Study Methods

 


H. J. Andrews Experimental Forest, OR

 


Cutting samples from a log

 



Cross-section from 7 year old Tsuga

 



Measuring outer diameter

 



Removing bark from the sample

 

 


Measuring longitudinal thickness of bark

 

 


Cutting out a sub-sample for moisture content

 

General methods in determining decomposition rates

The general approach used at each of the sites differed (Table 2).  The three methods used, chronosequence, time series, and decomposition vectors are defined below.

Chronosequence Approach. The chronosequence approach was used at Fraser, Frissell ridge, Klamath, and Sequoia to determine decomposition rates from logs in which the sampling occurred only once.  In a chronosequence one ages many pieces in various states of decay and examines how a parameter such as density changes through time.  This is a substitution of space for time, and while not as precise as a time series experiment (where individual pieces are followed through time) it provides a good first approximation of decomposition rates (Harmon et al. 1998). 

Time Series. In a time series a cohort of fresh logs is placed out and individual pieces are followed through time (Harmon et al .1998). Typically a log is sampled once, after a planned interval expires.  As different logs are sampled over time the trend in decomposition is revealed.  It is more accurate than the chronosequence approach, but takes a great deal more time to examine the full range of changes.  This method was used at H. J. Andrews, Pringle Falls, and Warm Springs.   

Decomposition Vectors.  The final method used to study decomposition rates is the decomposition vector method (Harmon et al 2001).  It represents, in some ways, a combination of chronosequence and time series.  It involves sampling a series of logs based on age (from a chronosequence) or decay class, and then resampling the same logs after a set period of years.  This allows one to more accurately determine changes in both density and volume over the full range of decay conditions. We used this method at Cascade Head, Fraser, Sequoia, and Wind River.

Time Series Method

The major time series study was conducted at HJAF.  The methods used to determine log characteristics in our analysis are described in detail by Harmon (1992) and Harmon et al (1995).  In September 1985, live, healthy trees of four conifer species (Abies amabilis, Pseudotsuga menziesii, Thuja plicata, and Tsuga heterophylla) and one hardwood (Alnus rubra) were felled, cut into long logs, and placed on the forest floor at seven study sites.  The conifer logs used in the study averaged 5.5 m in length, 52 cm in diameter, while the Alnus logs were 2.5 m in length, 25 cm in diameter.  To characterize initial density and volume of each tissue (i.e., outer bark, inner bark, sapwood, and heartwood) in a log, a cross-section 8 to 10-cm thick was removed from the end of each log.  The cross-sections were photographed and these were digitized to estimate the volume of outer bark, inner bark, sapwood, and heartwood.  Wood density in the case of conifers was determined by cutting regular shaped blocks on a table saw, measuring external dimensions, and dry weight. For Alnus, pie-shaped slices were removed, and the volume determined using the formula for a sector after measuring the thickness, radial length and angle of the pie slice (i.e., sector).   Inner bark density was determined by measuring the dimensions of bark peeled from the cross-sections.  Outer bark densities were determined using water displacement volumes.  All densities were calculated as oven dry weight (7 days at 55 C) divided by green volume.  Logs were resampled at annual to 4 year periods, with 5 to 6 cross-sections removed from each of 6 logs in the case of conifers and 3 logs in the case of Alnus.  At the resampling the initial methods were repeated to determine the change in density and total log volume.     

The other times series examined at PFEF and WSR were less intensive than the HJAF studies. Freshly fallen trees were cut into logs, typically 2 to 5 m long and cross-sections removed from each end.  For each cross-section the diameter, mean longitudinal thickness, circumference covered by bark, radial thickness of bark and mean radial depth of decay were recorded.  The total mass of the bark and wood for each cross-section was weighed on a portable electronic scale with a range of 1 to 6000 g (Ohaus Model CT6000), and then 50-200 g subsamples were removed to determine the moisture content.  These subsamples were weighed and then dried in an oven at 55 C until their mass was stable (usually 7 to 14 days).  Total dry weight was calculated as the fraction of dry material in the subsamples multiplied by the total wet weight.  The volume of the bark or wood was determined from the dimensional data recorded.  Density of bark or wood was calculated as the dry weight divided by the volume of the cross-section.  The overall cross-sectional density was calculated similarly, but using the total dry weight and volume.  The weighted mean density for each log was calculated.  Logs were resampled at 2 to 10 year intervals with 3 cross-sections removed and the methods for volume, mass, and density repeated.   

Chronosequence and Decomposition Vector Methods 

For logs sampled using the chronosequence and/or decomposition vector methods individual logs were selected to represent as many ages or decay stages as possible.  For decay classes we sought to have at least 3 replicates of each decay class.  For chronosequences we also sought to have at least 3 replicates of each decay class, although to represent as even a distribution in ages as possible, some decay classes had more replicates.  Logs were created from a range of processes, although for chronosequences we avoided the use of logs generated from snag fall.  In the case of logs left as a result of harvest or thinning, we examined the state of decay to make sure it roughly corresponded to the time left on the ground.  Pieces that were either too decayed or too sound were not used as they might have been added at a time other than the activity date noted in the records.  Trees adjacent to logs that resulted from natural mortality were examined for the presence of scars that may have been created when the log fell.  In the case of logs in very advanced stages of decomposition it was often impossible to find a tree fall scar or any other means to date the time of death.  Priority was given to very decayed logs with the possibility of dating the time of death, however, some logs were selected without dates.  To estimate the initial density of dead trees, we sampled a limited number of freshly killed, standing dead trees.  For the most part these trees still had needles attached and it was likely that they died in the past year.   

Once logs had been selected for sampling, they were marked with an aluminum tag and their location was plotted on a rough sketch map.  This allowed us to resample many of these logs and apply the decomposition vector method to narrow uncertainties about decomposition rates.   

Dates of tree death or cutting were taken from fall scars and or living stumps (Harmon et al. 1986).  In the case of logs added by harvest or thinning, we corroborated these dates by aging living stumps within the unit sampled.  Fall scars and the upper portions of living stumps were removed with a chainsaw and air-dried.  These cut surfaces were then sanded with fine grit sandpaper and examined under a binocular microscope at 10-X magnification.  Rings formed after the scars were counted twice and if the counts differed were counted a third time.   

For each log sampled we noted key characteristics that could be used to define a decay class system.  Physical characteristics recorded included: the presence of leaves, twigs, branches, bark cover on branches and boles, sloughing of wood, collapsing and spreading of the log (indicating the transition from round to elliptic form), degree of soil contact, friability or crushability of wood, color of wood, and if the branch stubs could be moved (Harmon et al 1987).  Biological indicators such as moss cover, fungal fruiting bodies, or presence of insect galleries were also noted although past experience shows they are too variable at a local scale to be useful in separating decay classes (Harmon et al 1989).   

To determine rates of decomposition we determined the remaining mass of each log.  Current mass remaining was determined from the current density and volume of the log.  The current volume was determined by measuring the length and diameter systematically along the bole from the base to the top as well as the locations where cross-sections were removed.  For each log segment defined by the end diameters or the cross-sections, we calculated the volume as a frustum of a cone (Harmon et al 1989).   

The current density of the wood and bark was determined by removing cross-sections along the length of each log (Figure 3).  For sound trees we removed 3 to 4 cross-sections.  In the case of extremely decomposed or very short logs, only 2 cross-sections were removed.  For each cross-section the diameter, mean longitudinal thickness, circumference covered by bark, radial thickness of bark and mean radial depth of decay were recorded.  Diameters were usually measured using a diameter tape, however, in the case of extremely old logs the cross-sections were too decomposed to be measured with this method.  We therefore excavated very decayed cross-sections and used a ruler to measure the long and short axes of the cross-section from the remaining portions of the log.  The total mass of the bark and wood for each cross-section was weighed on a portable electronic scale with a range of 1 to 6000 g (Ohaus Model CT6000), and then 50-200 g subsamples were removed to determine the moisture content.  When cross-sections had a range of decay or moisture conditions, we took samples from the different areas in rough proportion to the area in each condition.  These samples were weighed and then dried in an oven at 55 C until their mass was stable (usually 7 to 14 days).  Total dry weight was calculated as the fraction of dry material in the subsamples multiplied by the total wet weight.  The volume of the bark or wood was determined from the dimensional data recorded.  Current density of bark or wood was calculated as the dry weight divided by the volume of the cross-section.  The overall cross-sectional density was calculated similarly, but using the total dry weight and volume.  The weighted mean density for each log was calculated.  The densities from each cross-section in a log were weighted by their cross-sectional area, so that the smaller, upper cross-sections contributed less to the log average than the larger, lower cross-sections.

Statistical Analysis 

The primary statistical analysis used for the chronosequence and time series data was linear regression.  To calculate the decomposition rate constant, the estimated age of the logs was used as the independent variable and the density of the log (adjusted for fragmentation losses) was used as dependent variable.  The form of the regression equation used to calculate the decomposition rate constant was: 

ln (Densityt)= ln (Density0) -k t, 

where ln is the natural logarithm of Densityt or Density0, the density at time t and 0, respectively and k which is the decomposition rate constant.  

In the case of time series that did not have enough sample times for regression analysis (N<3) we used a modification of the regression equation:  

-k = (ln [Densityt/Densityt-i])/i 

where Densityt or Densityt-i, the density at time t and the initial sampling time t-i, respectively and i is the interval between samples.  The values of k calculated were then averaged to estimate the overall species rate of decomposition (Figures 4 and 5).   

For logs sampled using the decomposition vector method, we also averaged the decomposition rates for the decay classes.  However, rather than use a simple arithmetic average, we weighted the value of each decay class according to the length of time logs spend in each decay class. While we did not have exact information on the duration of each decay class for each species, the residence time in decay class 2 is typically twice that of decay class 1.  The same geometric relationship generally holds for class 3 and 2.  Class 4 and 5 are typically have the longest residence times, and these are approximately equal. As the weightings for the decay classes are relative, we therefore divided the total life span of a log such that decay classes 1-5 accounted for 4, 9, 17, 35, and 35% of the life span; respectively.  Decomposition rates were assigned to the midpoints of each decay class’s age, a decomposition rate of 0 was assigned to year 0 and relative year 100 (when no decomposition would be expected) and rates of decomposition were interpolated between the decay class midpoints.  We then used these observed and estimated decomposition rates to estimate the remaining mass for each relative year of the logs life span.  To estimate the weighted average or integrated decomposition rate, we calculated the reciprocal of the sum of all the remaining masses.  Essentially this is equivalent of determining the decomposition rate by knowing the steady-state store of logs (Log Massss) with a steady input of 1 unit of mass of logs:

K=1/Log Massss.

While this approach is usually applied to field data (e.g., Sollins 1982), it can also be used to give the average decomposition rate from logs of different ages (Figure 6).

To examine environmental controls we calculated the temperature response as a Q10  (quotient 10) or the increase with a 10C change in temperature. This relationship is: 

kT=Q10(T-10/10)

Where kT is the decomposition rate at temperature T (C) and 10 C is the base temperature. In this formulation kT is 1 when T is 10 C.  

Continue  

Return to main page