Incorporating LIDAR and Landsat-based disturbance metrics into fine-grain nearest-neighbor imputation maps of vegetation composition and structure

Background/Question/Methods

Spatially explicit information (i.e. maps) of forest vegetation composition and structure are needed at fine spatial scales across large spatial extents for resource management, ecological research, monitoring, and policy formulation. Nearest-neighbor (NN) imputation is an effective method integrating forest inventory plots with spatial predictors such as climate, topography, and satellite imagery to map detailed forest attributes across large areas. NN imputation maps have been used in many regions, but have focused on mid- to large-scale applications in forest planning and policy. We incorporated precise GPS geo-referencing of forest inventory plots with spatial predictors including Light Detection and Ranging (LiDAR) and Landsat-based disturbance metrics to develop fine-scale NN imputation maps of forest vegetation composition and structure for ~500,000 ha of the Deschutes National Forest in eastern Oregon, USA. Our primary research questions asked: 1) what combinations of spatial predictors (i.e. gridded climate data, topographic data, tasseled cap bands from Landsat TM, disturbance onset, magnitude, and recovery metrics derived from multi-date Landsat TM, and LiDAR derived metrics of vegetation structure) best predict vegetation composition and structure?, and 2) does decreased grain size (10m, 30m, 90m cell size) of spatial predictors and or sampling grain (subplot versus plots) improve NN imputation accuracy?

Results/Conclusions

NN imputations were conducted by relating field plots to spatial predictors using constrained ordination (CCA), and vegetation attributes from the nearest plot in CCA space were imputed to mapped pixels. Assessments of NN imputation models were based on variation explained in CCAs and multiple accuracy diagnostics (error matrices, kappa coefficients, RSMD, differences in cumulative distribution functions, mapped differences in modeled area estimates, etc.) Species composition was best described by a combination of Landsat TM tasseled cap bands, climate, and topography. LiDAR structural metrics and Landsat TM disturbance metrics did not greatly improve accuracy of species composition models, nor did decreased grain size of spatial predictors or field plots. LiDAR derived structural metrics and Landsat TM disturbance metrics in combination had the best predictions of live vegetation structure, while Landsat TM disturbance metrics alone best predicted dead vegetation structure (i.e. snags). Decreasing grain size of spatial predictors and field plots improved accuracy of vegetation structure predictions. Results suggest fine-scale NN imputation models of vegetation composition and structure can be achieved with sufficient predictive accuracy to be useful for a wide range of natural resource applications where stand-level spatially explicit data is required across large regions.