsurface flow distances are considerably shorter than those in the top elevation band (Figure S4). On the other hand, when harvesting occurs above H54, such as for scenario 50T, there is very little increase in basin-wide runoff efficiency and the general increase in Qp is related strictly to an increase in input fluxes (Figure 8c).

Figure 8.

Summary of peak annual discharge response to forest harvesting showing of 48-h antecedent input flux (Fn48) versus peak annual discharge (Qp) comparing control to scenarios (a) 50B, (c) 50T, and (e) 100; and scatterplots of the relative change in Fn48 versus the relative change in Qp for scenarios (b) 50B, (d) 50T, and (f) 100. Local polynomial regression lines are plotted to indicate general trends.

[38] Figure 8b, 8d, and 8f explores the direct relationship between relative changes in moisture input inline image and ΔQp; again, general trends are indicated with local polynomial regression. The correlation between ΔQp and ΔFn48 is positive, but the relationship appears non-linear for scenarios 50B and 50T (although the relationship is arguably linear for ΔFn48 < 0.4). The scatter about the estimated trends in Figure 8 is attributed to antecedent effects beyond 48 h. In particular, the non-linear behavior in the response of ΔQp to ΔFn48 for 50B and 50T is influenced by a few events of ΔFn48 > 0.4 that produce only moderate ΔQp. In comparing the response of ΔQp for scenarios 50B to 50T (i.e., same AH), it is again evident that for a given change in Fn48 peak flow has a more sensitive response when harvesting occurs in the bottom elevation band.

[39] The tendency for ΔQT to increase with decreasing ZH (Table 3) does not agree with the general concept that harvest effects tend to increase with increasing harvest elevation in snow-dominated catchments, which is predicated on the hypothesis that peak annual discharge is generated when the snowline has retreated above the H60 to H50 elevations [Garstka et al., 1958; Gluns, 2001; Schnorbus and Alila, 2004a]. However, these studies dealt with watersheds that had much larger relief than those in UPC and measurable climate gradients that affected snow accumulation and caused a distinct snowline recession. In UPC climate gradients are undetectable, relief is moderate, and snowpack accumulation and depletion is relatively spatially uniform. As such, the “elevation effect” appears to be due primarily to the discontinuity in drainage density that exists around H54 (Figure 1). Hence, spatial variability in harvest response appears to be dominated by spatial variability in drainage density. The DHSVM UPC application is parameterized such that surface runoff, either from infiltration excess or soil saturation, is a rare occurrence (which is considered reasonable given a lack of any physical evidence of significant surface runoff). Consequently, runoff through the soil is predominantly simulated as lateral sub-surface flow, rendering runoff and discharge sensitive to hillslope length and drainage density. We acknowledge that the occurrence of surface runoff or subsurface runoff via some other mechanism (e.g., macropore flow) would make runoff more efficient and cause the basin to be less sensitive (or insensitive) to drainage density. Nevertheless, the current DHSVM application has been shown to effectively represent the spatial and temporal dynamics of groundwater in UPC catchments, hence the assumption of runoff as predominantly via sub-surface matrix flow seems valid [Kuraś et al, 2011].

[40] The sensitivity of forest harvesting to variation in drainage density between the top and bottom elevation bands is simulated in the absence of roads. Forest roads, however, are known to increase basin drainage density and affect hillslope routing, potentially altering runoff efficiency and synchronization [Jones and Grant, 1996; Wemple et al, 1996; Bowling and Lettenmaier, 2001; La Marche and Lettemaier, 2001]. Forest roads have also been identified as possible mechanism for diverting water from small drainages [Kuraś et al., 2012]. It is reasonable to presume that the addition of roads, as would be the case in an operational setting, may potentially alter the sensitivity of the basin to forest removal. This would likely be particularly evident if roads are added to the top elevation band, where there currently is no definable surface drainage network.

4.4. Single Event Case Study: Temporal and Spatial Changes in Water Input

[41] The peak discharge response to changes in input moisture fluxes, specifically snowmelt, and its dependence upon elevation in 240 Creek, is investigated in greater detail using an event-based examination of the period leading up to and including the peak discharge event of SY43. This event was selected as it highlights the complexity of the peak flow response with respect to AH, ZH, and T as summarized in Table 5. The area-average daily melt rates for the top and bottom elevation bands and basin-wide for the control, 50B and 50T scenarios during May of SY43 are compared in Figure 9. Melt rates in Figure 9 are calculated from the difference in SWE over the preceding 24 h period. Hourly discharge for each scenario and hourly precipitation are shown in the bottom panel of Figure 9. For the control scenario the melt rates from both elevation bands are roughly equivalent and synchronized (Figure 9a) and both peak on 23 May, which coincides with the occurrence of peak annual discharge. In general, the shape of the hydrograph correlates closely to the shape of the melt curves, and the occurrence of precipitation appears to be of secondary importance. From 10 to 19 May, discharge tends to lag melt by 27 h (rxy = 0.57), although high cross-correlations also exist at lags of 6 and 51 h. During the period 20 to 24 May runoff efficiency is increased and maximum cross-correlation shifts to a lag of only 5 h (rxy = 0.58), although high cross-correlation still occurs at a lag of 27 h (rxy = 0.57).

Figure 9.

Area-average daily melt rate in bottom elevation band (below H54), in top elevation band (above H54), and basinwide for (a) the control, (b) scenario 50B, (c) scenario 50T, and (d) hourly discharge and precipitation for 30 April to 31 May SY43.

Table 5. Peak Discharge Response by Treatment Scenario for SY43
ScenarioQp (m3/s)aΔQp (%)T (years)b inline image (m3/s)cΔQT (%)
  1. a

    Simulated annual maximum.

  2. b

    Estimated by taking QT = Qp and inverting fitted GEV function for respective scenario.

  3. c

    Estimate of control QT using Control GEV function and given T.

Control1.13 1.7  
10B1.26122.11.242
10T1.07−51.51.052
20B1.28222.61.225
20T1.04−81.30.995
30B1.50332.81.3412
30T1.12−11.41.048
40B1.64453.01.3819
40T1.12−11.41.0111
50B1.76563.31.4125
50T1.1841.41.0315
1001.81602.21.2545

[42] For scenario 50B increased melt from the bottom elevation band begins on 16 May, although melt rates from both elevation bands remain synchronized and the basin-average melt rate and hydrograph still peak on 23 May (Figure 9b). The substantial ΔQp of 56% for scenario 50B (Table 5) is, however, more reflective of the relative increase in daily melt rate generated in the bottom elevation band (60%) than that which occurs for the basin as a whole (30%). Due to melt synchronization and high runoff efficiency during late May the magnitude of ΔQp in SY43 for scenarios 10B, 20B, 30B, and 40B correlates positively with AH (Table 5). In a similar study of the Redfish Creek watershed in the south-central interior of BC, Whitaker et al. [2002] and Schnorbus and Alila [2004a] also found that the sensitivity of the peak flow response depends upon the degree of snowmelt runoff synchronization between elevation bands following forest harvesting. Using empirical evidence, Jones and Perkins [2010] also identified the occurrence of snowmelt synchronization following forest harvesting as a mechanism in increasing the magnitude of rain-on-snow peak flow events in large basins in the Oregon Cascades.

[43] Possibly due to spatial variation in forest structure and topographic aspect (which affects solar radiation), forest harvesting in the top elevation band (scenario 50T) rapidly advances area-average melt much earlier than occurred with scenario 50B; melt is increased quite early on 6 May (Figure 9c) and the accelerated melt shifts the Qp from 23 to 16 May. During this time runoff routing is still relatively inefficient and the increased discharge tends to lag melt by 1 to 2 days. Desynchronization of melt between the two elevation bands reduces basin-wide snow melt by 23 May. Nevertheless, the hydrograph peak that occurs on 23 May, which now reflects melt mostly from the bottom elevation band, is of similar magnitude as the control scenario. As a result, when Qp control scenario (23 May) and Qp 50T scenario (16 May) are compared, ΔQp is only 4% (Table 5), although the discharge peak on 16 May is increased by 44%. This tendency for desynchronization of runoff results in negative ΔQp for AH ≤ 40%, despite increased water input fluxes (Table 5).

4.5. Chronology-Based Versus Frequency-Based Discharge Comparison

[44] In snow-dominated watersheds such as those at UPC, forest harvesting alters the hydrological processes that generate peak discharge. This not only changes peak discharge magnitude, but also the inextricably linked frequency of the event within the new AMS (e.g., Table 5). In effect, control and treatment peak annual discharge events that define a certain return period may derive from completely dissimilar antecedent conditions that occur within different years. A comparison of peak flow magnitudes based on chronological pairing, as conducted in traditional paired watershed studies, does not account for this nonlinear and inverse relation between the magnitude and frequency of floods, i.e., the flood frequency distribution [Alila et al., 2009]. For example, when treatment discharge events for SY43 are compared to the control discharge event for SY43 ( inline image), ΔQp ranges from 12% to 60% for harvesting in the bottom elevation band, and −5% to 4% when harvesting occurs in the top elevation band (Table 5). However, following harvesting, both the magnitude and return period of the treatment events change. Treatment events of SY43 tend to become less frequent (T increases) when harvesting occurs in the bottom elevation band (the return period almost doubles between the control and 50B scenarios), but become more frequent (T decreases) when harvesting occurs in the top elevation band (Table 5). Hence, when treatment events for SY43 are now compared to control events on the basis of equal return period (where inline image now varies from 0.99 to 1.41 m3/s) relative peak flow changes now become consistently positive and range between 2 and 45% across all scenarios.

[45] Although there are some examples to the contrary [e.g., Jones and Grant, 1996], it would seem that the consensus of past paired-basin research is that the relative increase in peak discharge decreases with increasing event magnitude [Thomas and Megahan, 1998; Beschta et al., 2000; Troendle et al., 2001; Calder, 2005; Moore and Wondzell, 2005; Bathurst et al., 2011a, 2011b; Birkinshaw et al., 2010]. However, all these studies assess peak flow impacts as a function of event magnitude by comparing the relative change in chronologically-paired events, i.e., those events that either occurs in the same year (for annual maximum events) or those that occur in response to the same storm. For example, Figure 10a shows both the scatter plot and a linear regression of treatment (scenario 50B) on chronologically-matched control peak discharge (as, for instance, would be the case in a typical ANCOVA analysis). The postharvest regression equation (with 95% confidence bounds on the parameters) is: inline image. The regression intercept parameter is not significant (but, admittedly, would be at a significance level slightly weaker than 0.05), and the slope parameter is not significantly different from 1.0. This suggests that in this example the post-harvest regression is not significantly different from the pre-harvest regression (i.e., harvesting does not have a significant effect on peak flow). Note that for this experiment, the pre-harvest regression is simply the 1:1 line.

Figure 10.

Plot showing the change in scenario 50B peak annual discharge versus control discharge (bottom axis in all panels) and control return period (top axis in all panels). (a) Chronology-paired changes that compare events that occur in the same year (Qp; treatment return period not equal to control return period) and showing fitted linear regression (LR) with 90% prediction uncertainty, (b) frequency-paired comparison of events of equivalent frequency (QT; treatment return period equal to control return period) using an empirical approximation (equivalent ranks) and showing fitted GEV function with 90% prediction uncertainty, and (c) comparison of relative changes in ΔQT and ΔQp, shown both as individual points and as the difference between respective fitted GEV and LR functions, including prediction uncertainty.

[46] When control and treatment peak annual discharge is instead compared on the basis of event rank (which is used as a surrogate for frequency), as shown for scenario 50B in Figure 10b (see also Figure 6), results indicate that the post-harvest relationship is now significantly different from the pre-harvest 1:1 reference line (i.e., harvesting does have a significant effect on peak flow). The GEV comparison also shows that ΔQT increases from 20 to 40% with increasing T (Figure 10b). As well, individual empirical ΔQT values exhibits substantially less scatter than the ΔQp values, and ΔQT is consistently positive while ΔQp often has negative occurrences (Figure 10c); i.e., results for ΔQT are not only different but also considerably less ambiguous than those for ΔQp. As the purpose of this analysis is to expressly examine the impact of forest harvesting on the practically more important peak flow frequency and magnitude, only the frequency paired methodology (i.e., Figure 10b and ΔQT curve in Figure 10c) provides the relevant result [Alila et al., 2009, 2010].

5. Conclusion

[47] The DHSVM in conjunction with synthetic weather data was used to investigate the immediate impact of forest harvesting upon the peak annual discharge flow regime for 240 Creek, a headwater stream located in south-central BC. Statistical properties of the GEV distributions used to describe the response of eleven hypothetical harvest scenarios and a control scenario indicate that both harvest area (as a proportion of basin area) and harvest elevation (given as either above or below the median elevation) affect the annual peak flow frequency curve. Comparison of peak flow quantiles for return periods (T) ranging from 1.003 to 100 years demonstrate the added dimension of event frequency when considering peak discharge impacts. It was shown herein that for all scenarios, when annual peak discharge between the control and a treatment watershed are compared on the basis of equal frequency, a relative increase in peak annual discharge occurs consistently across the full range of return periods. Both large and small annual peak flow events are increased following harvesting. From a frequency perspective, events of a given magnitude are therefore expected to become more frequent following harvesting and, inconsistent with the prevalent hydrological wisdom, the larger the event the greater its increase in frequency. Depending upon the statistical test used (i.e., comparing distributions or comparing individual quantiles), peak flow impacts do not become statistically significant until harvest area exceeds 20% or 30% of the original forested area.

[48] The peak flow regime of 240 Creek is dominated by snowmelt. As such, forest harvesting will affect peak discharge in general via changes in the energy available at the snowpack surface, particularly net radiation, which in turn will affect melt and runoff rates, and runoff synchronization. Changes in antecedent rainfall are not a dominant influence. Consequently, in both elevation bands post-disturbance peak flows become progressively larger in magnitude with increasing harvest area. The dependence of peak flow response to elevation appears to be an artifact of the channel morphology of the watershed rather than to climatic gradients affecting snowpack and energy supply. Clear-cutting in the bottom elevation band tends to produce the largest peak flow response due to high drainage density; flow lengths are short and peak flow impacts are in direct response to both increased melt rates and increased runoff efficiency. Harvesting in the top elevation band affects peak discharge magnitude generally via increased melt runoff that augments runoff generated in the bottom elevation band.

[49] The results of this work and that of Schnorbus and Alila [2004a] and Kuraś et al [2012] suggest that the response of the peak flow regime to forest harvesting is sensitive to even small differences in basin physiography. Hence, future work could attempt to assess the functional relationship between ΔQT and such physical basin properties as elevation, topographic exposure (slope/aspect), spatial distribution of forest type and disturbance, and drainage density [e.g., Green and Alila, 2012].

Acknowledgments

[50] This work was funded by Forest Innovation Investment grant R02-15 and NSERC grant RGPIN 194388-11 to Y. Alila. Forest Renewal BC and the BC Forest Service funded collection and analysis of the UPC data. The authors acknowledge the ongoing commitment of the Okanagan-Shuswap Forest District and Weyerhaeuser Company Ltd. to research at UPC. The authors thank Rita Winkler and David Spittlehouse for their feedback on an earlier draft of this manuscript. The authors are also grateful to the constructive comments provided by three anonymous reviewers.

Ancillary

Advertisement