Krahwinkler M.P. Machine learning based classification for semantic world modeling: support vector machine based decision tree for single tree level forest species mapping

RWTH Aachen University, 2013. — 355 p.Robotic applications more and more expand into unstructured terrains. The new applications require detailed automatically generated models of the environment. Semantic world models map the environment and the semantic meaning of the objects in a virtual model. These models and their connection to the real world allow for precise navigation based on landmarks as well as planning, simulation and control of robots in a virtual testbed and in the real world. Semantic world models are generated based on a classification that assigns each instance of the model to the appropriate category. Supervised classifiers, which were developed in the context of machine learning, are used for this process. Data mining allows for the analysis of the input data sources and the detection of features that are essential for the desired classifications. Image processing algorithms are used in the preprocessing of raster data. The combination of these methods with techniques that were developed in geosciences provides the basis for new applications in unstructured natural terrains. A high intraclass and a low interclass variability are a major challenge in the classification of natural environments like forests. An approach for object-based tree species classification at single tree level for large areas was developed in this study. A support vector machine based binary decision tree was created, which is characterized by high resolutions, large scale applicability, high flexibility, low demand for manual parameterization and short classification times. The influence of several configurations and input data specifications was analyzed and the results provide valuable information as decision support for future applications. The comparison with commonly used algorithms proves the strength of the developed method, which is especially suitable for hierarchical classifications and can also be used in spectral analysis, letter detection or medical imaging.