Remote sensing can provide a relatively low-cost approach to large scale assessment of forest productivity which enables investigations into spatial and temporal trends in forest growth. However, the connection between remote sensing products and scalable field metrics is not well understood. Much of the existing research has focused on homogeneous, single species forests, with limited remote sensing inputs for model calibration or field data to assess accuracy of productivity predictions. To explore this approach across mixed northeastern forests, we compared models of annual basal area increment (BAI) using widely available remote sensing data products as well as ancillary spatial data layers to capture site, stand, and relative habitat suitability both across the landscape and for specific northeastern species. A stepwise multiple linear regression model was used to develop both full, and species specific BAI growth models based on dendrochronologies from 132 plots located across Vermont and New Hampshire, linked with 10 years of MODIS derived NDVI and NDWI metrics. Species that dominated at least 5 plots across all dendrochronology field sites were used for species specific calibrations. These species included: Abies balsamea, Acer saccharum, Betula alleghaniensis, Betula papyrifera, Fagus grandifolia, Picea rubens.
Results/Conclusions
We found that a landscape scale model for all species was not accurate, accounting for only 16 percent of the total variability in BAI, with an accuracy of only 46%. However, when individual species were modeled, accuracy and stability increased significantly. A 3-term model was able to predict BAI on plots dominated by Betula papyrifera, with an r2 = 0.69, RMSE = 1.353 and average accuracy of 61%. Picea rubens dominated plots were modeled using 6 terms with an r2 = 0.59, RMSE = 4.18 and average accuracy of 62%. A 7-term model on Acer saccharum dominated plots predicted BAI with an r2 = 0.33, RMSE = 4.47 and average accuracy of 54%. These results indicate that modeling forest productivity across a heterogeneous landscape is difficult based on the complexity of spectral characteristics in mixed stands, variability across the landscape and diversity of factors influencing tree growth on a micro-scale. However, species specific models can be built and applied across these landscapes with sufficient accuracy to inform spatial and temporal patterns in forest productivity.