2.1. Study Area
The study area included a wide range of vegetation conditions and management and disturbance histories in southwest Oregon, USA, that spanned six of Cleland et al.’s [21
] ecoregion sections (Figure 2
). The area was defined as most of the forested lands in six counties (Coos, Curry, Douglas, Jackson, Josephine, and Klamath) between 42.0° and 43.9° N latitude and 120.9° and 124.6° W longitude. Elevation ranges from 0 to 2800 m (933 ± 550 mean and standard deviation) and annual precipitation from 29 to 367 cm (130 ± 58 mean and standard deviation, based on Daymet) [22
] (Supplemental Figure S1
). Climax vegetation series range from western hemlock (Tsuga heterophylla
) and tanoak (Lithocarpus densiflorus
) in the Coast Ranges, to Douglas fir (Pseudotsuga menziesii
) and Oregon white oak (Quercus garryana
) in the interior valleys of the Klamath Mountains, to mountain hemlock (Tsuga mertensiana
) in the higher elevations of the Western Cascades, to Ponderosa pine (Pinus ponderosa
) and western juniper (Juniperus occidentalis
) in the Eastern Cascades and Modoc Plateau [23
]. The total land area is 4.9 million ha, of which 84% is forested (based on NFI) [24
]. Of the forested lands, 41% is managed by the National Forest Systems (NFS), 17% by the Bureau of Land Management (BLM), 38% by private owners, and the remaining 4% by other federal, state, and local agencies (Supplemental Figure S2
). Since the mid-1990s, federal management has emphasized protection and restoration of forest structural diversity through partial cutting, while private owners have tended to harvest more frequently and intensively [25
]. Nine percent of the forestland in the area is in a protected status (e.g., wilderness, national park, or national monument). Forest fires are not uncommon; one of the largest in Oregon state recorded history occurred in the study area in 2002 (the Biscuit Fire, which covered 200,000 ha).
2.2. Plot Data
The ground-measured data used in this study were collected under two separate forest inventories: (1) the United States Department of Agriculture (USDA) Forest Service’s Forest Iinventory and Analysis (FIA) periodic inventory [26
], and (2) the Pacific Northwest Region of the USDA Forest Service’s Current Vegetation Survey (CVS) inventory [29
]. Both inventories used a probability-based sample design; the FIA sample consisted of a randomized systematic grid at a 5.47 km spacing on lands that were not managed by NFS or western Oregon BLM, while the CVS sample consisted of a systematic square grid at a 5.47 km spacing across NFS lands, and a denser grid at a 2.74 km spacing outside of designated Wilderness areas, providing a sample density of one plot per 3000 and 750 ha, respectively. A single, nationally consistent forest inventory approach, referred to as “annual FIA,” was instituted across all ownerships in Oregon starting in 2001 where the grid consisted of randomly selected, previously measured plots (where available) within 2400 ha hexagons [24
]. Although western Oregon BLM lands also had a CVS-style inventory, remeasurement data were not available at the time this study was initiated, and their plot sample was not included in the selection criteria for annual FIA plots. Nevertheless, we expect that the selected inventories capture a broad range in vegetation conditions, disturbance types, and carbon trajectories to be useful for a proof-of-concept for this landscape. The FIA periodic inventory measured plots that qualified as forested (i.e., land areas ≥0.4 ha and >36 m wide that support or previously supported ≥10% tree canopy cover and were not primarily managed for a nonforest use). FIA plots were measured in the 1980s (1985–1988) and remeasured in the 1990s (1995–1999), for a mean remeasurement interval of 10 years. CVS plots were installed in 1993–1997 and remeasured 1997–2007, with a remeasurement period that ranged from 1–14 years with a mean of 7.1 years.
Plots from either inventory that were forested and measured twice, and were on the same grid intensity (one plot per 3000 ha, to ensure a balanced representation of vegetation conditions), were selected for analysis. In addition, only plots that had also been measured in annual FIA through 2007 were used, as these had been manually checked for accuracy of plot-based land-use and disturbance information for related studies. Furthermore, plot locations were never adjusted, so some plots straddled boundaries between different stands or land-use types. While this is not a problem for inventory estimation, it can be a problem for constructing models using imagery because different stands can have different reflectance characteristics and experience disturbance at different times. As a result, only plots where at least four-fifths of the plot area was in a single forested condition were used as a compromise between stronger signal fidelity and a reduced, biased sample size. Finally, 9% of the remaining CVS plots were excluded because the remeasurement occurred less than 5 years after plot installation, which was deemed too short a period to accurately reflect change in live tree biomass and shorter than any planned future inventory remeasurement cycle. These criteria resulted in 416 CVS plots and 260 periodic FIA plots being selected (n = 676). The mean and standard deviation of the remeasurement period for this set of plots was 10.0 ± 1.4 years.
The periodic FIA plot design consisted of five variable-radius points, while the CVS design consisted of nested fixed-radius subplots of different sizes around five points. This analysis focused on changes in the live tree forest component; calculations of tree biomass, and change components of growth, mortality, and harvest, were consistent among datasets. Estimates of aboveground live-tree woody biomass were based on regional FIA models of the sum of bole, bark, and branch biomass for each tree measurement. Bole biomass (ground to tip) was calculated from regional species-specific volume equations and wood densities [30
]. Bark and branch biomass were calculated from regional species-specific equations selected from Means et al. [31
] and documented in USDA Forest Service [32
]. An expansion factor derived from the fixed-area plot size or the tree’s diameter and variable-radius prism factor (depending on tree selection method) was used to convert individual tree biomass to an area basis (Mg ha−1
). We summed bole, bark, and branch estimates on each plot and multiplied by 0.5 for an estimate of above-ground live woody carbon (AWC). Net change in AWC between the first and second measurements (time 1 and time 2, respectively) was calculated by subtraction and referred to as ΔAWC. Changes in size and status of individual trees tracked on the plots was compiled into components of growth, mortality, and harvest. Additional information on data compilation can be found in Gray and Whittier [33
] and Gray et al. [34
]. We also included a few commonly available calculated variables from inventory plots as potential predictor variables (Table 1
). The annual growth rate for each live tree at time 1 was available from prior measurements, increment cores, or modeling [35
], from which an estimated increase in stand AWC due to growth was calculated (GROWMODEL). Stand age was estimated in the field as the average age of overstory dominant trees, based on increment cores. We estimated site productivity in terms of production of wood at culmination of mean annual increment (MAI, m3
) from measurements of site index trees (i.e., height and age) on each plot [36
Disturbance codes were collected in the field on most plots or gleaned from written descriptions provided by field crews (disturbance coding was more complete and comprehensive during annual FIA measurements, which included coding prior disturbances that may have been described during the periodic or CVS inventories). Crews noted the type and timing of harvest or fire that occurred and also noted damage or mortality to live trees from animals, insects, disease, or weather. Plot-recorded disturbances were grouped for analysis into harvest, fire, fire and harvest, and other (including plots where crews coded that cutting occurred in the stand but none of the measured trees on the plot were cut). Plots that experienced both fire and cutting were subsequently grouped with harvest. Plant species composition observed in the field was used to classify plots to plant association using regional Forest Service guides [37
]; we grouped these into plant association zones (PAZs) for analysis. Although most PAZs are named after conifer species, 15% of the plots in this study were dominated by hardwood tree species. Basic plot attributes summarized by PAZ are shown in Supplemental Table S1
2.3. Remote-Sensed Data
The satellite data used to generate predictor variables of AWC change were created from annual Landsat time series of radiometrically normalized Landsat Thematic Mapper and Enhanced Thematic Mapper Plus images from 1984–2007 following method in Kennedy et al. 2010 [40
]. The radiometrically normalized time series were then visually inspected at each inventory plot location using the TimeSync (TS) tool [12
] to manually segment the time series between vertices of change in slope and classify segment types as disturbance, recovery, or stability, and, for segments that were disturbed, identify disturbance agents.
To model change in AWC (ΔAWC) with the aid of TS, we calculated the tasseled cap angle (TCA) from TC greenness and brightness described in Powell et al. [41
] as TCA = arctan(TCG/TCB), and TC distance (TCD) as first described in Duane et al. [42
], where TCD = (TCG2
. In a similar study, Pflugmacher et al. [43
] found TCA to be related to vegetation cover, useful in characterizing change in forest biomass, and TCD to be related to forest structure and composition, useful in characterizing current conditions. For each plot evaluated, we temporally smoothed the spectral data between segment vertices by fitting an ordinary least-squares linear regression on spectral values as a function of time. Starting with the first, segments were fitted in successive order to assure continuity (e.g., the fitted value from the end vertex of the first segment was used to fit the second segment). For abrupt disturbance segments, the fitted end vertex from the previous segment was used for that segment’s start vertex instead of the observed spectral value. Fitting was done independently for TCA and TCD.
Finally, for each field plot, we extracted the spectral values and time-series metrics associated with the 30 × 30 m Landsat pixel that contained the plot center. Because a 3 by 3 mean filter had been applied to all spatial data, the effective ground area sampled was 90 × 90 m around the plot center. The result was a series of segments for each 90 m area with annual reflectance values interpolated between the segment ends (vertices). Some of the change segments were attributed with the disturbances coded as delay, mechanical other, other non-disturbance, and stress and tended to be of low magnitude; these were grouped for analysis as “Other”.
TS segments, or portions thereof, were extracted for analysis based on the dates of plot measurement. For stable and recovery segments, the segments were truncated if necessary to coincide with the plot measurement years. For disturbance segments that coincided with the year of plot measurement, the plot disturbance codes and crew field notes were examined to determine whether to include the disturbance in the time series for the plot or not.
Predictor variables were calculated from the TS data to characterize the magnitude and duration of disturbance or recovery processes, based on those used by Pflugmacher et al. [44
]. Disturbance-related variables were keyed to the greatest disturbance in the time series (in this case, between the plot measurement dates), determined as the disturbance segment with the greatest change in TCA values (Table 1
). Other variables describe TCA and TCD values and trends prior to and following the greatest disturbance. Time series without disturbance were described by the current condition, current trend, and last monotonic trend variables.
2.4. Statistical Analyses
We compared the classification of disturbance types between the plot data and the TS analysis. Agreement rates of different groupings were calculated. The means and distributions of plot and TS change variables were generated to investigate the characteristics of discrepancies between the methods.
We used multiple linear regression analyses to explore the relationships between AWC at time 2 and ΔAWC between measurements (responses), and the plot- and TS-derived variables (predictors) as:
is the estimated plot-level AWC (or ΔAWC) for each plot j
is the value for each independent variable i
for each plot j
, and a
are coefficients estimated by the regression model.
Two types of analyses were done: (1) using only remotely sensed variables and (2) using remotely sensed variables and plot measurements at time 1, including net tree growth rates (net growth is growth of live trees minus mortality). The tree productivity variable MAI, although based on plot measurements, was also included in the first model because spatial models of MAI were available in the region [45
]. MAI integrates climate, topography, and soils in a variable directly meaningful to live tree carbon flux, making those additional variables superfluous.
Prior to building regression models, we removed variables with similar information as determined by a correlation coefficient >0.5 with other predictor variables. We used stepwise model selection with a minimum p
-value of 0.15 on each type of analysis. To avoid over-fitting, and because most of the models still had large numbers of variables (n
> 10), we selected the model with the lowest value for Schwarz’ Bayesian information criterion [46
]. Separate change-analysis models were run for TS-disturbed and -undisturbed plots since the relevant variable sets were different, and predicted values were combined to calculate overall model accuracy.