Tropical forest area measured from global land-cover classifications: Inverse calibration models based on spatial textures

Abstract

Retrieving area estimates from broad scale land-cover maps is generally inaccurate due to the effect of spatial aggregation on class proportions. In a previous study, we tested a method to calibrate area estimates of tropical forest cover by inverting a model of the influence of the forest spatial fragmentation on the spatial aggregation bias, as characterized by two nested regression models. This was based on a sample of high resolution land-cover classifications, distributed across the tropical belt. In this study, improvements of this previous model are sought, first, by better accounting for the spatial variability of landscape characteristics using texture measures and, second, by integrating spatial information in the mixed pixel estimator—that is, the modeling of spectral mixtures at the scale of coarse resolution pixels as a function of the proportion of land-cover types. These improvements were tested using NOAA's Advanced Very High Resolution Radiometer data at 1.1 km resolution, Landsat Thematic Mapper-based classifications, and data simulated at the 250 m resolution of the forthcoming Earth Observing System's Moderate Resolution Imaging Spectroradiometer (MODIS). The integration of spatial information into a correction model to retrieve fine resolution cover-type proportions from coarse resolution data can improve by up to 35% the reliability of the estimates. The results also demonstrate that the integration of spatial information in the mixed pixel estimator controls for the variability due to different landscape characteristics. This study improves our capability to estimate tropical forest cover from coarse resolution remote sensing data at a global scale.