The advent of deep learning technologies givessatellite imagery analysis birth to unprecedented achievements tovarious tasks. Especially, change detection is one of the attentivefields regarding to remote sensing as a unique task to compare thepaired images. While a great amount of works deals with changedetection in pixel level to generate change map, its labelling costto train the model in data driven manner is extremely high in thatit should be annotated in pixel level as well and it is sensitive topixel level distortion. Instead of change maps, scene level changedetection only classifies whether the newly coming image hasdifferent contexts or not especially when the system has targetobjects in the scene with comparably low labelling cost and con-sidering overall contexts. However, only few works address scenelevel change detection and are yet unexplored with multiple targetobjects. In this end, we propose a two-phase framework to screenout the redundant same images compared to the reference timepoint image. Instead of using image features or object featuresonly, we utilize scene representation graph and train on ourproposed GNN architecture as to compare graphs representingimages with multiple objects. Due to lack of perfect matchingdataset, we verify our proposed framework on correspondinglymatchable datasets and show the performance improvement onscene change type classification by 13% includingmovecasesover the baseline.