GeoPython2019

»Machine Learning for Land Use / Land Cover Statistics of Switzerland«
2019-06-25, 09:15–09:45, Room 1

This project demonstrates a powerful prototype to classify land use / land cover statistics for an entire country using a Deep Learning approach for aerial imagery processing and a Random Forest architecture for data fusion with time series and other auxilliary datasets.

The Swiss land use / land cover (LULC) statistics ("Arealstatistik") is produced by the Swiss Federal Statistical Office and is an important instrument for long-term spatial observation. The statistics have been collected periodically since the 1980s and have been based on the same survey method: aerial photographs of Switzerland are overlaid with a regular grid of 100x100 meters and a team of trained interpreters determines the land cover and land use classes at each grid intersection. The final product contains more than 4 million sample points.

The area statistics are produced with great personnel effort in periodic intervals of 9 years. The goal of the project ‘AI pilot for area statistics’ is to partially automate the interpretation task using ML methods. The developed prototype employs CNN-based supervised learning for land use/cover classification of aerial images. The CNN output is fused with various auxiliary data (cadastral survey information, altitude, satellite-derived time series etc.) and classified using the Random Forest method. The obtained results show great potential of the proposed approach in partial automation of data interpretation.