Adapting machine learning methods for coarse resolution land cover classification

Land cover influences numerous biogeophysical processes at the interface of the land surface and atmosphere. Increasingly, models are being used to better understand the global climate system. Such models require accurate characterization of land cover at global and regional scales. To provide more accurate maps of land cover, classification of coarse resolution remote sensing data is widely used. Although classification techniques have been commonly applied to remotely sensed data in the past, classification problems currently being addressed by the global remote sensing community pose new challenges. To meet these challenges, nonparametric machine learning algorithms are being used. However, these algorithms are less well understood for remote sensing applications than conventional approaches.; The main objective of this dissertation is to improve understanding of nonparametric classification algorithms in remote sensing applications, and by extension, to improve global land cover mapping techniques. The issues examined address identifying and reducing uncertainty in coarse resolution land cover maps produced using nonparametric classification algorithms. The research draws heavily on a new algorithm known as boosting, which is used in conjunction with supervised classification algorithms. This dissertation provides four main results. First, pixel-scale uncertainty measures derived using boosting are reliable indicators of classification error, as demonstrated using several validation data sets. As a result, spatially explicit estimates of the likelihood of misclassification can be provided with classified maps. Second, a Bayesian approach for incorporating ancillary information with nonparametric classification algorithms improves classification accuracy. This approach reduces the sensitivity of classification predictions to the class frequency distribution of the training sample. Third, estimates of sub-pixel cover from several machine learning algorithms are more accurate than those produced using linear spectral mixture analysis because the nonparametric algorithms are better able to accommodate within-class spectral variability. Finally, comparison of four currently available global land cover products, including one produced using techniques developed in this dissertation, against validation data from the United States Forest Service for California shows systematic overprediction of forest cover. Future global mapping efforts will benefit from identifying problems with accurate, high resolution validation data.