Geospatial data automatic processing and its final representation using machine learning techniques

Resumen

Geospatial data can be obtained from different sources and devices. From photogrammetry, to laser scanners, moving through depth sensors and satellite imagery. The inclusion of affordable high precision and accuracy devices led to the incorporation of this source of information to new industry sectors. But the amount of generated data by those devices can be sometimes overwhelming. Industry and users require a simplification of the geospatial data, sometimes by simply reducing the final representation and others by preprocessing it extracting significant information. This PhD thesis includes works proposing new methodologies which mix the machine learning and computer vision techniques to offer solutions to the final user and the industry necessities. As a starting point, some of the developed works consist in getting familiar with the geospatial data in the shape of 3D point clouds and it's analysis. As a follow up, this PhD dissertation proposes new techniques to help in the plan of the data acquisition and survey, with focus on saving time and costs while maintaining data quality. In addition the most relevant work comes with the automatic point cloud registration of cloud-to-cloud and cloud-to-model. Registration of geospatial data encompass the use of different methodologies since one of the most challenging features is to be able to obtain the alignment without the use of artificial targets or the use of any other sensor apart from the 3D geometric coordinates. Results were proved by conducting exhaustive tests in several datasets. All the techniques are suitable to obtain quality results suitable for the final user. The final solutions proposed in this research work can be easily applied to final applications to the user and the industry sector. The final representation is proved to be simplified by the automatic processing of the proposed methodologies allowing the final user to obtain results easily and more suitable to its needs. Furthermore, machine learning and computer vision techniques were validated along the whole PhD thesis, proving to be worth of consideration for the geospatial data processing sector.