Abstract. Change detection applications from satellite imagery can be a very useful tool in monitoring human activities and understanding their interaction with the physical environment. In the past few years most of the recent research approaches to automatic change detection have been based on the application of Deep Learning techniques and especially on variations of Convolutional Neural Network architectures due to their great representational capacity and their state-of-the-art performance in visual tasks such as image classification and semantic segmentation. In this work we train and evaluate two CNN architectures, UNet and UNet++, on a change detection task using Very High-Resolution satellite images collected at two different time epochs. We also examine and analyse the effect of two different loss functions, a combination of the Binary Cross Entropy Loss with the Dice Loss, and the Lovász Hinge loss, both of which were specifically designed for semantic segmentation applications. Finally, we experiment with the use of data augmentation as well as deep supervision techniques to evaluate and quantify their contribution in the final classification performance of the different network architectures.

No abstract

Abstract. Semantic segmentation is an active area of research with a wide range of applications including autonomous driving, digital mapping, urban monitoring, land use analysis and disaster management. For the past few years approaches based on Convolutional Neural Networks, especially end-to-end approaches based on architectures like the Fully Convolutional Networks (FCN) and UNet, have made great progress and are considered the current state-of-the-art. Nevertheless, there is still room for improvement as CNN-based supervised-learning models require a very large amount of labelled data in order to generalize effectively to new data and the segmentation results often lack detail, mostly in areas near the boundaries between objects. In this work we leverage the semantic information provided by the objects’ boundaries to improve the quality and detail of an encoder-decoder model’s semantic segmentation output. We use a UNet-based model with ResNet as an encoder for our backbone architecture in which we incorporate a decoupling module that separates the boundaries from the main body of the objects and thus learns explicit representations for both body and edges of each object. We evaluate our proposed approach on the Inria Aerial Image Labelling dataset and compare the results to a more traditional Unet-based architecture. We show that the proposed approach marginally outperforms the baseline on the mean precision, F1-score and IoU metrics by 1.1 to 1.6%. Finally, we examine certain cases of misclassification in the ground truth data and discuss how the trained models perform in such cases.

Abstract. Over the past few years, many research works have utilized Convolutional Neural Networks (CNN) in the development of fully automated change detection pipelines from high resolution satellite imagery. Even though CNN architectures can achieve state-of-the-art results in a wide variety of vision tasks, including change detection applications, they require extensive amounts of labelled training examples in order to be able to generalize to new data through supervised learning. In this work we experiment with the implementation of a semi-supervised training approach in an attempt to improve the image semantic segmentation performance of models trained using a small number of labelled image pairs by leveraging information from additional unlabelled image samples. The approach is based on the Mean Teacher method, a semi-supervised approach, successfully applied for image classification and for sematic segmentation of medical images. Mean Teacher uses an exponential moving average of the model weights from previous epochs to check the consistency of the model’s predictions under various perturbations. Our goal is to examine whether its application in a change detection setting can result in analogous performance improvements. The preliminary results of the proposed method appear to be compatible to the results of the traditional fully supervised training. Research is continuing towards fine-tuning of the method and reaching solid conclusions with respect to the potential benefits of the semi-supervised learning approaches in image change detection applications.