A sample design for Landsat-based estimation of national
trends in forest disturbance and regrowth
Robert
E
Kennedy, USDA Forest Service, PNW Research Station, robert.kennedy@oregonstate.edu
(Presenting)
Warren
B
Cohen, USDA Forest Service, PNW Research Station, warren.cohen@oregonstate.edu
Gretchen
G
Moisen, USDA Forest Service, Rocky Mountain Research Station, gmoisen@fs.fed.us
Samuel
N
Goward, Department of Geography, University of Maryland, sgoward@umd.edu
Michael
Wulder, Canadian Forest Service, Pacific Forstry Centre, Victoria, BC, mwulder@pfc.cfs.nrcan.gc.ca
Scott
L
Powell, USDA Forest Service, PNW Research Station, scott.powell@oregonstate.edu
Jeffrey
G
Masek, Biospheric Sciences Branch, NASA’s Goddard Space Flight Center, jeffrey.g.masek@nasa.gov
Chengquan
Huang, Department of Geography, University of Maryland, cqhuang@geog.umd.edu
Sean
Healey, USDA Forest Service, Rocky Mountain Research Station, shealey@fs.fed.us
Dense temporal stacks of Landsat images have great potential for describing the dynamics of forest disturbance and regrowth, particularly when spectral data are linked with forest inventory data, but require significant time and expense to process and analyze. Here we describe a stochastic, design-based sampling strategy used to identify 23 Landsat scenes in which we are characterizing detailed forest dynamics over the past three decades. The sample was required to meet several competing goals, including capture of diverse forest types and disturbance regimes, minimization of effort expended in low-forest-area scenes, flexibility for expansion of the sample size in the future, and preferential inclusion of scenes where significant prior research had been accomplished. Sample units were defined as the non-overlapping area of Landsat scenes on the WRS-2 grid, were divided into eastern and western sample frames, and were attributed with forest type and area from a recent national-level forest type map. For each frame, 100,000 randomized, ordered lists of scenes were chosen and scored according to the competing goals of the project using a target number of samples in the east and the west. The minimal set of these lists that best balanced all goals and that included each scene in the frame at least once was identified, and from this set a single ordered list was randomly chosen. Probabilities of inclusion for each scene were calculated from the proportion of lists from the final set that included that scene. This strategy allows use of unequal-probability estimators in a design-based estimation paradigm, while also ensuring that a full range of conditions can be used in a model-based estimation paradigm.