Semi-automated classification of very high resolution imagery using convolutional neural networks

Abstract

New challenges in remote sensing require the design of a pixel classification method that once trained can generalise to other areas of the Earth (Maggiori etal. 2017a). Much of the extraction of features from satellite imagery is performed by human analysts and with that comes time delays, increased cost and increased errors (Mnih 2013).

Convolutional Neural Networks are a new and emerging technology for image classification and come as a result of significant advances in the field of Computer Vision and Deep Learning. The convolutional layers within the network ”learn” features which may be based on colour as with traditional pixel-based classification, but also create edge detectors and various other feature extractors that could exist in a region of pixels that would not be recognised by traditional pixel by pixel based classifications.The aim of this project is to produce a semi-automatic, pixel-based, image classification process that reduces the human element thus reducing costs, errors and time.

The model was trained using a labelled building dataset with a pixel size corresponding to 0.5m on the ground. The building dataset was created for Chicago, Vienna, Austin, Kitsap and Tyrol from cadastral datasets and the RGB images were captured in different illuminations and aspects. Overall the trained model performed best in Chicago and Vienna with an overall accuracy of 22% and 27% respectively. Accuracy this low means the model can not be used to classify building, non building for another area of interest. However, the framework is now in place to build upon this project and continue training to fine tune the model to produce more reliable results