This study integrated field-based FIA forest inventory data and Landsat 8 imagery to model and map forest characteristics to inform restoration activities in the critically endangered longleaf pine (Pinus palustris) ecosystems of the southeastern United States. U.S. Forest Service Forest Inventory and Analysis (FIA) field plot data were related to normalized Landsat 8 imagery using general linear regression and machine learning models. Multiple raster surfaces of forest metrics, including basal area (ft2 ac-1), stand density (trees ac-1), forest type and occurrence of longleaf pine, were created for a 28.8 million acre study area in Florida, Georgia and Alabama. Specific methods to improve modelled relationships over previous studies came from the use of forest type results to focus basal area and stand density estimates on areas where tree species groups of interest were present. In addition, raster surface cells outside the range of predictive LANDSAT variables were not estimated, eliminating model extrapolation and exposing a challenge of using FIA data in this way.
Results/Conclusions
Basal area and stand density linear model results demonstrate a strong relationship between Landsat 8 imagery and FIA field plot data. The sixteen linear models of basal area and stand density covering our study area have R2 values ranging from 0.315 to 0.543, with an average of 0.421. The overall accuracy of the eight softmax neural network forest type models ranged between 0.78 and 0.87, averaging 0.83.
In total this project successfully produced 35 moderate resolution (30m) spatially explicit raster surfaces of estimated forest type, longleaf occurrence, basal area and stand density. These raster surfaces can be integrated with existing vector, tabular and raster datasets to inform and prioritize longleaf pine conservation and restoration efforts in the study area.