Figure 1. Tree-stem volume as a function of spectral features. Study area Orivesi, number of observations 324. r is the correlation coefficient.
² VTT Automation, Space Technology,
Remote Sensing P0 Box 13031,
FIN-02044, VTT, Finland
E-mail Firstname.Lastname`vtt.fi
A new methodology for estimating the biomass (organic matter) of conifer-dominated Boreal forests has been developed in regions where ground data are limited. The principal models are first computed using ground measurements and high-resolution satellite data. The spectral models are then directly applied to a calibrated AVHRR image mosaic covering the entire area of interest. The method was quantitatively tested in Finland and demonstrated on an area reaching from the west coast of Norway to the Ural mountains.
Boreal coniferous forests or Taiga forests form a continuous circumpolar vegetation zone. They are the largest terrestrial biome on Earth (Syrjänen et al. 1994). The mineral soil lands and peatlands of the Boreal forests form a significant pool in the global carbon cycle. It has been estimated that in Boreal forests there are 31 10(exp 12) kg of carbon stored in the trees alone (Kuusela 1990). The Boreal forests are dominated by conifers: spruce, larch, fir and pine. The foreseen climate warming would cause northward movement of the tree line. The amount of biomass close to the tree line should be a good indicator of the effects of the climatic changes and changes in the global carbon cycle.
The primary objective of this study was to exploit optical remote-sensing data to develop a new biomass-estimating methodology for conifer-dominated Boreal forests on mineral soil lands. Methods were also developed for estimating the proportion of broad-leaved trees in the total biomass.
Three study areas were used:
Orivesi was the primary area for model development to estimate forest characteristics using ground data and a Landsat TM image. At Lammi, similar models to those of Orivesi were computed to test how much the parameters change when another image and ground data are utilised. In the third area, the models developed using ground data and Landsat TM image were applied to the NOAA AVHRR weather-satellite mosaic.
Ground data for the Orivesi area were taken from sample plots of the National Forest Inventory. The ground data for the Lammi area were stands from the forest management plan maps of private forest estates.
To estimate the biomass, models were first computed for tree stem volume because this is closely connected with the biomass (the organic matter of the trees) (Nihlgärd 1972, Mälkönen 1974) and because stem volume is measured on the ground in practical forestry. Models were developed to transform the tree stem volume into dry organic matter for the trees, including bark, branches and roots. The organic matter in the undergrowth vegetation was also estimated.
The Landsat TM image for the Orivesi area was acquired on 21 June 1985.
The Landsat image for the Lammi area was recorded on 27 July 1989.
Images from several years had to be selected for the NOAA AVHRR mosaic due to cloud cover, a total of fifteen AVHRR Local Area Coverage images eventually being used to construct the mosaic over the study area. The images had been acquired during the summers of 1990 to 1993, during which the phenological phase of the forest ecosystems is assumed to have been relatively stable.
The AVHRR images were first rectified using the orbital parameters of the satellite. The spectral effect caused by different angles between the Sun and the surface was corrected using the cosine transformation. The rectification was further adjusted using a second rectification based on the measurement of control points. Two images, one reaching far west and another, reaching far east, were first rectified to the map coordinate system using ground control points. Then, the other images were coregistered to the coordinate system of the two reference images.
A semi-physical approach was chosen for biomass estimation. In this approach, the principal models are first computed using ground measurements and Landsat TM data. An image mosaic is made from the AVHRR data. The relationship between Landsat data and AVHRR data is determined using the intensities of the original spectral channels. One benefit of this methodology is that a forest characteristic model, suitable for use with Landsat data, can be directly applied to the AVHRR mosaic.
Linear multiple regression analysis was used to estimate the tree stem volume. Channel 3 (red) and channel 4 (near-infrared) intensities of Landsat TM were used as predictor variables, since their wavelength ranges corresponded to the wavelengths of the two first NOAA AVHRR channels. The spectral feature could be, in addition to an original spectral channel, a ratio of the channels or the NDVI (Normalised Differential Vegetation Index). However, the NDVI models could not be directly applied to the AVHRR data because of the non-linearity of the predictor variables.
The models used to estimate the amount of broad-leaved trees were computed in two ways:
A relative atmospheric correction method was implemented separately for each AVHRR image. The method used the intensities of water in line direction of an AVHRR image. A relative calibration method was used to match the intensities of AVHRR channels 1 and 2 and Landsat TM channels 3 and 4 (Häme 1991, Olsson 1994). The matching was done to make it possible to directly apply forest characteristic models, computed for Landsat data. Two approaches for the radiometric matching were tested:
An AVHRR image from 26 July 1992 (in which the atmospheric effects had been corrected and the intensity matching to the Landsat TM image had been done) was used as a master image to make the AVHRR image mosaic. Intensities of other AVHRR images in which atmospheric correction had been made were matched to this master image using mean and standard deviation matching. After the relative calibration, the actual mosaic was computed as weighted averages of pixel intensities of overlapping images. Pixels with lower intensities had higher weights.
Two mosaics were produced: the 'regression mosaic' and the 'mean and standard deviation matching mosaic' depending on the TM/AVHRR matching method.
After the corrections, a systematic increase in intensities towards the east was detected in the AVHRR 1 channel. Intensities were corrected using a linear scaling procedure in an east west direction. The AVHRR 1 intensities of similar forests in the Urals and Finland were scaled the same way. In the near-infrared channel AVHRR 2, no systematic errors in intensities could be detected.
Relationships between stem volume and reflected optical radiation
Figures 1 and 2 indicate the general relationship between
coniferous forest biomass and reflected optical radiation. The
correlation is negative throughout the optical-wavelength area,
including the near-infrared radiation, which is an opposite
result to the situation found over agricultural lands and
grasslands (Tucker et al. 1975). Because of the very similar
relationship between biomass and reflected red and near-infrared
radiation, the vegetation index, i.e. the ratio between near-
infrared and red, is a very poor indicator of the forest biomass.
Figure 1. Tree-stem volume as a function of spectral features.
Study area Orivesi, number of observations 324. r is the
correlation coefficient.
Figure 2. NDVI computed from the mosaic. NDVI scale from low to
high values: blue-green-yellow-red.
Table 1 shows computed biomass models. The models for the Orivesi and Lammi areas are similar. For instance, the models to estimate total and coniferous stem volume (V and V conifer ) using the red channel (models I, II, V and VI in Table 1) coincide surprisingly well. The results indicate that a biomass model using the red channel TM3 as a single predictor is rather robust to the presence or absence of broad-leaved trees. Models for coniferous stem volume using vegetation indices as predictor variables are very poor (models IX, X).
Table 1. Summary of the models for biomass estimation
Model Predicted Coeff. Coeff. Number
identifier variable Constant- for TM3 for TM4 R² of obs. Area
---------------------------------------------------------------------------------
I V 562 - 21.5 none 0.20 345 Orivesi
II V 522 - 22.1 none 0.15 1595 Lammi
III V 567 - 15.1 - 2.80 0.25 345 Orivesi
IV V 474 - 9.58 - 3.20 0.23 1595 Lammi
V Vconifer 572 - 22.0 none 0.23 250 Orivesi
VI Vconifer 566 - 23.1 none 0.15 685 Lammi
VII Vconifer 590 - 16.4 - 2.87 0.27 250 Orivesi
VIII Vconifer 520 - 11.2 - 3.25 0.21 685 Lammi
IX Vconifer 265 - 46.1 M4/TM3 0.029 260 Orivesi
X Vconifer 242 - 213 NDVI 0.019 250 Orivesi
XI Vbroadleaf 31.5 - 2.69 0.591 0.093 345 Orivesi
XII Vbroadleaf 31.7 - 4.33 1.17 0.12 1595 Lammi
XIII Vbroadleaf -20.8 11.8 TM4/TM3 0.079 345 Orivesi
XIV Vbroadleaf -40.7 19.8 TM4/TM3 0.13 595 Lammi
XV BroProp -0.505 0.240 TM4/TM3 0.36 345 Orivesi
XVI BroProp -0.533 0.233 TM4/TM3 0.29 1595 Lammi
V Stem volume (m³/ha) including all tree species
Vconifer Stem volume (m³/ha) using ground data strongly dominated by
(BroProp 50.05)
Vbroadleaf Stem volume (m³/ha) of broadleaved deciduous trees
BroProp Proportion of deciduous trees of the total stem volume of a plot or a stand
The highest coefficient of determination (R²) values can be seen in models to estimate the proportion of the broad-leaved trees from the total stem volume (models XV and XVI). Note that these models can give high values even if the total biomass is low because the values are relative. The predictor variable in these models was actually the widely used 'vegetation index', i.e. ratio of near-infrared and red.
Models were also computed for the coniferous plots using all combinations of Landsat TM channels as predictors. The best combinations are shown in Figure 3. The value of the coefficient of determination (R²) hardly increased after two predictor variables. According to this test, the AVHRR optical channels are the two optimal channels for coniferous biomass estimation.
Figure 3. Performance (R²) of best Landsat TM models to
estimate tree-stem volume in area Orivesi, coniferous forests,
number of observations 250.
Applying the biomass models to the AVHRR mosaic
Models developed
using Landsat TM image and ground data from the Orivesi area were
applied to the final AVHRR image mosaic. The results were tested
against the National Forest Inventory data for Finland. The test
data were obtained from the Yearbook of Forest Statistics, which
includes results of the inventory by 20 forestry board districts
(Aarne (Ed.) 1992). The total land area of the forestry board
districts was 30.4 million hectares.
The correlations between the measured stem volume and the estimated volume of the forestry board districts were high. However, most models somewhat overestimated the stem volume. The exceptions were the models for broad-leaved trees, which gave huge underestimates. The main reason for such poor results was the lack of training data for the broad-leaved trees.
The highest correlation with the total stem volume, r=0.98, was obtained using a model where the only predictor was the red channel (model I in Table 1). However, it was a poor model for separating forest from non-forest. The overestimation of the stem volume for Finland as a whole was 14 percent.
The model in which both red and near-infrared channels were predictors and in which the TM/AVHRR matching was done using means and standard deviations (model III in Table 1, Figs. 4 and 5), was able to separate forest and non-forest better than the red-channel model. This model underestimated biomass in forestry board districts which had a significant proportion of broad- leaved trees. Overestimates occurred in areas with small lakes and in mountain valleys.
Figure 4. Estimated (black bar) and measured (grey bar) total
stem volume by Forestry Board Districts in Finland. Model III
(Table 1) applied to mean and standard deviation TM/AVHRR
matching mosaic. Correlation between estimated and measured stem
volumes r=0.94; overestimation for Finland as a whole 8%.
Despite certain inaccuracies, the performance of the biomass estimation method developed in this study exceeded expectations. It was shown that the models computed using ground data and Landsat TM images can be successfully utilised with the AVHRR data. The modelling approach chosen in which the parameters of a physical model are computed using regression analysis seemed to be appropriate because the models worked reasonably well over the whole of Finland, and because models computed for two independent areas using two independent images were found to be similar. The models were biased, usually producing overestimates of biomass. Some bias was expected because no calibration of the estimates was made in this estimation procedure.
The red channel was the best single channel for biomass estimation. Models having a red channel as the only predictor gave good results and were robust when changing images and when using different combinations of conifers and broad-leaved trees. If such models are applied, the forested areas have to be separated first using both the red and near-infrared channels.
The vegetation index appeared to be inappropriate for estimating the biomass of the coniferous forests. It seems to be effective in biomass estimation only when the amount of biomass is related to the proportion of bare soil covered by the vegetation.
In the future, procedures for improving the radiometric calibration and atmospheric correction of the images will be developed. A method to calibrate the biomass estimates for bias reduction will be incorporated.
Figure 5. Application of model III to the AVHRR mosaic. light
yellow: zero biomass; turquoise: 1-25 m³/ha; green: 26-50
m³/ha; brownish green: 51-100 m³/ha; brownish red:
101-150 m³/ha; bright red: =150 m³/ha; black: water,
mountains and clouds. The mean stem-volume value in the figure
is 73.9 m³/ha for pixels with biomass estimate value above
0, which corresponds to organic matter of 42 552 kg/ha or 47 352
kg/ha including the undergrowth vegetation. The maximum stem
volume estimate is 472 m³/ha. The area of the forested
pixels (non-zero biomass) is 266 million ha, which gives the
total tree stem volume of 18 900 million m³ and organic
matter of 10 900 000 million kg, or 12 100 000 million kg
including the undergrowth. These indicative dry-weight values
were computed from the stem-volume values using formulae
developed in this study.
The majority of the funding for this study came from The Finnish Research Program on Climate Change (SILMU) of the Academy of Finland. We wish to thank the following individuals for their contributions to the study: Mr Kai Lindh, Mr Kimmo Syrjänen, Dr Vasily I. Sukhikh, Dr Thomas Hüusler and Dr Pamela Kennedy.
Aarne M (Ed.) 1992, Yearbook of Forest Statistics 1990-91, Folia Forestalia 790, 281 p.
Häme T 1991, Spectral interpretation of changes in forest
using satellite scanner images, Acta Forestalia Fennica 222,
111
p.
Kuusela K 1990, The dynamics of Boreal coniferous forests,SITRA Helsinki, 172 p. ISBN 951-563-274-9.
Mälkönen E 1974, Annual primary production and nutrient cycle in some Scots pine stands, Communicationes Institute Forestalis Fenniae 84(5).
Nihlgärd B 1972, Comparative studies on beech and planted spruce forest ecosystems in Southern Sweden, Lund Univ., PhD Thesis, 139 p.
Olsson H 1994, Changes in satellite-measured reflectances caused by thinning cuttings in Boreal forest, Remote Sensing of Environment 50: 221-230.
Syrjänen K, Kalliola R, Puolasmaa A & Mattsson J 1994, Landscape structure and forest dynamics in subcontinental European Taiga, Ann. Zoll. Fennici 31, 19-34.
Tucker C J, Miller L D & Pearson R L 1975, Short-grass prairie spectral measurements, Photogrammetric Engineering and Remote Sensing, 41, 9, 1175-1183.
ESA EOQ 52