Land Cover and Forest Structure in the St. Petersburg Region of
Russia:
Integrating Landsat with Forest Inventory Data
INTRODUCTION:
Russian forests have been identified as a potentially important
sink for carbon sequestration, but accurate information regarding the current
status of carbon storage is lacking (Krankina et al. 1996). Using detailed
information at the local forest level (Kukuev et al. 1997), remote sensing
techniques may be employed to extrapolate land and forest cover information
across the vast Russian landscape. The purpose of this study is to use Landsat
Thematic Mapper (TM) data in conjunction with the Russian Forest Inventory
System to characterize forest cover and biomass storage for a 76,850
km2 region around St. Petersburg.
OBJECTIVES:
- To map the land and forest cover of the St. Petersburg Region of Russia
for the purpose of carbon storage modeling.
- To test the usefulness of the Forest Inventory System data for continuous
mapping of forest age and biomass.
STUDY AREA:
The St. Petersburg region of northwestern Russia is located between 58° and
62° N and between 28° and 36° E. While the administrative region
occupies over 100,000 km2, much of that area belongs to the Baltic
Sea and Lake Ladoga, the largest lake in Europe. The influence of these water
bodies helps create a maritime climate for the region, with cool wet summers and
long cold winters. Mean temperature in July ranges from 16° to 18° C,
and in January it is -7° to -11° C. The landscape is typically frozen
from November until March, such that much of the annual precipitation of 600-800
mm falls as snow.
The natural vegetation of the area is southern taiga; major dominant conifer
species include Scots pine (Pinus sylvestris L.) and Norway spruce (Picea abies
(L.) Karst.), growing in both pure and mixed stands. After disturbance, these
species are often replaced by northern hardwoods, including birch (Betula
pendula Roth.) and aspen (Populus tremula L.). Most of the region is part of
the East-European Plain with elevations between 0 and 250 m amsl. The terrain
is flat and consists of ancient sea sediments covered by a layer of moraine
deposits. Toward the northwest, glacial features dominate the landscape and
bedrock topology is more prominent. Soils are mostly podzols on deep loamy to
sandy sediments.
The St. Petersburg region has a long history of agricultural and forest
management dating from the 18th century. Most of the forests are second-growth.
The land-use history is similar to that found in Scandinavia but distinct from
many other boreal regions, such as Siberia and Alaska. The human population of
the region is close to 7 million, with over 5 million people living in the city
of St. Petersburg (Krankina et al. 1998).
IMAGE PRE-PROCESSING:
Thirteen separate TM and one Multi-Spectral Sensor (MSS) images from 1986-1995
were used in this project. The presence of clouds required using images from
different years and different seasons. Before data analysis could begin, we
processed the images to apply geometric rectification and radiometric
normalization.
A cloud free image from the central part of the region acquired on 19 May 1992
was used as a 'core' scene for which all other scenes were spatially adjusted.
Adjacent scenes from the same decade were shifted to match the core scene in the
overlap region; images from the 1980's were matched to the 1990's images using
an automated tie-point selection procedure and second-order polynomial
transformation.
Each of the TM images was radiometrically normalized to the core scene by first
selecting 'no-change' pixels within an overlap area which represented water,
forest, and bright objects, and then using linear regression to calculate
band-by-band correction equations. For images that did not contact the core
scene, the procedure was performed with the next adjacent image.
FOREST INVENTORY DATA:
In the St. Petersburg region of Russia all lands under state forest management
(about 75% of all forest lands) are inventoried with detailed on-site surveys
every 10 years by the Northwest State Forest Inventory Enterprise. Two types of
forest inventory data from 1992-93 survey were used in this study. Stand-level
databases for three forests (Roschino, Tosno, and Volkhov) were used to map
forest stand attributes (forest species composition and biomass), while regional
summary inventory data was used for comparison with the region land-cover
classification.
The three ranger districts were selected to represent the variation in forest
stand attributes found within the region. Field crews surveyed each forest stand
polygon (a homogeneous patch of forest vegetation) delineated from air photos on
the entire territory of the selected forests (97,714 ha). The standard set of
data gathered in the field included site productivity and drainage, tree species
composition, mean height, diameter and age, canopy structure, wood volume and
characteristics of different types of land without tree cover (e.g., clearcuts,
bogs, meadows). Over 200 different variables measured or visually estimated in
the field were used to describe stand polygons, depending on land category
(Kukuev et al. 1997). The databases for all three forests included a total of
12,791 stand polygons. From these we selected about 1500 stand polygons, based
on size and spectral signals, and randomly subdivided them into testing and
training sets.
Biomass was calculated from forest inventory data using available allometric
equations (Alexeev and Birdsey 1998). Calculated biomass stores do not exceed
310 Mg/ha while in 64% of all stand records the biomass is between 100 and 200
Mg/ha. The range of biomass observations is well distributed among the conifer,
hardwood, and mixed species groups. In addition, the field data adequately
represent the limited variety of tree species and stand ages in St. Petersburg
region.
Regional summary of forest inventory provides an estimate of land cover for
6.101 million ha of "Forest Fund Lands" (this is a default land-use designation
of lands not assigned to other uses, i.e. agriculture, urban, mining, parks).
Because only a very small portion of inventory data was used in image
interpretation (about 15 thousand ha or <0.5% of the total forest area), we
consider the results of image classification and forest inventory summary to be
independent of each other.
LAND COVER MAPPING:
The land cover map of St. Petersburg region was constructed using a combination
of remote sensing procedures. Initially, each scene was subjected to an
unsupervised classification and labeled as one of ten cover types. Confused
clusters were reclassified until they could be defined. The individual scenes
were then mosaicked together using a decision rule that conserved forest over
other classes, so that clouds and clearings would not replace forest. Where
possible the more recent images (1994-1995) were used instead of earlier scenes.
After a full mosaic of the study area was prepared, extensive hand digitizing
was performed to distinguish agriculture and built environments from forest and
shrub. The boundaries of developed areas were delimited by expert judgement
based on shape, texture, and proximity. Similarly, mined bogs were digitized
where human alterations were apparent.
Once the forest area for the region was identified, the forest inventory
training polygons were used to drive a supervised classification to produce five
forest cover classes (pine, spruce, hardwood, mixed conifer, and mixed
conifer/hardwood), which were later collapsed into the three final forest
classes. A small area in the far northwest was not covered by TM imagery; the
MSS image was sufficient to classify it as forest, but not further.
Accuracy assessment of this map was done using 822 of the forest inventory
testing polygons as well as 300 randomly generated points in the non-forest
classes. The overal map accuracy is 70.9%, with most of the confusion found in
the mixed forest and shrub classes.
CONTINUOUS VEGETATION MODELLING:
Rather than mapping carbon storage via forest structure models (Cohen et al.
1996), we elected to calculate continuous biomass estimates directly from TM
values (Cohen et al. 2001). To reduce model bias, we used canonical correlation
analysis to produce an index of transformed TM values which was used in a
reduced major axis (RMA) regression to produce linear models. Unlike
traditional linear regression, which minimizes the sum of squared differences in
the Y direction, RMA regression minimizes the sum of squared residuals in both X
and Y directions (orthogonally), thus reducing bias of the fit (Curran and Hay
1986).
In the graph below, predicted and observed biomass values for the testing data
set are show by species group. Each species was modelled separately, yet the
average error is low throughout. Overall, this modelling approach yielded a
correlation of testing predicted vs. observed of 0.62, with a root mean squared
error of 43.5 t/ha. Measures of model bias and variance showed that the
predicted values effectively recreated the mean and variance of the
observations.
RESULTS:
The map below shows the application of biomass prediction models across the full
region. Estimates less than 1 t/ha have been truncated.
Extrapolating the total forest biomass of the region from the raw predictions
yields a result of 541 Tg, which is comparable to the Forest Inventory estimate
of 526 Tg. Other comparisons of the Landsat estimates and the inventory values
are shown in the table below.
SUMMARY:
A Landsat-based land and forest cover classification scheme was used to produce
a regional land cover map for the St. Petersburg Region of Russia. In addition,
forest inventory plots were used to calculate continuous forest biomass
estimates. Both products were in good agreement with inventory-based results.
While cloud cover, landscape conversion, and the absence of inventory data
outside of the core scene certainly confounded our results (Homer et al. 1997),
the overall maps provide a more complete coverage than the forest inventory
method, which did not include several types of tree-covered lands, such as
municipal parks, orchards, and other treed areas. Overall the satellite-based
methodology promises to be a cost-effective and efficient tool for gathering
carbon storage information at the regional scale.
REFERENCES:
- Alexeyev, V. A. and R. A. Birdsey. 1998. Carbon storage in forests and
peatlands of Russia. USDA Forest Service General Technical Report NE-244.
- Anonymous. 1995. Guidelines for inventory of the state forest lands of
Russia. Russian Federal Forest Service, Moscow, Russia (in Russian).
- Cohen, W. B., M. E. Harmon, D. O. Wallin, and M. Fiorella. 1996. Two
decades of carbon flux from forests of the Pacific Northwest. BioScience
46(11):836-844.
- Cohen, W. B., T. K. Maiersperger, T. A. Spies, and D. R. Oetter. In press.
Modelling forest cover attributes as continuous variables in a regional context
with Thematic Mapper data. International Journal of Remote Sensing.
- Curran, P. J., and A. M. Hay. 1986. The importance of measurement error
for certain procedures in remote sensing at optical wavelengths.
Photogrammetric Engineering & Remote Sensing 52(2):229-241.
- Homer, C. G., R. D. Ramsey, T. C. Edwards, Jr., and A. Falconer. 1997.
Landscape cover-type modeling using a multi-scene Thematic Mapper mosaic.
Photogrammetric Engineering & Remote Sensing 63(1):59-67.
- Krankina, O. N., M. Fiorella, W. Cohen, and R. F. Treyfeld. 1998. The use
of Russian forest inventory data for carbon budgetiing and for developing carbon
offset strategies. World Resource Review 10(1):52-66.
- Krankina, O. N., M. E. Harmon, and J. K. Winjum. 1996. Carbon storage and
sequestration in the Russian forest sector. Ambio 25(4):284-288.
- Kukuev, Y.A., O. N. Krankina and M. E. Harmon. 1997. The forest inventory
system in Russia. Journal of Forestry 95(9):15-20.
CREDITS:
This project is funded by the NASA Land Cover - Land Use Change Program (LCLUC).
The authors gratefully acknowledge the significant efforts of Rudolf Treyfeld of
the NW Forest Inventory Enterprise, St. Petersburg, Russia, and Gody Spycher of
the Oregon State University Department of Forest Science.
DIGITAL PRODUCTS:
Digital data layers (Arc/INFO Grid format or Imagine8.3 format) are available
for downloading below. The download files are Gzipped ARC/Info export files
(.e00.gz) or Gzipped Unix Tar Archives (.tar.gz). If your browser allows you to
save them to disk, you can then uncompress them in UNIX with the gunzip
<file.e00.gz> command, or in NT with WinZip. If your browser uncompresses the files
for you, you can either wait for the file to finish reading and then use the
File/Save As... command, or try a Right-hold/Save Link As... mouse
command to save the export file to your local workspace. Either way, you may
then Import the file to an ARC/Info coverage with ARC/Info or ArcView or
use tar xvf to extract the archive.
MAP & POSTER PRODUCTS:
** May require Adobe Acrobat v.4 **
| MAP: Land Cover of St. Petersburg Region, 1986-1995 |
Preview |
Adobe PDF (5
Mb) |
| POSTER: Integration of Satellite Imagery and Stand Inventory Data for
Mapping Carbon Stocks in Forest Ecosystems of Northwest Russia (Poster presented
at IBFRA 2000: The Role of Boreal Forests and Forestry in the Global Carbon
Budget, 8-12 May 2000, Edmonton, Alberta) |
Preview |
Adobe PDF (17
Mb) |
| POSTER: Land Cover and Forest Biomass in the St. Petersburg Region of
Russia: Integrating Landsat with Forest Inventory Data (Poster presented at IGBP
2001: The Global Change Open Science Conference, 10-13 July 2001, Amsterdam,
Holland) |
Preview |
MS Power Point (10
Mb) |
CONTACT:
Doug Oetter
Laboratory for the Application of
Remote Sensing in Ecology
Forestry Sciences Lab
3200 SW Jefferson Way
Corvallis, OR 97331
541-737-8417
oetter@fsl.orst.edu
Disclaimer
FSRN home /
LARSE
home
/ 30 July 01
