Geographic stacking: Decision fusion to increase global land cover map accuracy

Abstract

Techniques to combine multiple classifier outputs is an established sub-discipline in data mining, referred to as “stacking,” “ensemble classification,” or “meta-learning.” Here we describe how stacking of geographically allocated classifications can create a map composite of higher accuracy than any of the individual classifiers. We used both voting algorithms and trainable classifiers with a set of validation data to combine individual land cover maps. We describe the generality of this setup in terms of existing algorithms and accuracy assessment procedures. This method has the advantage of not requiring posterior probabilities or level of support for predicted class labels. We demonstrate the technique using Landsat based, 30-meter land cover maps, the highest resolution, globally available product of this kind. We used globally distributed validation samples to composite the maps and compute accuracy. We show that geographic stacking can improve individual map accuracy by up to 6.6%. The voting methods can also achieve higher accuracy than the best of the input classifications. Accuracies from different classifiers, input data, and output type are compared. The results are illustrated on a Landsat scene in California, USA. The compositing technique described here has broad applicability in remote sensing based map production and geographic classification.