A framework for creating and validating a non-linear spectrum-biomass model to estimate the secondary succession biomass in moist tropical forests

Abstract

Uncertainties remain in the use of remote sensing technologies to provide validated model-derived estimates of the biomass of the secondary succession (SS) forests in the Amazon Basin. The objectives of this study were to develop a modeling framework for creating a valid spectrum-biomass model to estimate the SS biomass, to assess the utility of the framework and the accuracy and validity of the model, and to identify the model’s determinants. Data sources for this study include 1992–1993 vegetation inventory data and 1991 Landsat Thematic Mapper (TM) data on the Altamira region of Para, Brazil, and 1994–1995 vegetation inventory data and 1994 Landsat TM data on the nearby Bragantina region. The allometric approach was used to estimate the biomass of the sampled sites based on the vegetation inventory data. A framework for the spectrum-biomass regression model was developed based on the estimated biomass of the sampled sites and the Landsat data. The framework includes (1) the pooling of data from Bragantina and the use of ANCOVA to justify this approach; (2) image calibration; (3) biomass data age-adjustment, (4) selection of independent variables, (5) regression model development, and (6) model assessment and validation. The cubic regression model with TM Band5-related predictors was found to best fit the data as evidenced by an adjusted R-squared value of 0.865, mean square error (MSE) of the model, and the analysis of residuals. Residual analysis showed that the model might yield a better estimation on a lower biomass values than on higher biomass values. In addition, further analyses identified several determinants that can impact the accuracy of the spectrum-biomass model. ANCOVA confirmed that the relationship between the biomass and the spectrum is independent of the Altamira and Bragantina regions, and that it was appropriate to pool sampled data from both regions in the proposed model. The model development methodology and the model produced from this research will be of value to researchers using the spectrum-biomass modeling approaches to estimate the biomass and study the SS rates in moist tropical forests.