OPTIMIZATION OF PARAMETER SETTINGS FOR MULTILEVEL IMAGE SEGMENTATION IN GEOBIA

摘要

In geographic object-based image analysis (GEOBIA) segmentation plays a key role since all consequent classification steps depend on its quality. For evaluating the quality empirical discrepancy and goodness methods are available. However, when optimizing parameter settings certain criteria have to be fulfilled, this in order to effectively facilitate the first step in GEOBIA. Besides being objective by making use of statistical methods the approach applied should optimize per class or group of classes, account for all degrees of freedom, use small parameter increments, and be preferably automated. The study presented in this paper combines a newly introduced 'area fitness rate' as a discrepancy method with an 'objective function' as a goodness method. By fulfilling the above criteria it is, amongst others, possible to follow the multilevel approach. Very similar results obtained for five different study sites demonstrate that the parameter settings found are indeed optimized. A visual check strengthens this finding further.