PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2014
143
Prediction of Forest Attributes with Field Plots,
Landsat, and a Sample of Lidar Strips: A Case
Study on the Kenai Peninsula, Alaska
Jacob L. Strunk, Hailemariam Temesgen, Hans-Erik Andersen, and Petteri Packalen
Abstract
In this study we demonstrate that sample strips of lidar in
combination with Landsat can be used to predict forest at-
tributes more precisely than from Landsat alone. While lidar
and Landsat can each be used alone in vegetation mapping,
the cost of wall to wall lidar may exceed users’ financial
resources, and Landsat may not support the desired level
of prediction precision. We compare fitted linear models
and k nearest neighbors (
kNN
) methods to link field mea-
surements, lidar, and Landsat. We also compare 900 m
2
and
8,100 m
2
resolutions to link lidar to Landsat. An approach
with lidar and Landsat together reduced estimates of residual
variability for biomass by up to 36 percent relative to using
Landsat alone. Linear models generally performed better
than
kNN
approaches, and when linking lidar to Landsat,
using 8,100 m
2
resolution performed better than 900 m
2
.
Introduction
Studies which demonstrate ways to use lidar in forest in-
ventory, mapping, and monitoring are now fairly common.
Investigators have modeled and estimated forest attributes
(Tonolli
et al.
, 2011; Strunk
et al.
, 2012a) classified forest
types (Pascual
et al.
, 2008) species (Kim
et al.
, 2009a; Zhang
and Qiu, 2012), and condition (Kim
et al.
, 2009b), delineated
stand boundaries(Sullivan
et al.
, 2009), and segmented upper
canopy tree crowns (Hyyppa
et al.
, 2001). These and most
other studies demonstrate approaches which rely on complete
lidar coverage for their area of interest (
AOI
). However, the
acquisition of lidar for an entire
AOI
is not always justifiable
due to high costs, especially for large
AOI
s.
Recently, interest has increased in approaches to estimate
forest attributes from a sample of lidar strips (or swaths)
(Gregoire
et al.
, 2011; Ståhl
et al.
, 2011; Andersen
et al.
,
2011a). Unfortunately, while a sample of lidar strips is less ex-
pensive than complete lidar coverage, a sample of lidar strips
is not directly suited to mapping. To map between the strips
requires an additional source of auxiliary information. One
option is to fill in the gaps between lidar strips using lower
cost reflectance information collected with a passive remote
sensing technology such as Landsat or aerial photography.
The use of lidar with alternate sources of remote sensing
has also been demonstrated in the literature; although most
efforts used combined lidar and spectral information when
both were available for the same areas (Packalén and Malta-
mo, 2006; Hudak
et al.
, 2006; Popescu
et al.
, 2004). Fit sta-
tistics for models developed in these studies did not appear
to appreciably improve when spectral information is used in
addition to lidar. Spectral information can provide improve-
ments in species differentiation over lidar alone (e.g., Ørka
et al.
, 2012). There are also examples of studies for which
lidar was only available for a subset of the
AOI
, while spectral
information was available over a broader area. Wulder
et al.
(2007) used Landsat Enhanced Thematic Mapper (
ETM+
) data
and successive profiling lidar measurement to study change
in vertical height over a period of time. Landsat data were
used to segment the region, and then lidar was used to assign
height data for the segments in successive lidar acquisitions.
The authors found this approach to be more effective for de-
tecting change than simply differencing the strips. A similar
approach by Andersen
et al.
(2011b) used Landsat and pola-
rimetric
SAR
to classify the landscape with a nearest neighbor
approach to classify the landscape. Scanning lidar data were
then used to estimate average biomass for the classes. The
approach was aimed at estimation (e.g., of the population
mean or total) rather than prediction (e.g., for mapping). A
similar approach by Chen and Hay (2011) for a small test area
compared multiple regression and support vector machines to
relate characteristics of image segments to lidar data; although
unlike in Andersen
et al.
(2011b), the image segments were
individual tree crowns.
In a study by Hudak
et al.
(2002) simulations were used to
look at estimation of canopy height from Landsat
ETM+
and
samples of lidar data for different numbers and configurations
of lidar samples. The authors compared a variety of approach-
es including geo-statistical models and were successful in
improving the precision of predictions for areas not covered
with lidar. However, with 2000 m being the greatest distance
between lidar measurements, it is not clear how well these
approaches would perform for lower lidar sampling intensities
(e.g., the sampling intensities used by Andersen (2009), Ander-
sen
et al.
(2011a), Gregoire
et al.
(2011), and Ståhl
et al.
(2011).
We consider the modeling approach to be of great impor-
tance in evaluating a prediction strategy, both in terms of per-
formance (precision) and utility. Two common approaches to
Jacob L. Strunk and Hailemariam Temesgen are with
the Department of Forest Engineering, Resources and
Management, Oregon State University, Corvallis, OR, 97331
).
Hans-Erik Andersen is with the USDA Forest Service PNW
Research Station, Seattle, WA.
Petteri Packalen is with the Faculty of Science and Forestry,
University of Eastern Finland, Joensuu, Finland.
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 2, February 2014, pp. 143–150.
0099-1112/14/8002–143
© 2013 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.2.143
