What is the potential of Artificial Intelligence in combination with GIS?
|Datum:||07 mei 2019|
Case: classifying asbestos roofs in the Dutch province of Drenthe using hyperspectral imagery and deep learning
For my bachelor thesis under the supervision of dr. Marco Wiering (Department of Artificial Intelligence, RUG) and Klaas Dijkstra (NHL Stenden University of Applied Sciences), I studied the possibilities of deep learning on hyperspectral imagery in Drenthe. The specific task at hand was the classification of asbestos roofs in the province of Drenthe, yet further applications of combining deep learning with geospatial data are also intriguing, for the overarching concept allows sufficient possibility for solving an abundance of use cases. For this specific case a pixel-wise segmentation method was developed in which the buildings were cut out of the dataset using the BAG (Basisregistratie Adressen en Gebouwen; open overheidsdata) and then fed to the deep neural networks. A comparison was made between a modified version of U-Net (with a reduced depth to be able to better handle the small building cutouts) and DeepLabv3+ with ResNet-101 as a backbone.
Using deep learning for the detection of objects or classification of materials on the basis of aerial imagery is a compelling use case. When enough ground truth data (of adequate quality) are available the networks can be sufficiently trained for the task at hand. However, if no ground truth data are available these have to be obtained first, which could mean that there is still plenty of fieldwork to be done. In optimal cases the ground truth can be built on top of an existing dataset together with one of the Dutch basisregistraties.
In this specific study it was found that it was quite hard to classify asbestos roofs but that the method itself and the pipeline used to work with the geodata is very promising. The fact that it was able to work with a large amount of data for a large geographical area (5.2TB for an entire province) is very promising for future applications. The biggest difficulty in the project were the material itself. The corrugated asbestos roofing sheets often contain only about 12% asbestos. That means that the remaining 88% of the sheet is cement, as the asbestos is bound into this material.
With respect to the results it was found that DeepLabv3+ was able to perform better than Spectral Angle Mapping on this task. A mean Intersection over Union of 0.41 was obtained with DeepLabv3+ when trained on a dataset with 1000 buildings versus a mean Intersection over Union of 0.32 for Spectral Angle Mapping. These are very promising results!
Combining deep learning with geospatial data is a powerful and ingenious way to obtain new insights and to serve as a basis for new policies. This makes it possible to classify objects like buildings, segment ground use (like crops) or detect features like pedestrian crossings. We are excited to explore the further possibilities and eager to support interested parties in utilizing this crossover of technologies to solve their (research) questions. Contact us at Geodienst@rug.nl to discuss the possibilities for your case.
The full thesis can be found on http://fse.studenttheses.ub.rug.nl/19092/1/AI_BA_2019_NickUbels.pdf.