Estimating and mapping forest variables of the Brazilian National Forest inventory to finer scales.

Abstract

Introduction: Brazil recently started its National Forest Inventory (NFI) following a broad-scale sampling of 1:100,000. Therewith, there is an increasing need for disaggregating forest variables at a more precise, finer scale. Objective: With assistance of remotely sensed data, the aim of this study was to test two methods for estimating and mapping, at a fine scale, forest components as wood volume, biomass and carbon. Methods: The analytical procedure has seven steps. First: collecting tree variables; NFI plot data from three meso-regions of the State of Parana were used. Second: forest modelling; necessary for estimating tree-level variables, as stem volume, tree biomass and carbon. Third: calculation of NFI’s sampling error and confidence interval (CI). Fourth: satellite imagery pre-processing; seven Landsat-8 OLI bands were used, therefore Digital Number (DN) was transformed into surface reflectance (physical variable). Fifth: imagery classification; object-oriented classifications were performed for delimiting classes of (i) early successional, and (ii) mid- to-late successional forest. Sixth: pixel-level modeling; two fine-scale methods for estimating forest variables at pixel-level were tested. The first of them uses a Surface Reflectance Model (SRM), which is a linear model that fits wood volume as function of surface reflectance from Landsat-8 OLI data, including Red, NIR, SWIR1, SWIR2, Brightness, and NDVI. The stepwise method was used to select the reflectance variables. Principal components of the selected variables were used as input variables. The second method is based on the regressionkriging (RK) geostatistical technique, in which latitude and longitude were the only two input variables. The first method (with SRM) was applied to wood volume, biomass and carbon. The second one (with RK) was applied to wood volume only. Seventh: Wood volume was extrapolated at coarse (without remote sensing techniques) and fine (with such techniques) scales, and then compared with each other. Results: The SRM method had a performance better than the RK method, though both of them provided, in general, similar results. The SRM estimated a mean of 123.8 m³ ha-1 and a total volume for the study area of 210,961,589 m³. The RK estimated a mean of 115.7 m³ ha-1 and total of 203,326,674 m³, i.e., value 3.7 % lesser than the SRM estimates. In relation to forest biomass, the SRM estimated a mean and total of, respectively, 120.1 Mg ha-1 and 217,736,862 Mg. For carbon, the mean was of 49.2 MgC ha-1 and total of 89,272,113 MgC, corresponding to 327,331,082 Mg of CO2-equivalent stocked in the study area. The total wood volumes estimated by both fine-scale methods fit into the CI obtained at a coarse scale. Conclusion: Despite SRM and RK were capable of producing overall accurate estimates of wood volume, SRM better represented spatially the wood volume. SRM also estimated accurately forest biomass and carbon.