Urban change detection from VHR images via deep-features exploitation

Land cover change detection from remote sensing data is a crucial step in the periodic environmental monitoring, but also in the management of emergencies. In particular, the availability of Very High Resolution (VHR) images enables a detailed monitoring on urban, regional or larger scale. Together with data, new methodologies able to extract useful information from them are needed. In the present work, a transfer learning technique is presented to produce change detection maps from VHR images. It is based on the exploitation of opportune deep-features computed by using some pre-trained convolutional layers of AlexNet. The proposed methodology has been tested on a data set composed of two VHR images, acquired on the same urban area in July 2015 and July 2017, respectively. The experimental results show that it is able to efficiently detect changes due to the construction of new buildings, to variation in roof materials or to vegetation cut that has made visible the underlying non-vegetated areas. Moreover, it is robust with respect to false positive because changes due to different occupation of parking areas or due to building shadows are not detected.