Land Classification Using Deep Neural Networks Applied to Satellite Imagery Combined with Ground-Level Images

With the exponential growth of georeferenced labeled and high-detailed imagery information of our world available on the web, combined with both computational power and the algorithms improvements from the last years, the land classification task becomes more reliable from the start, and thus it is possible to achieve sharper results when applying deep learning techniques to that imagery. Besides using exclusively high resolution data to train deep learning models, we can now apply it to terrain elevation models in order to have a more precise information about the visibility areas of a picture to better classify the land. Mapping the multiple sources of human activity involved in land or minimizing natural hazards, are two of many applications of a Geographic Information System (GIS), and our geographic knowledge discovery benefits from having this kind of information. The novel approach, behind the idea of my M.Sc. thesis of mapping land classification, relies on the usage of both satellite and ground-level images combined, applying visibility analysis and viewshed techniques, with the finality to determine and map the usage and coverage of the land to where those images belong.

In this context, the work that is expected covers :

• Data processing
• Implement Land Cover & Land Use Classification Model
• Visibility Analysis
• LC and LU Mapping
• Research Paper Submission for the 27th ACM CIKM
• Evaluation
• M.Sc. Dissertation

Supervision: The work will be co-supervised by Prof. João Moura Pires, from FCT/UNL, Prof. Brun Martins, IST/UL

Hosting Institution: IST/UL