Landslides pose a significant threat to society and infrastructure and their occurrence is projected to increase in many regions under the effect of climate change. There is an urgent demand for reliable monitoring of this natural hazard. The combination of spaceborne remote sensing data with state-of-the-art machine learning algorithms offers valuable tools for landslide detection in remote areas. However, a key limitation for the detection lies in the scale factor, especially for methods relying on pixel neighbourhoods. This study presents an innovative methodology which combines a spatial graph with SAR and multi-spectral change products. The graph integrates the flow direction based on the topography into the neighbourhood determination. This unique neighbourhood allows for the preservation of the unique shape and signature of an individual landslide. This paper compares the proposed graph neighbourhood to a common square window approach. Therefore, a RFC is trained with neighbourhood statistics from both approaches and applied to landslides of varying extent. A research area in New Zealand’s West Coast region is selected due to the continuous evolution of a single landslide over multiple events. The graph approach shows promising results, particularly for small-scale events which are successfully detected while being missed by the common window approach. Using the graph neighbourhood, we can even detect the smallest visible extent of the landslide at 2-3 pixels (30-45m) width. The main limitation of the proposed approach lies in the quality of the input data. Future work will focus on the improvement of the Sentinel-1 and Sentinel-2 pre-processing.