Exploring coarse-to fine-scale approaches for mapping and estimating forest volume from Brazilian National Forest Inventory data

Abstract

The aim of this study was to explore methods to: (1) enhance coarse-scale estimates of wood volume from National Forest Inventories (NFIs) data and (2) map them at finer scales. The 'standard' coarse-scale estimation extrapolates wood volume from clusters to the grid cell they represent, using the cluster's represented forested area (RFA) to predict the cell's forested area. Data from a subset of Brazil's NFI clusters were combined with Landsat-8 imagery to explore a new coarse-scale method, where forested area derived from image classification (FADIC) is used instead of RFA. The RFA- A nd FADIC-derived estimates of total volume were, respectively, 197.4 million m3 and 116.3 million m3. For fine-scale methods, volume was estimated and mapped at pixel level using: (i) surface reflectance-based models (SRMs), and (ii) regression-kriging (RK) and a RK model (RKM) whose inputs were latitude and longitude of pixels. The SRM-based method captured the mean and the general spatial trend of the volume well. The RK-based method also estimated the mean well, but it failed to predict higher and lower volumes. The SRM- A nd RK-based estimates of total volume were 211.7 million m3 and 203.3 million m3, an overestimate of 7 per cent and 3 per cent, respectively, of the 'standard' NFI estimate (197.4 million m3), though both estimates were within the 95 per cent confidence interval, meaning that both fine-scale methods yield total volume statistically similar to the 'standard' coarse-scale method.