Multi-sensor data fusion for estimating forest species composition and abundance in northern Minnesota

Abstract

The magnitude, duration, and frequency of forest disturbance caused by the spruce budworm and forest tent caterpillar in northern Minnesota and neighboring Ontario, Canada have increased over the last century due to a shift in forest species composition linked to historical fire suppression, forest management, and pesticide application that has fostered increased dominance of host tree species. Modeling approaches are currently being used to understand and forecast potential management effects in changing insect disturbance trends. However, detailed forest composition data needed for these efforts is often lacking. We used partial least squares (PLS) regression to integrate different combinations of satellite sensor data including Landsat, Radarsat-1, and PALSAR, as well as pixel-wise forest structure information derived from SPOT-5 sensor data (Wolter et al., 2009), to determine the best combination of sensor data for estimating near species-level proportional forest composition (12 types: 10 species and 2 genera). Single-sensor and various multi-sensor PLS models showed distinct species-dependent sensitivities to relative basal area (BA), with Landsat variables showing greatest overall sensitivity. However, best results were achieved using a combination of data from all these sensors, with several C-band (Radarsat-1) and L-band (PALSAR) variables showing sensitivity to the composition and abundance of specific species. Pixel-level forest structure estimates derived from SPOT-5 data were generally more sensitive to conifer species abundance (especially white pine) than to hardwood species composition. Relative BA models accounted for 68% (jack pine) to 98% (maple spp.) of the variation in ground data with RMSE values between 2.46% and 5.65% relative BA, respectively. Receiver operating characteristic (ROC) curves were used to determine the effective lower limits of usefulness of species relative BA estimates which ranged from 5.94% (jack pine) to 39.41% (black ash). These estimates were then used to produce a dominant forest species map for the study region with an overall accuracy of 78%. Most notably, this approach facilitated discrimination of aspen from paper birch as well as spruce and fir from other conifer species which is crucial for the study of forest tent caterpillar and spruce budworm dynamics in the Upper Midwest. We also demonstrate that PLS regression is an effective data fusion strategy for mapping composition of heterogeneous forests using satellite sensor data.