0888-B1

Monitoring Forests from Space: Hyperspectral and Kyoto Products

David G. Goodenough^[1], Andrew Dyk, K. Olaf Niemann, Hao Chen, Tian Han, Sarah McDonald and Greg Wilson

Abstract

Remote-sensing data can be used to create products to support national and international agreements on sustainable forest management and the Kyoto Protocol. Examples of products derived from satellite hyperspectral data and from multitemporal Landsat data are presented in this paper. Multitemporal Landsat data from 1985, 1990, 1996 and 2001 were orthorectified and used to create forest classifications and biomass estimates. The multitemporal products were used to create above-ground carbon maps, and reforestation, afforestation and deforestation maps. The above-ground carbon measurements were compared with those derived from a traditional forest inventory for our test site near Hinton, Alberta, Canada. The remote-sensing methods reported twice as much forest area, and half the biomass, as derived from the forest inventory. The total above-ground carbon results for the Hinton test site from the two methods were in general agreement.

With EO-1 Hyperion data of the Greater Victoria Watershed (on southern Vancouver Island, in British Columbia, Canada), forest species were classified to an accuracy of 90.0% correct. The Hyperion data were orthorectified to a positional accuracy of 10.1 m. Hyperspectral monitoring of forests can be used for forest inventory, forest health and forest chemistry.

Kyoto Products

In 1997, under the Kyoto Protocol, Canada negotiated to cut greenhouse gas emissions by 6%, of 1990 levels, between 2008 and 2012. This agreement, when ratified, will require Canada to report on the state of its forests, including providing information about carbon stored in forests, and about afforestation, reforestation, and deforestation (ARD). Thus, Canada needs effective and accurate measuring tools. We analyzed Landsat imagery to produce forest cover, change, aboveground carbon, and ARD products for a study area near Hinton, Alberta. The study area is approximately 2605 km², and is characterized by a mixture of undisturbed mature forests, regenerating forests, cutblocks, grasslands, roads, water bodies, and human habitations. The study area is shown in Figure 1.

Figure 1. Study area near Hinton, Alberta; from 1985 Landsat (5, 4, 3).

Multitemporal Landsat TM Image Acquisition

Four Landsat images from 1985, 1990, 1996, and 2001 were used to create a fused multitemporal Landsat image for the Kyoto-product analysis. Three images were acquired from LANDSAT-5, and one from LANDSAT-7. The geometric correction and orthorectification was performed on each of the Landsat scenes. The root-mean-square was less than 0.5 pixels. Image data were extracted corresponding to the Hinton study area. Image-to-image registration was performed on 1990, 1996, and 2001 images by using the 1985 image as the reference image.

No atmospheric correction was done on any of the Landsat images because of the lack of corresponding ground-truth information for the early-date Landsat acquisitions. To reduce the radiometric difference between the Landsat images, radiometric normalization was performed. The 1985 Landsat image was chosen as the reference image. The radiometry from the 1990 and 1996 Landsat TM data, and the 2001 Landsat ETM+ data, were compared to the 1985 Landsat TM image. Since there were no significant differences between the 1985 and 1996 radiometry, the radiometric normalizations were performed only on the 1990 and 2001 Landsat images.

Landsat Image Segmentation and Classification

Spatially based supervised classification methods (using segmentation) were applied to the Landsat images. This was done to increase the classification accuracy. The classified forested areas (softwood, hardwood, and mixed wood) were then used as inputs to generate the Kyoto Protocol products. Training classes were identified from the forest-inventory maps provided by the Alberta Forest Service and used to classify the Landsat images. These classes include conifer, mixed wood, regeneration, clearcuts, rocks, clear land, scrub conifers, scrub deciduous, mine, water, deciduous, cloud and cloud shadow. The conifer class contains at least 80% conifer and no more than 20% deciduous. The reverse is true for the deciduous class. All other forest types are labeled as Mixed Wood.

Segmentation was performed on the fused data set with edges defined by every TM band and date. Training segments for each class was selected and the remaining segments were classified by Euclidean distance from the adjusted training classes. Figure 2 shows the multitemporal classification image based on the 1985, 1990, 1996, and 2001 Landsat images. Table 1 lists the results of the assessment of classification accuracy. The overall classification accuracy was 86% correct or better.

Table 1. Results of the classification accuracy assessment.

Figure 2. Multitemporal classification image.

Aboveground Carbon

The forest layer containing conifer, deciduous, and mixed wood was created from the classification images and used to estimate the remotely sensed vegetation biomass and aboveground carbon. The permanent sample plot data from the test area should be used to develop the biomass relationship, which is a function of remote sensing vegetation indices, forest class, site index, and age class:^[2]

Biomass Volume = - 478.58 + 4.5041 × ND45 (Mt/ha)
ND45 = 128 × ((TM4* - TM5*) / (TM4 + TM5)) + 128
Carbon = Biomass Volume × 0.5 (Mt/ha)

However, these data were unavailable to us from Weldwood of Canada Limited (the timber licensee).

Therefore, the formulas derived from another test site with similar species (near Canal Flats in southeastern British Columbia) were applied (Gemmell and Goodenough 1992). The carbon estimates were calculated for each individual Landsat image and compared with those derived from the Canadian Forest Inventory (CanFI) for 1980. The CanFI entries were based on the Phase 2 Inventory of the Alberta Forest Service (1980) and the biomass calculations by (Penner et al. 1997). The carbon figures are in reasonable agreement, with the average aboveground carbon being 5.42 +/- 0.38 megatonnes (Mt), as estimated using remote sensing. However, the remote sensing forest classification resulted in a forest area 1.88 times larger than CanFI forest area. The Alberta forest inventory concentrates on the commercial forest sites. The remote sensing analysis includes all forests. The biomass estimates from the remote sensing data were 50% lower than those of CanFI. This may be an area dilution effect. Table 2 shows the decrease of the forest areas and the aboveground carbon over the period from 1985 to 2001. The average carbon (tonnes/ha) from each date is in good agreement. The carbon map product generated from 1985 Landsat imagery is shown in Figure 3.

Table 2. Normalized carbon estimates for single-data Landsat images.

Figure 3. Carbon image for 1985

Carbon Comparisons in Non-Changed Forested Areas

To determine the consistency of the carbon estimations from Landsat, we also calculated the aboveground carbon in the forested areas where no change occurred during the period 1985 to 2001.

Table 3 shows the carbon results in the unchanged forested areas. The carbon figures were in very good agreement, with a carbon average estimate of 3.83 +/- 0.06 Mt obtained from remote sensing.

Table 3. Normalized carbon estimates for unchanged forested areas.

Afforestation, Reforestation, and Deforestation

Afforestation, reforestation, and deforestation (ARD) are the required Kyoto Protocol products. The authors computed ARD from the classification images generated from the 1985, 1990, 1996, and 2001 Landsat images. The 1985 classification image was used as a reference image since it was the earliest Landsat image of the study area. Based on the ARD definition (Lempriere, T., 2001. Forests and the Kyoto Protocol: Update from COP7. Natural Resources Canada, Canadian Forest Service. Unpublished document), the classes in each of the four classification images were converted into four basic land-type classes: forest class, non-forest class, regeneration class, and unclassified. The ARD calculation was based on a three-date ARD sliding window so that 27 ARD permutations could be accepted (Goodenough et al. 2001). Figure 4 shows the spatial distribution of the ARD classes created from the 1990, 1996, and 2001 classification images.

Figure 4. ARD class image from the 1990, 1996, and 2001 moving frame.

The ARD statistics were calculated from the ARD images. Table 4 gives the detailed ARD statistics for year 1990, 1996, and 2001. The extent of afforestation was similar in each year. The total reforested area in 2001 was about ten times larger than in 1990. However, the total deforested area in 2001 was about two times larger than in either 1990 or 1996. There is still a need for international agreement on ARD definitions when remote sensing is used to assess compliance.

Table 4. ARD statistics in hectares for 1990, 1996, and 2001.

Hyperspectral Sensing

Examining the exchange between energy and forest-ecosystem features on the earth’s surface can provide insight into ecological processes at local and global scales (Dawson et al. 1999). Ecosystem processes such as photosynthesis, evapotranspiration, nutrient cycling, and net primary productivity can be better understood by assessing canopy biochemical content through analysis of chlorophyll, nitrogen, and lignin concentrations (Wessman et al. 1988; Curran 1989). Field sampling of vegetation biochemistry is not feasible on regional and global scales due to time and financial constraints; therefore, remote sensing of canopy biochemistry provides an alternative method of examining these processes on such scales (Treitz and Howarth 1999).

With improved sensor technologies, detailed information about the forest canopy can be derived from spectrometers on airborne (AVIRIS) or spaceborne (Hyperion) platforms. With higher signal-to-noise ratios, more spectral bands, and greater radiometric resolution, the biochemical status of the forest canopy can be examined. AVIRIS data have been used to validate analytical methods and to process satellite hyperspectral data from Hyperion.

Hyperion Data

Hyperion is the hyperspectral sensor carried on NASA’s EO-1 satellite. Radiometric corrections applied to our data are described in Han et al. 2002. Atmospheric correction was performed to determine the ground reflectances of features using the Environment for Visualizing Images’ (ENVI) Fast Line-of-Sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) module, developed by Spectral Sciences, Inc. Details of the geometric correction performed on the Hyperion, ALI (Advanced Land Imager), AVIRIS, and ETM+ data collected over the Greater Victoria Watershed District are described in Dyk et al. 2002. Correction accuracies ranged from 7.81 to 10.1m. Spectra extracted from the hyperspectral data were calibrated using ground spectra from an Analytical Spectral Device (ASD). A force-fit function was applied to Hyperion data to provide accurate reflectance measurements used for forest classification and forest chemistry.

Species Recognition with EO-1 and Landsat Imagery

The Greater Victoria Watershed District provided a detailed forest-cover geographic information system (GIS) database for the study. The GIS data were overlaid on 1-m resolution, black and white orthophotos during the delineation of the ground reference or “truth” polygons used for training for the supervised classification. Classes were selected that showed dominant species of forest cover, as well as other non-forest classes that matched the definitions of the Canadian National Forest Inventory (NFI) photo plots (Natural Resources Canada 2001). The forest cover in the common area consists predominantly of Douglas-fir stands of varying age class and densities, as well as a few stands where lodgepole pine, hemlock, red alder, or western redcedar are dominant.

Eleven MNF channels, with significant eigenvalues - MNF 2 to 12 - were selected for the Hyperspectral classification. For the ETM+ data, bands 1 to 5 and band 7 were used as input to the classification. For the ALI data, the nine multispectral bands 1p to 5 and band 7 were used for the classification.

The supervised classification results for the individual classes before aggregation of the training (truth) areas and the check areas can be found in Goodenough et al. 2002.These classes were aggregated to produce a new set of classes, as shown in Table 5. The order of sensors by classification accuracies for aggregated classes on the test data was: Hyperion 90.0%, ALI 84.8%, and ETM+ at 75.0%.

Table 5. Aggregated classification accuracies for ETM+, Hyperion, and ALI.

The classification results for the aggregated classes (ETM+, Hyperion, ALI) are depicted in Figure 5. The GIS reference data covered 45.03% of the area in common. In addition, Hyperion and ALI each missed a different portion of the corresponding ETM+ scene. Within the GIS reference data and ground plots, we are confident of the classification results.

Figure 5. Common area shown within red line for the 3 sensors (first image). ETM+ classification (second image). Hyperion classification (third image). ALI classification (fourth image).

Additional information can be found at the website for the Evaluation and Validation of EO-1 for Sustainable Development Project (EVEOSD): http://www.pfc.cfs.nrcan.gc.ca/aft/eveosd/index_e.html.

Conclusions

Remote sensing can be used to provide products to meet Canada’s reporting requirements for the Kyoto Protocol. These products include mapping aboveground carbon, afforestation, reforestation, and deforestation. Optical remote sensing can provide accurate measurements of forest area by forest type. Long wavelength polarimetric SAR could provide accurate biomass estimates. However, in the absence of these data, texture measures of optical imagery can be calibrated locally to provide estimates of forest biomass. For the Canadian Forest Service’s test site near Hinton, Alberta, the estimates of stability of carbon were checked for areas of forest that had not changed and were found to be stable within 2%. Estimates of carbon using multitemporal Landsat were similar to those obtained via traditional forest inventory methods, but two major differences occurred: the remote sensing results revealed twice the forest area and half the biomass compared to the inventory results.

Hyperspectral remote sensing can provide more accurate recognition of forest species, and bioindicators of forest chemistry, health, and stress. For a test site near Victoria, British Columbia, Canada, satellite hyperspectral data provided a forest species classification with an average accuracy of 90.0%.

Literature Cited

Curran, P., 1989. Remote sensing of foliar chemistry. Remote Sensing of Environment 30:271-278.

Dawson, T., P. Curran, P. North, and S. Plummer, 1999. The propagation of foliar biochemical absorption features in forest canopy reflectance: a theoretical analysis. Remote Sensing of Environment 67:147-159.

Dyk, A., D.G. Goodenough, A.S. Bhogal, J. Pearlman, and J. Love, 2002. Geometric correction and validation of Hyperion and ALI data for EVEOSD. pp. 579-583 in Proceedings of the International Geoscience and Remote Sensing Symposium (IGARSS '02), Toronto, ON, Vol. I. IEEE.

Gemmell, F.M. and D.G. Goodenough, 1992. Estimating timber volume from TM data: the importance of scale and accuracy of forest cover data. pp. 297-306 in 18th Annual Conference of the Remote Sensing Society. A.P Cracknell and R.A. Vaughan (editors). The Remote Sensing Society, University of Dundee, Scotland.

Goodenough, D.G., A.S. Bhogal, H. Chen, and A. Dyk, 2001. Comparison of methods for estimation of Kyoto Protocol products of forests from multitemporal Landsat. pp. 764-767 in Proceedings of the International Geoscience and Remote Sensing Symposium (IGARSS '01), Sydney, Australia, Vol. 2. IEEE.

Goodenough, D.G., A.S. Bhogal, A. Dyk, A. Hollinger, Z. Mah, K.O. Niemann, J. Pearlman, H. Chen, T. Han, J. Love, and S. McDonald, 2002. Monitoring forests with Hyperion and ALI. pp. 882-885 in Proceedings of the International Geoscience and Remote Sensing Symposium (IGARSS '02), Toronto, Ontario, Vol. 2. IEEE.

Han, T., D.G. Goodenough, A. Dyk, and J. Love, 2002. Detection and correction of abnormal pixels in Hyperion images. pp. 1327-1330 in Proceedings of the International Geoscience and Remote Sensing Symposium (IGARSS '02), Toronto, Ontario. IEEE.

Natural Resources Canada, 2001. National forest inventory design document. Canadian Forest Service, Pacific Forestry Centre. Victoria, BC.

Penner, M., K. Power, C. Muhairwe, R. Tellier, and Y. Wang, 1997. Canada's forest biomass resources: deriving estimates from Canada's forest inventory. Natural Resources Canada, Canadian Forest Service, Pacific Forestry Centre, Victoria, BC. Information Report BC-X-370.

Treitz, P. and P. Howarth, 1999. Hyperspectral remote sensing for estimating biophysical parameters of forest ecosystems. Progress in Physical Geography 23:359-390.

Wessman, C., J. Aber, D. Peterson, and J. Melillo, 1988. Remote sensing of canopy chemistry and nitrogen cycling in temperate forest ecosystems. Nature 335:154-156.