Graph Neural Network based Scene Change Detection Using Scene Graph Embedding with Hybrid Classification Loss

Soyeon Kim, Kyung-No Joo, Chan-Hyun Youn

Abstract: The advent of deep learning technologies gives satellite imagery analysis birth to unprecedented achievements to various tasks. Especially, change detection is one of the attentive fields regarding to remote sensing as a unique task to compare the paired images. While a great amount of works deals with change detection in pixel level to generate change map, its labelling cost to train the model in data driven manner is extremely high in that it should be annotated in pixel level as well and it is sensitive to pixel level distortion. Instead of change maps, scene level change detection only classifies whether the newly coming image has different contexts or not especially when the system has target objects in the scene with comparably low labelling cost and considering overall contexts. However, only few works address scene level change detection and are yet unexplored with multiple target objects. In this end, we propose a two-phase framework to screen out the redundant same images compared to the reference time point image. Instead of using image features or object features only, we utilize scene representation graph and train on our proposed GNN architecture as to compare graphs representing images with multiple objects. Due to lack of perfect matching dataset, we verify our proposed framework on correspondingly matchable datasets and show the performance improvement on scene change type classification by 13% including move cases over the baseline.