Monitoring Forest Cover Transitions using Landsat and Forest Inventory Data

ticketdonkeyΤεχνίτη Νοημοσύνη και Ρομποτική

25 Νοε 2013 (πριν από 3 χρόνια και 7 μήνες)

101 εμφανίσεις

Monitoring Forest Cover Transitions using Landsat and Forest Inventory Data

Olivier R. van Lier
, Joan E. Luther

and Donald G. Leckie

Sir Wilfred Grenfell College, Memorial University, 20 University Drive, Corner Brook,

NL, Canada, A2H 6P9; 709.637.4944;

Canadian Forest Service, Natural Resources Canada, 20 University Drive, Corner Brook,

NL, Canada, A2H 6P9; 709.637.4971;

Canadian Forest Service, Natural Resources Canada, 506 West Burnside Road, Victoria,

BC, Canada, V8Z 1M5; 250.363.0624;


Forested landscapes are dynamic ecosystems which produce a variety of ecological,
social and economic values. Up
date land cover (LC) information ensures a sound
foundation for managing forested landscapes while past LC information provides
cal trends and can thus assist in future management planning. Moreover,
quantifying spatial forest change information is important for assessing impacts of human
activities and environmental changes on ecosystem sustainability.

In this paper, we prese
nt methods to monitor forest transitions using satellite imagery
and forest inventory data. We demonstrate the applicability of the methods for the
Humber River Basin (HRB), Newfoundland, a region where there are land and resource
management issues centere
d on maintaining a sustainable timber supply, conserving
wildlife habitat, preserving high profile viewscapes, and protecting water supply areas.
First, we generate a time series of LC maps using a combination of Landsat MSS and
Landsat TM data. We then qu
antify forest cover transitions over time and produce
spatially explicit maps of change. Such information can provide decision
makers with an
early warning system that can be further linked to human activities and environmental

Materials and M

Study area and data preparation

The Humber River Basin study area occupies 9 290 km2 in coastal Western
Newfoundland and Labrador, Canada (Fig. 1). Contained within the Boreal Shield
Ecozone, the HRB is largely dominated by balsam fir (Abie
s balsamea (L.) Mill.), black
spruce (Picea mariana (Mill.) B.S.P.) and to a lesser extent, white birch (Betula
papyrifera March.). Primarily disturbed by forest harvesting and insect damage, the HRB
has also experienced a significant increase in socio
nomic development over the last
two decades. Although 85
90% of harvested areas regenerate naturally in this region,
economic developments result in permanent LC transitions.

The study area was represented by three Landsat TM scenes (0426, 0525
and 0526)
acquired for the years 1990, 2001 and 2007 (Fig. 1a) and a single Landsat MSS scene
(0526) acquired for the year 1976. All Landsat scenes were resampled to a 25 m pixel
size, manually registered, mosaicked and reprojected using manual ground cont
rol point
collection with Geomatica OrthoEngine® software (©PCI Geomatics, Richmond Hill,
Ontario, Canada). Root mean square errors were maintained below 0.35 pixels to
maintain sub
pixel precision for change detection. Prior to mosaicking, the 2001 Landsa
images; hereafter referred to as baseline, were radiometrically normalized to a 2001
day composite image using Theil
Sen robust regression (Olthof et al. 2005).
Similarly, the non
baseline 2007 images were normalized to a 2005 MODIS 32
site image. Non
baseline mosaics were then normalized to the baseline mosaic
using the band statistical comparison technique (Joyce and Olsson, 1999). Processing and
analysis was performed on all pairs of baseline and non
baseline mosaics.

Three disturbance maps were generated from the provincial forest inventory database
to reflect disturbances occurring for each temporal transition period (i.e. 1976
2001, 2001
2007). Disturbance maps were independently updated by
nd and Labrador’s Department of Natural Resources, Forestry Services
Branch, within forest management district offices. The HRB is intersected by provincial
management districts 9, 12, 13, 14, 15, 16, and Gros Morne National Park (Fig. 1b) and
the inventor
ied year varied though 1977 to 2003 from one district to another. The
inventories are updated on average every 13 years by interpreting 1:12,500 scale aerial
photographs. Disturbance polygons from the updated inventories (i.e. harvest, windthrow,
fire or i
nsect) were selected and stratified by year of occurrence to match the analysed
transition periods.

Fig. 1. Location of the Humber River Basin overlaid with a) Landsat TM scene

and b) forest management districts.

a) b)

Mapping land cover

Through a joint collaboration of Natural Resources Canada and the Canadian Space
Agency, the Earth Observation for Sustainable Development of Forests (EOSD) project
produced a nationa
l LC map of the forested areas of Canada (Wulder et al. 2003; 2008a)
which provided baseline coverage of the HRB. Through the EOSD project, the baseline
imagery was classified using unsupervised K
means clustering applied to individual
NDVI strata masks (w
ater, non
vegetated, low reflectance vegetation, high
vegetation) which were created during image pre
processing. Clusters were labelled as
one of 23 land cover classes through visual interpretation of multiple false color
composite image enhan
cements in combination with forest stand maps, ancillary data,
and local knowledge. Updating LC for the non
baseline imagery was accomplished by
first applying an unsupervised k
means clustering algorithm to each pair of baseline and
baseline imagery o
n bands 1 through 5, and 7 (Leckie et al. 2008). To mask change,
clusters were labelled as either change or no change. Unsupervised K
means clustering
was then applied solely to the non
baseline image bands and the resulting clusters were
merged and labell
ed as per the labelling process of the baseline. The mapping and
updating procedures were validated by comparing 294 ground survey plots sampled in
2008 with the 2007 LC product. Finally, for each LC map, the estimated areal
distribution of a) non
d classes, b) vegetated non
treed classes, and c) vegetated
treed classes were quantified.

Forest cover transitions

Forest transitions were quantified from LC transition matrices generated by cross
tabulating the LC maps for the three transition p
eriods. As suggested by Puyravaud
(2003), annual rates of change were calculated for LC types using the following
rate formula:



where Px is the percent

rate of change for a given LC type x, and Ax1 and Ax2 are x’s
areal estimate at time t1 and t2, respectively. The transition classes were validated by
comparing the transition of the 2001 LC map label to the 2008 ground survey plot label
with the 2001 to
2007 LC transition. Net forest cover depletion and regeneration maps
were then produced by mapping the transition of a treed cell to a non
treed cell, and a
treed cell to a treed cell, respectively. Percent cover of forest depletion and
regeneration we
re mapped for 10 ha grid cells.


The overall classification accuracy of the 2007 LC map at the vegetation type level
was 74.15% with a Kappa Index of Agreement (KIA) of 0.68. The vegetated treed
surface area comprised 70.6% of the total HR
B area in 1976; 70.3% in 1990, 67.1% in
2001 and 67.7% in 2007. The annual rates of change varied within transition periods. The
forest cover (i.e. aggregation of all treed LC types) changed at a rate of
0.03%/yr for the
1976 to 1990 transition period,
.42%/yr for the following transition period (i.e. 1990 to
2001) and increased at a rate of 0.16%/yr for the final transition period (i.e. 2001 to2007).
The coniferous LC type decreased at a rate of 0.21%/yr and was accompanied by
increasing change rates fo
r the vegetated non
treed (0.65%/yr), broadleaf (1.45%/yr),
mixed wood (0.61%/yr) and wetland treed (0.09%/yr) LC types.

The overall accuracy of the 2001
2007 transition classes (depletion, regeneration,
treed no change and non
treed no change) was 8
7.8% with a KIA of 0.83. We obtained
user accuracies of 91.7% for depletion, 89.7% for regeneration, 92.0% for treed no
change and 82.4% for non
treed no change. Disturbances mapped in the forest inventory
corresponded well with the treed LC to non
treed L
C transition (Fig. 2).

Fig. 2. Forest depletion mapped from a) forest inventory disturbance data and b) LC
transition analysis, for the 2001
2007 time period.


The implementation of land cover update methods and resulting generation of
multitemporal LC maps enabled post
classification analysis of areas of LC conversion
(i.e. change of LC type) or modification (i.e. change within a LC type) for a largely
sted region of western Newfoundland, Canada. These LC changes formed the basis of
further assessment of forest cover transitions and annual rates of change while the
sequence of generated LC maps portrayed the spatial and temporal pattern of these

The use of forest inventory data was an obvious source of information to validate our
LC and forest transitions. However, many issues exist when comparing pixel
based LC
maps with vector
based polygon interpretations of forest inventory (Wulder et al
. 2006).
Moreover, since the HRB covers several forest districts, it was impossible to represent
a) b)

the entire basin with a single time stamp of LC derived from forest inventory since the
forest inventory data interpretation years varied by management distric
t. Therefore, the
forest inventory was not well suited for long
term monitoring and reporting of change for
the HRB. On
other hand, the analysis of satellite remote sensing data, made it
possible to extract trend information that was otherwise difficul
t to generate from the
available forest inventory.


This study analysed forest cover transitions within the Humber River Basin from 1976
to 2007, using Landsat and forest inventory data. The LC updating procedure adopted
here was well adapt
ed to integrate continuous monitoring information on LC and forest
change to support resource management decision
making. The continuous Landsat
coverage allowed for spatially
explicit monitoring of change in land cover and was useful
in assessing trends t
o support local and regional resource management activities.

Literature Cited

Fernandes, R. A., & Leblanc, S. G. 2005.
Appropriate linear regression techniques for the
calibration of remote sensing models: When classical linear regression should not be
ed. Rem. Sens. Env. 95: 303

Joyce S. and H. Olsson. 1999. Long
term Monitoring with Temporal
Trajectories from Landsat TM Data. IUFRO conference on Remote Sensing and
forest monitoring. Rogow, Poland. June 1

Leckie, D., Cranny, M., Henley,

M., Luther, J., van Lier, O. and E. Malta. 2008. Satellite
Analysis Procedures for National Scale Land Cover Map Update. Proceedings of the
10th International Circumpolar Canadian Symposium on Remote Sensing, June 2
Whitehorse, Canada. 11 p.

Olthof, I.
, D. Pouliot, R. Fernandes and R. Latifovic. 2005. Landsat
7 ETM+ radiometric
normalization comparison for northern mapping applications. Rem. Sens. Env. 95:

Puyravaud, J. P. 2003. Standardizing the calculation of the annual rate of deforestation.

Forest Ecol. Manag. 177: 593


Wulder, M., J. Dechka, M. Gillis, J. Luther, R. Hall, A. Beaudoin and S. Franklin. 2003.
Operational mapping of the land cover of the treed area of Canada with Landsat data,
EOSD land cover program. For. Chron. 79 (6): 10


Wulder, M., White, J., Luther, J., Strickland, G., Remmel, T. and S. Mitchell. 2006. Use
of vector polygon for the accuracy assessment of pixel
based land cover maps. Can. J.
Rem. Sens. 32(3): 268


Wulder, M. A., White, J. C., Cranny, M., Hall,

R. J., Luther, J. E., Beaudoin, A.,
Goodenough, D. G. and J. A. Dechka. 2008. Monitoring Canada’s forests. Part 1:
Completion of the EOSD land cover project. Can. J. Rem. Sens. 34 (6): 549