ForestERA - Vegetation modeling tools

home > tools > vegetation modeling

ForestERA Vegetation Modeling Tools

Information about forest composition and structure is essential to analyses of fire dangers and risks to wildlife habitat. Using remotely-sensed data, and reliable ground data, the ForestERA Project has produced data layers describing forest conditions over more than five-million-acres in the southwest. These layers — dominant overstory vegetation; basal area, tree stem density, and canopy cover — are the primary inputs for our wildlife and fire models, and underlie the landscape analyses that help managers and all stakeholders understand forest conditions, prioritize restoration efforts, and anticipate and mitigate for the likely effects of forest treatment and wildfire on watersheds and wildlife habitats. The descriptions that follow are for the Western Mogollon Rim (WMPALA) study area.

Dominant Overstory Vegetation

The dominant overstory vegetation layer is a map of the most common species that characterize a given area, and is based on the dominant woody vegetation. The nine categories of dominant overstory vegetation are: (1) open grasslands, shrublands, and barren areas, (2) ponderosa pine (Pinus ponderosa), (3) quaking aspen (Populus tremuloides), (4) mixed-conifer, including mixtures of spruces, firs, pines and/or Douglas fir (Pseudotsuga menziesii), (5) pinyon-juniper (Pinus edulis and Juniperus spp.), (6) juniper-dominated mix (with ponderosa and/or pinyon pines), (7) ponderosa pine/quaking aspen, (8) ponderosa pine/Gambel oak (Quercus gambelii), and (9) mixed-conifer/ quaking aspen. This layer was created using a classification tree methodology (Breimann et al. 1984) and has a pixel resolution of 90m (0.81 ha or 2 acres). To develop the layer, we used forest structural data from ground plots as training data, and derived predictive data layers from multi-temporal Landsat 7 Enhanced Thematic Mapper (ETM) satellite imagery and digital elevation models (DEMs). A ten-fold cross-validation assessment (a type of internal accuracy check, see glossary) of this layer indicated an overall accuracy of 87%, meeting standards of accuracy suggested by Thomlinson et al. (1999) for vegetation classifications derived from remote-sensing data.

Canopy Cover

Canopy cover is a measure of the proportion of ground area that is covered by a vertical projection of tree crowns. The canopy cover layer is reported as percent and has a pixel resolution of 30m (0.09 ha or 0.22 acre). The layer was developed directly from a mosaic of digital orthophoto quadrangles (DOQs), by applying a fractal concentration value-area method (Xu et al. 2006), similar to multifractal methods widely used in geology (Cheng et al. 1994). This methodology permits us to differentiate between tree crown, shadow, and non-crown areas, at high resolution, making it possible to derive highly accurate estimates of canopy cover over large areas. We assessed the accuracy of this layer using ground data collected on 200 plots (of 17.5m radius or 0.1 ha) at 18 sampling locations across the Western Mogollon Plateau focal area. Based on linear regression model with intercept constrained to zero, we found a statistically significant and unbiased relationship between the ground data estimates of canopy cover and the predicted canopy cover values in the GIS layer (d.f. = 199, r² = 0.545, P < 0.001, m = 1.02). Uncertainty analysis indicated that 90% of the predicted values for canopy cover were within 13% of their true value.

Basal Area and Tree Stem Density

Basal area, the total cross-sectional area of trees in a stand, and tree stem density, the number of trees per unit area, are commonly used measures of tree density. Both are important variables for predicting fire hazard and distribution of certain wildlife species. We considered only those trees/shrubs that had a diameter greater than 2.5cm in both the basal area (reported in m2/ha) and tree stem density layers (reported in units of stems/ha).

The basal area and tree stem density layers each have a pixel resolution of 90m (0.81 ha or 2 acres). Although the resolution of the original imagery was 30m (0.09 ha or 0.22 acres), we found that the accuracy of these layers was significantly improved when they were resampled to 90m. The layers were created using a regression tree methodology (Breimann et al. 1984), and a machine-learning algorithm known as “boosting” (Bauer & Kohavi 1999). We developed the layers using forest structural ground plots as training data, and derived predictive data layers from multi-temporal Landsat 7 Enhanced Thematic Mapper (ETM) satellite imagery and digital elevation models (DEMs).

Our accuracy assessment compared predicted basal area and tree density values in the GIS layers, to external data collected on nearly 600 plots at 63 sampling locations spread across the western Mogollon Plateau. Based on linear regression models with intercept constrained to zero, we found unbiased and statically significant relationships between values measured on the ground and the predicted values for those attributes in the GIS layers (basal area: d.f. = 62, r² = 0.508, P < 0.001, m = 1.09 and tree density: d.f. = 62, r² = 0.584, P < 0.001, m = 0.99). Uncertainty analysis indicated that 90% of the predicted values for basal area were within 9 m2/ha (40 ft2/acre) of their “true” value (as determined by ground measurement), and 90% of the predicted values for tree stem density were within 200 stems/ha (80 stems/acre) of their true value. We think that the application of advanced synthetic aperture radar (ASAR) data will improve the accuracy of our layers. We will be testing this approach in the near future.

Using High Resolution Spatial Imagery

We were interested in creating vegetation structure layers with greater accuracy and higher spatial resolution, so we developed a crown-delineation methodology for use with high-resolution multispectral imagery. Our preliminary tests, over a smaller study area of Anderson Mesa used imagery from the QuickBird satellite (0.7 m panchromatic; 2.4 m multispectral resolution). Results were promising (Hampton et al. 2003). We were able to obtain highly accurate estimates of the number of trees in a given area, the species of those trees, and the size of their crowns. Crown size is highly correlated with trunk diameter and tree height, so this methodology also can be used to estimate basal area and stand height. However, high-resolution imagery is expensive and the methodology is time-intensive, so, broader implementation of these new methods will require additional resources. We used this method to delineate all of the individual tree crowns within a single Quickbird image. We assessed the results using a set of measurements taken at 13 ground plots within the extent of that image. We identified and measured all trees within each ground plot, and then compared those data with the corresponding locations in the Quickbird image. For example, ground data indicated the presence of four ponderosa pines and 13 oaks within one of these plots. Using algorithms we had developed to delineate tree crowns and calculate stem density from the Quickbird imagery, we identified four ponderosa pines and 12 oaks within the same area. Similar results were obtained at the other 12 plot locations.

References

Anderson, M. C., J. M. Watts, J. E. Freilich, S. R. Yool, G. I. Wakefield, J. F. McCauley, and P. B. Fahnestock. 2000. Regression-tree analysis of desert tortoise habitat in the central Mojave Desert. Ecological Applications 10: 890-900.

Bauer, E. and R. Kohavi. 1999. An empirical comparison of voting classification algorithms: bagging, boosting, and variants. Machine Learning 36: 105-139.

Breiman, L., J. J. Friedman, R. A. Olshen, and C. J. Stone. 1984. Classification and Regression Trees. Wadsworth and Brooks Publishing, Monterey , California .

Dralle, K., and M. Rudemo. 1997. Automatic estimation of individual tree position from aerial photos. Canadian Journal of Forest Research 27: 1728-1736.

Franklin, S. E., A. Maudie, and J. Lavigne. 2001. Using spatial co-occurrence texture to increase forest structure and species composition classification accuracy. Photogrammatic Engineering and Remote Sensing 67: 849-855.

Foody, G. M. 2002. Status of land cover classification accuracy assessment. Remote-sensing of the environment 80: 185-201.

Hansen, M., R. Dubayah, and R. Defries. 1996. Classification trees: an alternative to traditional land cover classifiers. International Journal of Remote Sensing 17: 1075-1081.

Hyyppa, J., H. Hyyppa, M. Inkinen, M. Engdahl, S. Linko, and Y. Zhu. 2000. Accuracy comparison of various remote sensing data sources in the retrieval of forest stand attributes. Forest Ecology and Management 128: 109-120.

Thomlinson, J. R., P. V. Bolstad, and W. B. Cohen. 1999. Coordinating methodologies for scaling landcover classification from site-specific to global: steps toward validating global map products. Remote Sensing of the Environment 70: 16-28.

Xu, B., P. Gong, and R. Pu. 2003. Crown closure estimation of oak savannah in a dry season with landsat TM imagery: comparison of various indices through correlation analysis. International Journal of Remote Sensing 24: 1811-1822.