Land Cover Classification in Combined Elevation and Optical Images Supported by OSM Data, Mixed-level Features, and Non-local Optimization Algorithms

摘要

Land cover classification from airborne data is considered a challenging task in Remote Sensing. Even in the case of available elevation data, shadows and strong intra-class variations of appearances are abundant in urban terrain. In this paper, we propose an approach for supervised land cover classification that has three main contributions. Firstly, for the cumbersome task of training data sampling we propose an algorithm which combines the freely available OpenStreetMap data with the actual sensor data and requires only a minimum of user interaction. The key idea of this algorithm is to rasterize the vector data using a fast segmentation result. Secondly, pixel-wise classification may take long and be quite sensitive to the resolution and quality of input data. Therefore, superpixel decomposition of images, supported by a general framework on operations with superpixels, guarantees fast grouping of pixel-wise features and their assignment to one of four important classes (building, tree, grass and road). Particularly for extraction of street canyons lying in the shadowy regions, high-level features based on stripes are introduced. Finally, the output of a probabilistic learning algorithm can be post-processed by a non-local optimization module operating on Markov Random Fields, thus allowing to correct noisy results using a smoothness prior. Extensive tests on three datasets of quite different nature have been performed with two probabilistic learners: The well-known Random Forest and by far less known Import Vector Machine are explored. Thus, this work provides insights about promising feature sets for both classifiers. The quantitative results for the ISPRS benchmark dataset Vaihingen are promising, achieving up to 94.5% and 87.1% accuracy on superpixel and on pixel level, respectively, despite the fact that only around 10% of available labeled data were used. At the same time, the results for two additional datasets, validated with the autonomously acquired training data, yielded a significantly lower number of misclassified superpixels. This confirms that the proposed algorithm on training data extinction works quite well in reducing errors of second kind. However, it tends to extract predominantly huge and easy-to-classify areas, while in complicated, ambiguous regions, first type errors often occur. For this and other algorithm shortcomings, directions of future research are outlined.