Hyperspectral (narrow-band) remote sensing offers new opportunities for mapping vegetation type and function. Using imagery of the boreal forest (Canada), we explored the ability of a hyperspectral sensor (AVIRIS) to map vegetation properties by taking advantage of pigment and water absorption features. Two vegetation mapping techniques were applied. The first uses a combination of reflectance indices, e.g. NDVI, as inputs in a maximum likelihood classification routine. In the second classification procedure, laboratory acquired leaf spectra were utilized in spectral fitting procedures to map pigment abundance in the landscape. The resulting images were used in a statistical classification routine to map the distribution of vegetation cover types. Both of these techniques were able to differentiate stand age and vegetation type at accuracies comparable or superior to other well established classification methods. Further investigation is focusing on developing more quantitative derivations of pigment and water content at the landscape level for the purpose of modeling carbon and water vapor fluxes.
 
 

A major concern in the Boreal Ecosystem-Atmospheric Study (BOREAS) has been the accurate representation of forest cover types which are used as inputs for ecosystem and climate models (Sellers, et al. 1997). At the regional and sub-regional scales, Steyaert et al. (1997) derived an AVHRR (1-Km pixel resolution) vegetation classification based on a combination of field observations and an unsupervised cluster analysis on NDVI. Hall et al. (1997) have developed a Landsat TM physical based classification that uses canopy reflectance models to classify vegetation types and biophysical parameters. These studies produced superior results to those of solely statistical methods; however, low accuracy for fen and dry conifers, two important cover types in the boreal forest, were reported. Narrow-band airborne hyperspectral sensors, AVIRIS (224 spectral bands) and CASI, have provided the opportunity to explore different functional and structural expressions of vegetation such as, NDVI, chlorophyll content, etc. that can be used to improve on the results of previous research (Fig. 1). Using CASI imagery, Zarco-Tejada and Miller (In press) developed an unsupervised classification routine that took advantage of red edge parameters to differentiate vegetation types. With an overall accuracy of 61 % and the ability to map the fen cover type, their study provided a significant improvement over the TM classification. However, this algorithm was unable to classify the deciduous type and differentiate between wet and dry conifers. Thus, continued improvements in this area are still needed.

FIGURE 1: The spectral information in an AVIRIS image cube allows the exploration of various functional and structural expressions of vegetation. In this case, a 224 band AVIRIS image, July 21, 1994, has been submitted to a Spectral Mixture Analysis (SMA) routine using leaf spectral types (representing high levels of anthocyanins, chlorophylls, or carotenoids) as endmembers. The intermediate images, representing relative levels of these pigment classes, can then be combined in a single RGB image. Subsequent classification procedures (e.g., maximum likelihood) can then be applied to this final image to produce maps of vegetation type linked to functional properties.
 
 

Using six 1994 AVIRIS images, Spring (April 19), Fall (September 16) ,and Summer (July 21), for the BOREAS Southern Study Area, near Prince Albert, Saskatchewan (Fig. 2) were converted to reflectance using a modified MODTRAN algorithm developed by Green (1996). The images were georeferenced in ENVI 3.0 (Research Systems) using a panchromatic Landsat TM image as a base map, obtained from the BORIS (Goddard Space Flight Center) BOREAS archive. The images were then mosaicked and resampled to produce 30-meter pixel images encompasing an area bounded by the following UTM zone 13 coordinates (Fig. 3):
 
 

NW corner = 515,986.77 easting, 5977,008.48 northing

NE corner = 525, 858.62 easting, 5,976,998.53 northing

SW corner = 515, 986.77 easting, 5,961, 258.44 northing

SE corner = 525, 858.62 easting, 5, 961, 258.44 northing
 
 

The three resulting images were used to derive seven images of reflectance indices (Table 1, Fig. 4) that are indicators of pigment content, photosynthetic rates, canopy structure and water content. The images were then combined to produce a seven-band image cube.
 
 

SSA forest cover data from the Saskatchewan Environment and Resource Management, Forestry Branch (SERM-FBIU) and processed by the BORIS staff science team into binary raster files were obtained from the BORIS database. The area corresponding to the three seasonal index image and the summer SMA leaf types image was extracted from the SERM binary raster file and vectorized.

Leaf pigment spectra for 11 leaf types containing different levels of chlorophyll, anthocyanins and/or carotenoids were obtained using a Unispec spectrometer (PP Systems) (Fig. 5). The spectra was used to produce a spectral library of leaf types which also included a soil and shadow spectrum. The 224 band reflectance image for summer (July 21, 1994) was submitted to a Spectral Mixture Analysis (SMA) routine using the leaf type endmember spectral library. This resulted in a 13 band image, with each band representing the abundance of an endmember for each pixel in the image cube.
 
 

 TABLE 1: Using these seven reflectance indices, seven "index images" were derived. These seven images were then combined into three seven-band images corresponding to three AVIRIS sampling dates in Spring (April 4), Summer (July 21) and Fall (September 16).
 
 

The results of the two vegetation classification methods are presented in Figures 8-11. A visual comparison of the index based classification, the leaf type based classification and the SERM and TM classifications indicates that in general the index and pigment based classifications tend to correspond better to SERM than the TM image. Another obvious, observation is that standing water category is missing in the classification using indices.

 FIGURE 6: SERM-FBIU vegetation map produced from aerial photography and field observations. The cover categories have been aggregated to six functional vegetation classes: dry conifers, wet conifers, mixed, deciduous, fen, and disturbed.

FIGURE 11: Summer (July 21, 1994). Maximum likelihood classification using an 11 leaf type image cube as input. The image cube was the output of a Spectral Mixture Analysis (SMA) routine run to obtained the abundance of the leaf types in the study area.
 

The results are presented in Tables 2-6. The best overall accuracy was obtained for the spring index-based classification, 53.6%; followed by the summer index based image, 53%, the summer pigment-based image, 52.8%, and the fall index based image, 47.5%. In contrast, the TM classification only classified 42% of all pixels accurately. Zarco-Tejada and Miller (In press) reported an overall accuracy of 61.5% in their classification but without classifying deciduous vegetation or distinguishing dry and wet conifers. By aggregating the dry and wet conifers into a single class, conifers, the overall accuracy of the indices and pigment based classification yielded results closer to those reported by Zarco-Tejada and Miller – between 52.8% (for the fall index-based classification) and 60.4% for the spring index-based classification (Tables 7-11). Additionally, both the index and pigment classifications are performing better in classifying correctly the fen vegetation type than the Zarco-Tejada and Miller and the TM classifications. Fen accuracies ranged between 69.4% for the summer index classification and 76.4% for the summer pigment-based classification.

As indicated from the above results, the maximum likelihood classification for the fall image that used the combination of indices yielded the best classification results. In order to determine which of the indices employed had the greatest influence in the results, a series classifications were attempted in which all possible combinations of indices were utilized. A selected number of representative training classes were chosen for this exercise and these were submitted to the same maximum likelihood routine. From this analysis, it was concluded that modified NDVI and the sum of the reflectance of the green region have the most distinguishing power. 59.64% of all training class pixels were classified correctly. The next most important index was the normalized difference water index (NDWI), followed by NDVI, WBI, PRI, and finally, the red and green ratio. Overall, 77.93% of the training pixels were correctly classified.
 
 

TABLE 7: Accuracy assessment for the Spring index-based Maximum Likelihood classification when the dry and wet conifer types are aggregated.

TABLE 8: Accuracy assessment for the Summer index-based Maximum Likelihood classification when the dry and wet conifer types are aggregated.

TABLE 10: Accuracy assessment for the Summer leaf pigment types Maximum Likelihood classification when the dry and wet conifer types are aggregated.

TABLE 11: Accuracy assessment for the Landsat TM Spectral Trajectories classification when the dry and wet conifer types are aggregated.

By transforming an AVIRIS image cube into reflectance index and leaf type maps and using them as inputs in purely statistical classification routines, e.g. maximum likelihood, functional vegetation types in the boreal forests of Canada have been correctly classified with success rates up to 53.6%. These results are higher than the TM classification, used as a parameter in numerous flux models in the BOREAS science project. On the other hand, they are lower than those reported by Zarco-Tejada and Miller (In press). However, these two new classification methods are able to classify all major vegetation classes present in the study area, with the exception of standing water. Additionally, it should be noted that our accuracy assessments were based on the SERM vegetation maps, which were derived from field observation and air photography, with their own inherent errors.

The initial results presented here are indicative that pigment and water content expressions can detect functionally significant features in the landscape. Further investigation is focusing on developing more fundamental, quantitative derivations of pigment and water content at the landscape level for the purpose of modeling carbon and water vapor fluxes.
 
 

Gamon J.A. , Surfus J. S. In press. Assessing leaf pigment content and activity with a reflectometer. New Phytologist.

Gao, B-C. 1996. NDWI – A normalized difference index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment, 52:155-162.

Gitelson, A., and M.N. Merzlyak. 1996. Detection of red edge position and chlorophyll content by reflectance measurements near 700 nm. Journal of Plant Physiology 148: 501-508.

Green, R.O., Roberts, D.A., and Conel, J.E. 1996. Characterization of the atmosphere and iversion of AVIRIS calibrated radiance to apparent surface reflectance. in Summaries of the 6th Ann. Jet Propulsion Laboratory's Airborne Earth Science Workshop, R.O. Green (ed.). JPL 96-4, Vo. 1: 135-146, Pasadena, CA.

Hall, F. G., D. E. Knapp, and F. Huemmrich. 1997. Physically based classification and satellite mapping of biophysical characteristics in the southern boreal forest, J. Geophys. Res., 102, D24, 29569-29580.

Peñuelas, J. et al. 1997. Estimation of plant water concentration by the reflectance Water Index WI (R900/R970). International Journal of Remote Sensing 18: 2869-2875.

Richards, J. A., Remote Sensing Digital Image Analysis, Second Edition, Springer-Verlag, Germany, 1994.

Sellers, J. P. et al. 1997. BOREAS in 1997: Experiment overview, scientific results, and future directions, J. Geophys. Res., 102, D24, 28731-28769.

Zarco-Tejada, P. J. and J. R. Miller. In press. Land cover mapping at BOREAS using red edge spectral parameters from casi imagery. Journal of Geophys. Res.
 
 

We would like to thanks John S. Surfus and Miriam Schimdts for their help in the processing of the imagery used in this study. This research was conducted under a BOREAS NASA grant.