CD-17 Abstract

Validating, Scaling and Parameterizing a Forest Regrowth Model for the Amazon Region Using Aircraft and Spaceborne Sensors and GIS

Diogenes Salas Alves, INPE/DPI (SA-PI)
Mark J Ducey, University of New Hampshire (US-PI)

This proposal is in response to the NASA-Carbon Cycle Science and LBA-Ecology Program request for remote sensing research and the development of methods and datasets for exploring the response of ecosystems to disturbance at the regional scale.  We propose a four-step, incremental approach directed toward the spatially explicit modeling and mapping of forest regrowth potential for the Amazon region. Each of the four steps will make a significant contribution to current understanding of the response of ecosystems to disturbance at the regional scale.  Developing an ability to predict forest regrowth potential has considerable implications for our understanding of carbon dynamics in a future characterized by increased conversion of old-growth Amazonian forests and the subsequent abandonment of many areas originally cleared for agricultural activities. A central focus of our approach is the development of remote sensing approaches for quantifying vegetation recovery and changes in biomass following disturbance, determination of the optimal scale for these approaches, and testing of disturbance-specific parameters that may influence rates of forest regrowth in Amazonia.

Step 1 is the production of preliminary forest regrowth potential maps for the region using an empirical model of biomass accumulation in global secondary forests (Johnson et al.2000; see Appendix B).  The maps are a spatially explicit application of model algorithms within a GIS framework constrained by readily available regional datasets on climate and soil texture.  Step 1 products include maps illustrating potential biomass accumulation in Amazonian regrowth forests at 5-year intervals (Zarin et al. submitted; see Appendix C).

Step 2 is the definition of a set of normalized spectral indices of forest regrowth optimized for the Amazon region.  These indices will be derived from MASTER and multi-temporal Landsat TM/ ETM+data and calibrated against field measurements of regrowth structure and estimates of canopy height and vertical distribution. The ideal resolution for these spectral indices will be quantified. Expected products from Step 2 include forest regrowth data bundles that will contain age and land use history for each stand, temporal sequences of regrowth indices, derived canopy height, canopy profiles and canopy structural properties for a subset of regrowth stands.

Step 3 is the testing of the reliability of the preliminary maps (Step 1 product) and the remote sensing indices of regrowth structure (Step 2 products). We will employ traditional validation methods to compare the spatially explicit model predictions with forest inventory data not included in the development of the model.  Validation is used here in a strictly statistical sense.  Using independent data, we will assess the bias and variance characteristics of the global model when used as a regional predictor, including partitioning the variance components into local (within-site) and regional (between-site) components.  And although the model was developed with reference to regrowth biomass, we will also test its applicability to the prediction of regrowth structure as expressed by the spectral indices defined in Step 2.  AIRSAR estimates of biomass will be tested and compared to assess between-site and seasonal differences in SAR based biomass retrievals. Expected products from Step 3 are spatially explicit accuracy assessment of regrowth maps for biomass and structure, based on both field and remotely-sensed data. 

Step 4 is the refinement of the global model to enhance its regional applicability by including known disturbance-specific parameters shown to explain a significant amount of variance between measured and modeled regrowth biomass and structure.  We will test parameters that describe disturbance type, size, perimeter:area ratio and land-use history, including the number of agricultural cycles and persistence of the various stages of vegetation cover.  Remote sensing based observations are required to adequately test the importance of these disturbance-specific parameters with a sufficient number of observations. A subset of the forest inventory data will be used to calibrate this regional enhancement of the model.  Model selection will employ cross-validation with reserved independent data, or resampling cross-validation techniques, to avoid overfitting and to assess expected prediction errors.  The resulting spatial pattern of predictions will also be compared to the spatial pattern of the spectral indices defined in Step 2. The principal products of Step 4 will be a regionally refined and accuracy-assessed forest regrowth potential model and derivative maps that will provide our best predictions (and corresponding error estimates) for the regional responses of biomass and forest structure to current and future land-use patterns in Amazonia.