Land use describes the socio-economic function of an area of the earth surface (e.g. settlement, agricultural, water bodies etc.). Normally, a land use object consists of many different land cover elements. Land cover is related to the physical material of the earth surface, e.g. building, asphalt, grass, tree. Commonly, the information of land use is collected and represented in the form of polygons in geospatial databases, often acquired and maintained by national mapping agencies. For applications such as urban management and planning, up-to-date land use information is of high importance. However, the rapid development, e.g. of cities, makes the geospatial databases outdated quickly. Therefore, there is a demand for the automatic verification of geospatial databases. This can be achieved by comparing the database content to current remote sensing data to check whether the information contained in the database is still correct. Errors thus identified can then be corrected. This thesis deals with the automation of the database verification process. Due to the fact that the object catalogue of geospatial databases is frequently constructed in a hierarchical manner, this thesis tries to predict land use in multiple semantic levels hierarchically and simultaneously. To achieve that goal, a hierarchical deep learning framework is proposed which applies a two-step strategy. First, given high-resolution aerial images, the land cover information is determined. To achieve this, an encoder-decoder based convolutional neural network (CNN) is proposed. Second, the pixel-wise land cover information and the aerial images serve as input for another CNN to classify land use. To guarantee consistency of the classification results with the class hierarchy, two strategies are proposed. The first one is called joint optimization (JO). Using this strategy, predictions are made by selecting the hierarchical tuple over all levels which has the maximum joint class score, providing consistent results across the different levels. The second strategy is to use the predictions at the finest level to control the predictions at the coarser ones, which is called fine-to-coarse (F2C). To evaluate the performance and to investigate the strengths and limitations of the proposed methods, extensive experiments are conducted using five datasets: Hameln, Schleswig, Mecklenburg-Vorpommern (MV), Vaihingen and Potsdam. In terms of land cover classification, the proposed CNN achieves overall accuracies between 85% and 90% when dealing with class structures of 6 to 10 land cover classes. Using the Vaihingen and Potsdam datasets, the proposed CNN achieves results on par with state-of-art methods, but only requires 1% of unknown parameters compared to these methods. In terms of hierarchical land use classification, both the JO and F2C strategies achieve very good results. At the coarsest level, in which only four classes are differentiated, the overall accuracy can reach about 95%. As the semantic level increases, i.e. with an increasing number of classes to be discerned, the categories are more and more difficult to be correctly differentiated. In terms of overall accuracy, a drop of 10% - 15% between level Ⅰ and level Ⅱ (14 classes), and of about 5% between level Ⅰ and level Ⅱ (21 classes) is observed. Furthermore, it is found that object size has an impact on the classification. For large objects, the classification accuracy is higher than the one for small polygons.
展开▼
