A Multivariate Approach for Using Satellite Imagery to Map the Composition and Structure of Forests Susceptible to Insect Disturbance: Application to the Simulation of Carbon Dynamics in Northern Minnesota and Ontario
Abstract
Compared to other forest disturbances, insects and disease influence the largest area of forests in both the U.S. and Canada, affecting an estimated 50 million acres in the U.S. with economic costs over $1.5 billion. The successful understanding and modeling of ecosystem impacts of insect disturbances (especially for carbon dynamics) requires good knowledge of the spatial distribution, density and structure of host species on the landscape. In this study, we mapped the distribution of host species for the spruce budworm ( Choristoneura fumiferana) to facilitate landscape scale planning and modeling of outbreak dynamics. Spruce budworm is one of the most destructive indigenous pests in sub-boreal and boreal spruce-fir forests in the United States and Canada. Although periodic outbreaks are part of the natural cycle in these forests, traditional forest management practices may be responsible for increasing the frequency and severity of outbreaks. Currently, accurate spatially explicit forest structure data for such endeavors remains a persistent challenge and considerable research has focused on using remote sensing to identify methodologies to facilitate accurate estimation of stand volume and/or biomass. We used multi-temporal, multi-seasonal Landsat data and over 230 ground truth plots (and 220 additional validation plots) to map basal area (BA), for over two million hectares of forest in northern Minnesota and neighboring Ontario. BA was mapped both overall and for two spruce budworm host tree species ( Picea glauca and Abies balsamea) using partial least squares (PLS) regression applied to raw spectral bands, various spectral derivatives, and ground truth data. Results of the PLS regression yielded reasonable estimates of overall forest BA with an adjusted R2 of 0.62 and RMSE 4.67 m2 ha-1. White spruce relative BA had an adjusted R2 of 0.88 (RMSE 12.57 m2ha-1) and balsam fir relative BA had an adjusted R2 of 0.64 (RMSE 6.08 m2ha-1). The method also produced estimates for proportional cover of deciduous and evergreen species, with each having adjusted R2 values of 0.86 (RMSE 9.89 and 9.78 m2ha- 1, respectively). Because ground based measurements were placed largely in forest stands containing spruce and fir, modeled results show considerable confusion with non-target conifers, such as pines and cedar. Research is currently aimed at improving results by expanding ground-based measurements to include more non-target forest types to strengthen models. PLS regression has proven to be an effective data fusion tool for regional mapping of forest structure within spatially heterogeneous forests. Ongoing research is aimed at including other large-format sensors, such as Radarsat, to expand the capacity for modeling regional forest structure.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2007
- Bibcode:
- 2007AGUFM.B43C1443T
- Keywords:
-
- 0430 Computational methods and data processing;
- 0439 Ecosystems;
- structure and dynamics (4815);
- 0476 Plant ecology (1851);
- 0480 Remote sensing