Remote detection of forest biophysical parameters in Northern Minnesota using a multi-sensor approach

Insects and disease affect large areas of forest in the U.S. and Canada. Understanding the ecosystem impacts of such disturbances requires knowledge of host species distribution patterns on the landscape. I used partial least-squares (PLS) regression and an automated variable reduction process with spectrally coarse, but temporally rich, Landsat sensor data to estimate and map the distribution and abundance of host tree species for the two dominant forest insects: spruce budworm and forest tent caterpillar. The analysis yielded estimates for total forest basal area (BA) with R2 = 0.62 and RMSE = 4.67 m 2ha-1. Estimates of relative BA for Abies balsamea (R2 0.64, RMSE 6.08 m2ha-1 ) and Picea spp (R2 0.88, RMSE 12.57 m2ha-1) were encouraging.The technique was extended to estimate detailed forest structure using 5 and 10 m SPOT imagery based on raw SPOT bands, common derivatives, and neighborhood metrics calculated from images of forest cover at the two resolutions. Resulting pixel-wise estimates of hardwood and coniferous forest canopy diameter (R 2 = 0.82 and 0.93, RMSE 0.62 and 0.47 m), bole diameter (R 2 = 0.82 and 0.90, RMSE 2.92 and 3.75 cm), height (R2 = 0.69 and 0.92, RMSE 1.27 and 1.59 m), crown closure (R 2 = 0.52 and 0.68, RMSE 5.49 and 6.02%), live crown (R2 = 0.58 and 0.81, RMSE 0.96 and 1.25 m), and BA (R2 = 0.71 and 0.74, RMSE 2.47 and 4.58 m2ha-1) demonstrate that multi-resolution imagery can be used to achieved estimates of forest structure that approach or exceed lidar-level accuracy.Finally, synthetic aperture radar (SAR) data (Radarsat-1 and PALSAR), Landsat data, and the SPOT-5 structure results were combined to produce a single PLS regression model to estimate the relative BA of all major forest species simultaneously. Precision among estimates was higher using multiple sensor data than Landsat data alone. In fact, SPOT structure and SAR image variables were among the dominant predictors for most species. Thus, PLS regression has proven to be an effective modeling tool for regional characterization of forest structure within spatially heterogeneous landscapes using spectrally coarse satellite sensor data.