Elsevier

Remote Sensing of Environment

Volume 154, November 2014, Pages 61-73
Remote Sensing of Environment

Improving estimates of forest disturbance by combining observations from Landsat time series with U.S. Forest Service Forest Inventory and Analysis data

https://doi.org/10.1016/j.rse.2014.08.005Get rights and content

Highlights

  • A Landsat time series (LTS) approach is used to improve FIA disturbance estimates.

  • Analyst-led LTS interpretation increased the measurement frequency of FIA plots.

  • FIA and LTS both detected upper-canopy disturbance while only FIA saw below-canopy.

  • Combining FIA and LTS observations led to a 65% gain in detected disturbance.

  • Post-stratifying with a Landsat disturbance map led to more precise estimates.

Abstract

With earth's surface temperature and human population both on the rise a new emphasis has been placed on monitoring changes to forested ecosystems the world over. In the United States the U.S. Forest Service Forest Inventory and Analysis (FIA) program monitors the forested land base with field data collected over a permanent network of sample plots. Although these plots are visited repeatedly through time there are large temporal gaps (e.g. 5–10 years) between remeasurements such that many forest canopy disturbances go undetected. In this paper we demonstrate how Landsat time series (LTS) can help improve FIA's capacity to estimate disturbance by 1.) incorporating a new, downward looking response variable which is more sensitive to picking up change and 2.) providing historical disturbance maps which can reduce the variance of design-based estimates via post-stratification. To develop the LTS response variable a trained analyst was used to manually interpret 449 forested FIA plots located in the Uinta Mountains of northern Utah, USA. This involved recording cause and timing of disturbances based on evidence gathered from a 26-year annual stack of Landsat images and an 18-year, periodically spaced set of high resolution (~ 1 m) aerial photographs (e.g. National Aerial Image Program, NAIP and Google Earth). In general, the Landsat data captured major disturbances (e.g. harvests, fires) while the air photos allowed more detailed estimates of the number of trees impacted by recent insect outbreaks. Comparing the LTS and FIA field observations, we found that overall agreement was 73%, although when only disturbed plots were considered agreement dropped to 40%. Using the non-parametric Mann–Whitney test, we compared distributions of live and disturbed tree size (height and DBH) and found that when LTS and FIA both found non-stand clearing disturbance the median disturbed tree size was significantly larger than undisturbed trees, whereas no significant difference was found on plots where only FIA detected disturbance. This suggests that LTS interpretation and FIA field crews both detect upper canopy disturbances while FIA crews alone add disturbances occurring at or below canopy level. The analysis also showed that plots with only LTS disturbance had a significantly greater median number of years since last FIA measurement (6 years) than plots with both FIA and LTS disturbances (2.5 years), indicating that LTS improved detection on plots which had not been field sampled for several years. Next, to gauge the impact of incorporating LTS disturbances into the FIA estimation process we calculated design-based estimates of disturbance (for the period 1995–2011) using three response populations 1.) LTS observations, 2.) FIA field observations, and 3.) Combination of FIA and LTS observations. The results showed that combining the FIA and LTS observations led to the largest and most precise (i.e. smallest percent standard error) estimates of disturbance. In fact, the estimate based on the combined observations (486,458 ha, +/− 47,101) was approximately 65% more than the estimate derived solely with FIA data (294,295 ha, +/− 44,242). Lastly, a Landsat forest disturbance map was developed and tested for its ability to post-stratify the design-based estimates. Based on relative efficiency (RE), we found that stratification mostly improved the estimates derived with the LTS response data. Aside from insects (RE = 1.26), the estimates of area affected by individual agents saw minimal gain, whereas the LTS and combined FIA + LTS estimates of total disturbance saw modest improvement, with REs of 1.43 and 1.50 respectively. Overall, our results successfully demonstrate two ways LTS can improve the completeness and precision of disturbance estimates derived from FIA inventory data.

Introduction

In addition to being one of the primary drivers of the net terrestrial carbon budget (Harmon, 2001, Kasischke et al., 2013), forest disturbance also plays a critical role in affecting biodiversity (Dornelas, 2010), wildlife habitat (Gibson et al., 2013) and the surface energy balance (O'Halloran et al., 2011). With rising global temperatures (Easterling et al., 2000, Intergovernmental Panel on Climate Change, 2014) and an ever growing human population (Raftery, Li, Ševčiková, Gerland, & Heilig, 2012) poised to alter the frequency and severity of disturbance regimes across the globe (Bradford et al., 2013, Dale et al., 2001, Westerling et al., 2006), improved monitoring of forest disturbance, especially at the landscape scale has become increasingly important. Because forest disturbance manifests at a variety of spatial and temporal scales (Asner, 2013) and has varying impacts which affect the canopy, understory and forest floor, effective monitoring will likely require a hybrid approach, where detailed field measurements collected in a probabilistic sample are combined with frequent, repeated observations made by remote sensing satellites.

In the United States, the Forest Service Forest Inventory and Analysis (FIA) program (http://www.fia.fs.fed.us/) collects detailed field measurements which are used to produce timely and accurate estimates of a wide range of forest attributes. These attributes, which include forest area and volume, are used to provide information on the current status and health of forests over varying geographical extents. Typically, a probabilistic sample of inventory plots is used to estimate amount of forest area while maps derived from remote sensing have been used both as strata applied to reduce variance (i.e. for “post-stratification”, see McRoberts, Wendt, & Liknes, 2005) and to delineate forest location and extent (McRoberts & Tomppo, 2007). Reporting population estimates at the state and county level has long been a primary function of FIA. However, with increasing stress on forested ecosystems (van Mantgem et al., 2009, Weed et al., 2013), a new emphasis has been placed on improving FIA's capacity to report on how much forest is changing, where it is changing and what is changing it. This includes developing a flexible monitoring system which can estimate annual trends, report on different geographic areas (e.g. administrative units, watersheds, or ecoregions), as well as incorporate different spatial layers for use in post-stratified variance reduction.

Presently, monitoring the status and trends of forest disturbance can be challenging for inventory programs as they often have long temporal gaps (e.g. several years to decades) between plot measurements. For example, in the western U.S., a 10 year cycle of measurements is required for the network of FIA plots to fully sample the entire forested landscape. Since forest disturbance and recovery dynamics vary irregularly (annual rates < 3% per year; Masek et al., 2013) over space and time, both the frequency (time) and spatial coverage (space) of observation determine how well a sample captures a rare event like disturbance (Patterson & Finco, 2011). Therefore, it would greatly benefit FIA if all of its plots could be monitored and updated annually, thus providing estimates of disturbance which better reflect the current condition of the forest landscape. This would be especially beneficial in the western states, where annual updating would give FIA better capacity to track and respond to episodic insect outbreaks and fires. Based on this need, there is potential for FIA to improve its change monitoring capabilities by incorporating more frequent observations from optical remote sensing satellites such as Landsat.

With a 40 + year historical archive and a 16 day repeat cycle, Landsat imagery offers an excellent data source for monitoring forest disturbance over large areas (Hansen & Loveland, 2012). The 30 m spatial grain and 6 reflective bands are capable of capturing many types of forest disturbance, especially those that impact the upper canopy (Cohen & Goward, 2004). Now that the entire Landsat archive is freely available it has become economically feasible to monitor disturbance over large areas using satellite time series. The increased accessibility of Landsat data has led to the emergence of several new automated algorithms which are capable of mapping historical disturbance using the spectral response from multiple image dates (Huang et al., 2010, Jin et al., 2013, Kennedy et al., 2010, Zhu et al., 2012). Thanks in part to greater access to supercomputers such as the NASA Earth Exchange (NEX, https://nex.nasa.gov/nex/), these algorithms are now being run over increasingly larger areas (Hansen et al., 2013). For example, the North American Forest Dynamics (NAFD) project (Goward et al., 2008, Masek et al., 2013) is currently using the Vegetation Change Tracker (VCT, Huang et al., 2010) algorithm to produce a wall-to-wall forest disturbance map for the conterminous U.S. With the availability of Landsat-based disturbance maps on the rise, it is important to determine their utility for improving the precision of FIA estimates through post-stratification.

In addition to stratification, Landsat data have also recently emerged as an effective backdrop for collecting plot-level reference information on disturbance and land cover change. Visualization tools such as TimeSync (Cohen, Yang, & Kennedy, 2010) demonstrate how a trained analyst can use the spectral and temporal information from Landsat along with other spatial data to record timing and cause of most natural and anthropogenic disturbance events. Such ancillary data include high spatial resolution photos from NAIP and Google Earth; fire polygons from Monitoring Trends in Burn Severity (MTBS, www.mtbs.gov; Eidenshink et al., 2007); and disturbance grids compiled by the Landfire program (www.landfire.gov; Vogelmann et al., 2011). Because of Landsat's long historical record, the analyst interpretation approach has surfaced as one of the best (and only) methods for collecting reference data over the full range (20–40 years) and interval (annual) of most time series maps. In addition to validation, analyst-based LTS observations can also be used for design-based estimation if collected over a statistically-based set of sample locations such as FIA plots. In fact, it is possible that combining the improved frequency and spatial coverage of LTS observations with the detailed, but less frequently acquired FIA field data might provide the most complete picture of where and how disturbance is impacting the greater landscape. In general, combining LTS and field-based observations in a design-based framework is an emerging technique for estimating (or predicting) landscape scale disturbance dynamics (e.g. see Plugmacher, Cohen, Kennedy, & Yang, 2014) and for validating maps of disturbance (Olofsson, Foody, Stehman, & Woodcock, 2013). Although our focus here is on FIA data, it is important to recognize that LTS observations, both in the form of maps and analyst-interpretations can be similarly applied to other sets of probabilistically collected field measurements.

Traditionally, Landsat data and its derivatives (e.g. National Land Cover Dataset, NLCD) have been used by FIA for post-stratification, a statistical technique for using spatial data to reduce the variance of design-based parameter estimates (McRoberts, Tomppo, & Nӕsset, 2010). The literature suggests that static parameters such as forest area and volume have typically gained the greatest benefit from stratification (McRoberts et al., 2006, Nelson et al., 2002), while dynamic parameters reflecting forest change (e.g. growth, mortality and removals) have had much more limited success (Brooks et al., 2013, Gormanson et al., 2003). It is possible that stratification (which is based on agreement between map and sample plots) could improve if the FIA response variable were measured annually over the full landscape and if a satellite disturbance map could be developed which accurately captures the prevailing disturbance dynamics of a region (e.g. agents and severity).

Therefore, the two main objectives of this study are to: 1.) Use LTS disturbance observations to augment FIA plots and determine how supplemental plot histories impact design-based estimates of disturbance, and 2.) Determine whether a Landsat disturbance map can improve (i.e. increase precision of) design-based estimates through post-stratified variance reduction. To address these objectives, we first use an analyst-led LTS interpretation approach (similar to TimeSync, Cohen et al., 2010) to collect supplemental disturbance history information on 449 forested FIA plots located in the Uinta Mountains of northern Utah, USA. In addition to annual Landsat imagery, a time series of high spatial resolution air photos from NAIP and Google Earth are used to expand the interpretation approach to include detailed tree-level measurements of mortality occurring below the resolution of the Landsat data. To better understand how the LTS and field observations relate to one another we use the cells of an agreement matrix to divide the plots into 3 disturbance populations (e.g. 1.) Both LTS and FIA find disturbance, 2.) Only FIA finds disturbance and 3.) Only LTS finds disturbance) which are used to compare distributions of live and disturbed tree size (i.e. height, HT and diameter at breast height, DBH) and number of years between LTS and FIA measurements using non-parametric Mann–Whitney tests. Next the FIA and LTS observations of disturbance are used individually and combined together to generate design-based estimates of disturbed area (for the period 1995–2011) for specific agent classes (e.g. insect, disease, fire, harvest, weather, mechanical, and animal) and their total. To test the potential for improving the precision of the derived estimates through post-stratification, we use a multi-stage approach to develop a Landsat forest disturbance map which better captures the subtle, low magnitude spectral changes associated with recent insect outbreaks. We evaluate the maps' capacity to increase the precision of the plot-based estimates using the measure of relative efficiency (RE = variance from simple random sampling/variance from post-stratification; see McRoberts, Holden, Nelson, Liknes, & Gormanson, 2006), which can be interpreted as the factor by which the size of a simple random sample would have to increase by to achieve the same precision as stratification. Testing the capacity of remote observations to improve disturbance estimates over a broad range of forest types and causal agents may provide a template for future integration of field and LTS observations within FIA's monitoring and reporting system.

Section snippets

Study area

The 1.6 million ha study area consists of the non-overlapping portion (or Thiessen Scene Area, TSA) of Landsat scene path 37 row 32, which covers nearly the entire Uinta Mountain range located in the northeast corner of Utah, USA (Fig. 1). Approximately 56% of the study area is forested including broad expanses of conifer and mixed deciduous forest, as well as stands of pure aspen at high elevations and sparse Pinyon-Juniper woodlands on drier, lower elevation slopes. According to the most recent

Comparing FIA field and Landsat time series disturbance observations

Over the 17 year period of study, FIA field crews observed a total of 140 discrete disturbance and treatment (i.e. harvest) events (on 126 out of 449 plots), of which the vast majority (76%) were attributed to insects and disease (Table 1). Fires were the next most prevalent agent (8%), followed by animal (5%), weather (4%), other (3%) and harvest (2%). The types and proportion of disturbances detected on the ground by FIA were similar to those observed by analyst interpretation of the LTS data (

Disturbance observations

In this study, a new LTS methodology was used to integrate various scales of data to improve estimates of disturbance from a design-based sample of FIA plots located in the Uinta Mountains of Utah, USA. This involved using an analyst to record forest disturbances based on visual evidence gathered from annual Landsat images, NAIP photography, and other ancillary spatial data sets (e.g. MTBS, Landfire) (Fig. 2). One advantage to this approach is that it allows all plots to be measured annually

Conclusion

In this study we demonstrate two ways LTS can improve traditional, design-based forest inventory estimates of disturbance derived from FIA data. First, a 26 year satellite time series was used to collect supplemental disturbance history information on 449 forested FIA plots located in the central Uinta Mountains of Utah, USA. This involved using a trained analyst to interpret and record the cause and timing of all disturbances occurring on each FIA plot between 1995 and 2011. As the study area

Acknowledgments

This research was funded by the National Aeronautics and Space Administration (NASA) Terrestrial Ecology Program through the North American Forest Dynamics (Phase 3) project. Additional support was provided by the Interior West Region of the US Forest Service's Forest Inventory and Analysis (FIA) program, FIA's Techniques Research Band (TRB), and NASA's Applied Science Program. Additional technical assistance was provided by the US Forest Service's Remote Sensing Application Center (RSAC) and

References (61)

  • R.E. McRoberts et al.

    Remote sensing support for national forest inventories

    Remote Sensing of Environment

    (2007)
  • R.E. McRoberts et al.

    Statistical inference for remote sensing-based estimates of net deforestation

    Remote Sensing of Environment

    (2012)
  • G.W. Meigs et al.

    A Landsat time series approach to characterize bark beetle and defoliator impacts on tree mortality and surface fuels in conifer forests

    Remote Sensing of Environment

    (2011)
  • P. Olofsson et al.

    Making better use of accuracy data in land change studies: Estimating accuracy and area and quantifying uncertainty using stratified estimation

    Remote Sensing of Environment

    (2013)
  • J.E. Vogelmann et al.

    Monitoring gradual ecosystem change using Landsat time series analyses: Case studies in selected forest and rangeland ecosystems

    Remote Sensing of Environment

    (2012)
  • M.A. Wulder et al.

    Opening the archive: How free data has enabled the science and monitoring promise of Landsat

    Remote Sensing of Environment

    (2012)
  • Z. Zhu et al.

    Continuous monitoring of forest disturbance using all available Landsat imagery

    Remote Sensing of Environment

    (2012)
  • G.P. Asner

    Geography of forest disturbance

    Proceedings of the National Academy of Sciences

    (2013)
  • B.J. Bentz et al.

    Climate change and bark beetles of the western United States and Canada: Direct and indirect effects

    BioScience

    (2010)
  • E.B. Brooks et al.

    Improving the precision of dynamic forest parameter estimates using Landsat

  • E.B. Brooks et al.

    On-the-fly massively multitemporal change detection using statistical quality control charts and Landsat data

    IEEE Transactions on Geoscience and Remote Sensing

    (2014)
  • W.G. Cochran

    Sampling techniques

    (1977)
  • W.B. Cohen et al.

    Landsat's role in ecological applications of remote sensing

    BioScience

    (2004)
  • P.R. Coppin et al.

    Digital change detection in forest ecosystems with remote sensing imagery

    Remote Sensing Reviews

    (1996)
  • R.L. Czaplewski et al.

    Classification accuracy for stratification with remotely sensed data

    Forest Science

    (2003)
  • V.H. Dale et al.

    Climate change and forest disturbances

    BioScience

    (2001)
  • V. DeBlander et al.

    Utah forest health conditions 2011

  • M. Dornelas

    Disturbance and change in biodiversity

    Philosophical Transactions of the Royal Society Biological Sciences

    (2010)
  • D.R. Easterling et al.

    Observed climate variability and change of relevance to the biosphere

    Journal of Geophysical Research

    (2000)
  • J. Eidenshink et al.

    A project for monitoring trends in burn severity

    Fire Ecology Special Issue

    (2007)
  • Cited by (51)

    • Forest disturbance characterization in the era of earth observation big data: A mapping review

      2024, International Journal of Applied Earth Observation and Geoinformation
    View all citing articles on Scopus
    View full text