Optimal segmentation of high spatial resolution images for the classification of buildings using random forests

In the application of machine learning to geographic object based image analysis, several parameters influence overall classifier performance. One of the first parameters is segmentation size—for example, how many pixels should be grouped together to form an image object. Often, trial and error methods are used to obtain segmentation parameters that best delineate the borders of real world objects. Several attempts at automated methods have produced promising results, but manual intervention is still necessary. Meanwhile, numerous measures of segmentation quality have been defined, but their relationship to classifier performance is not then directly shown. For example, as measures of segmentation quality improve, do classification results improve as well? Our work considers the problem of building classification in high resolution aerial imagery of urban areas. Based on user defined training polygons generated with or without a reference segmentation, we have found several measures of segmentation quality and feature performance that can help users narrow the range of appropriate segmentations. Furthermore, our work finds that given this range, performance of machine learning algorithms remains relatively constant for any given segmentation as long as features used for classification are chosen correctly. We find that the range of scale parameters capable of producing an accurate classification is much broader than typically assumed and trial and error methods for finding this parameter may be an acceptable approach.