engleski

Change Detection Based on a Convolutional Neural Network for Remote Sensing Imagery

Change detection (e.g., detection of illegal buildings, water area supervision, natural disaster assessment, urban planning expansion research, etc) in remote sensing imagery is a critical and challenging task. It consists of comparing a registered pair of images of the same region and identifying the parts where a change has occurred. In recent years, image processing methods based on Convolutional Neural Networks (CNNs) have achieved very good results. In the field of semantic segmentation, CNNs are widely used due to its high efficiency and accuracy. Therefore, in this paper, UNet-based architecture is proposed for change detection (CD) between pairs of satellite images. The PlanetScope satellite imagery including four spectral bands (i.e., red, green, blue, and near infra-red) and a spatial resolution of 3.7m was used for CD. After manually labelling of training samples has been done, both sets of images have been divided into 512x512 pixel raster tiles and they have been concatenated to create an image with 8 bands which will be used as an input to UNet network. Gathered dataset has also been augmented by combining mirroring and rotating each raster input as well as its respective label. UNet network has been trained on both augmented and non-augmented datasets. The main contributions of this paper are twofold. The first contribution which is presented in this paper is that UNet CNN can detect changes between co-registered image pairs which were acquired with a same sensor at a different dates. The second contribution presented in this paper is that through data augmentation it is possible to improve the F1- score for a small dataset.

change detection, PlanetScope, convolutional neural networks, UNet, data augmentation

nije evidentirano