Beyond Hand-Crafted Features in Remote Sensing

A basic problem of image classification in remote sensing is to select suitable image features. However, modern classifiers such as AdaBoost allow for feature selection driven by the training data. This capability brings up the question whether hand-crafted features are required or whether it would not be enough to extract the same quasi-exhaustive feature set for different classification problems and let the classifier choose a suitable subset for the specific image statistics of the given problem. To be able to efficiently extract a large quasi-exhaustive set of multi-scale texture and intensity features we suggest to approximate standard derivative filters via integral images. We compare our quasi-exhaustive features to several standard feature sets on four very high-resolution (VHR) aerial and satellite datasets of urban areas. We show that in combination with a boosting classifier the proposed quasi-exhaustive features outperform standard baselines.

Representative Publications:

Features, color spaces, and boosting: New insights on semantic classification of remote sensing images
Tokarczyk, P., Wegner, J.D., Walk, S., Schindler, K.
IEEE Transactions on Geoscience and Remote Sensing, vol. 53(1), 2015, pp. 280-295

Beyond Hand-Crafted Features in Remote Sensing
P. Tokarczyk, J. D. Wegner, S. Walk, K. Schindler
ISPRS Workshop on 3D Virtual City Modeling (VCM 2013), Regina, Canada, 2013

An evaluation of feature learning methods for high resolution image classification
P. Tokarczyk, J. Montoya, and K. Schindler
22nd ISPRS Congress, Melbourne, Australia, 2012

Contact person: Piotr Tokarczyk and Jan Dirk Wegner