Heterogeneous Data Mining of Earth Observation Archives: Integration and Fusion of Images, Maps, and In-situ Data

Kurzfassung

The amount of Earth Observation (EO) data is in constant growth due to the proliferation of EO missions in space and the continuous evolution of their instruments. The catalog of EO products is enriched by the high diversity of the imaging sensors. Along with the imagery data, EO products accommodate different metadata containing several parameters related to the image, the satellite and the instrument. In addition, we can also consider as EO information the data derived from third party systems not directly related with satellite EO products. An example are the widely spread Geographic Information Systems (GIS), which store map information that can be used for different purposes during EO image analysis. In this heterogeneous Big Data scenario, the main challenge is not only to provide better and more efficient algorithms, but also to design and implement tools that allow a greater exploitation of the available information. In line with the challenge, this thesis focuses on the integration, mining and exploitation of a wide range of EO heterogeneous data in order to efficiently extract valuable information for a better understanding of EO image content. The presented Heterogeneous Data Mining (HDM) system prototype overcomes the limitations of previous systems by including multispectral images, Synthetic Aperture Radar (SAR) images, and digital maps in an accelerated active learning algorithm. The learning stage of the algorithm is based on naive Bayes Classifiers which make use of posterior probabilities of a user-defined semantic label given a query image. This accelerated algorithm opens new ways for knowledge-driven information mining systems to Big Data scenarios. In conjunction with the learning algorithm, the HDM concept also contains a probabilistic search method based on the distances between the elements being used for the calculation of the posterior probabilities and image Bag of Words (BoW) in the database. Additionally, a multilayer system for heterogeneous geospatial data analytics is introduced. The system manages data from the source and performs several transformations in order to enable the integration of remote sensing, cartographic and in-situ data. Specifically, we use as in-situ data the results from the Land Use/Cover Area frame Survey (LUCAS). This survey monitors the state and change dynamics in land use and cover in the European Union. The system is tested in different scenarios and used for the development of a data mining methodology to filter and validate land cover changes recorded in multitemporal in-situ surveys. Our final effort focuses on the visual exploitation of the integrated heterogeneous EO data. By combining the results obtained from automatic analysis methodologies with interactive visualization tools, one can navigate and understand the EO data more efficiently.