Improving the Efficiency of Land Cover Classification by Combining Segmentation, Hierarchical Clustering, and Active Learning

Acquiring training samples for supervised machine learning methods to automate land cover classification is very labor intensive. The ever-increasing amount of available remote sensing data makes it necessary to reduce these costs. By adding a segmentation approach to hierarchical clustering and active learning, we create a new and unique method for land cover classification. In the segmentation stage, we adapt the simple linear iterative clustering algorithm to use the spectral angle instead of the Euclidean distance. The bisecting k-means algorithm creates the hierarchical clustering. Finally, we employ an active learning strategy based on the active queries method. We further investigate the effects of incorporating local density information into the active learning scheme. Here, we use probabilistic active learning. In this framework, the sample's label and its posterior probability are modeled as two random variables and weighted with the local density to estimate the impact of the sample's future label. We conduct eight experiments to test the influences of the method's different parameters and stages. Evaluation is done on three aerial remote sensing datasets for urban and vegetational regions in Germany: Abenberg, Potsdam, and Vaihingen. The method is compared to three state-of-the-art methods. Our results show that we are able to achieve the same overall accuracy while decreasing the amount of needed training samples by 95%.