Improving global land cover characterization through data fusion

Global-scale land cover characterization has advanced from a spatial resolution of 1 × 1° in the mid-1990s to 30 × 30 m resolution to date. However, some mapping challenges exist persistently regardless of the increasing spatial resolution. Data fusion has been proved as an effective way of improving land cover characterization. Here we applied a machine learning-based data integration approach for improving global-scale forest cover characterization. The approach employed six coarse-resolution (250–1000 m) global land cover maps as input and various regional, higher-resolution land cover data-sets as reference to build regression tree models per continent. The average error of 10-fold cross validation of the regression tree models varied between 7.70 and 15.68% forest cover and the r2 varied between 0.76 and 0.94, indicating the robustness of the trained models. As a result of data fusion, the synthesized global forest cover map was more accurate than any input global product. We also showed that other major vegetative land cover types such as cropland, woodland, grassland, and wetland all exhibit similar magnitude of discrepancies as forest among existing land cover maps. Our developed method, because of its type- and scale-invariant feature, can be implemented for other land cover types for improving their global characterization. The ensemble approach can also be internalized for improving data quality when generating a global land cover product, where multiple versions can be produced and subsequently integrated