The increased occurrence of catastrophic events caused by climate change-induced mass movements affects human welfare and results in huge economic losses, making the development of early warning systems (EWSs) crucial to everyday decisions of local governments and communities. The state-of-the-art (SOTA) EWSs can only provide information on the level of direct danger of mass movement that each village is exposed to. This is the result of highly sensitive and coarse classifications and aggregated predictions at regional level, potentially leading to a poor perception of risk and inadequate local management plans. Also, this does not account for the indirect effects of mass movements on the local communities at social and economic levels, e.g., a town that can be cut out of the transportation and communication systems because of landslides blocking the roads around it is currently not considered by SOTA EWSs.

To overcome this issue, we developed a novel machine learning scheme that investigates the environmental (hydrological and geological) characteristics of mass movements and the connectivity information of formal settlements and road network data in terms of graph structures. In particular, we study the interaction of the probabilistic mass movement susceptibility (derived from the environmental properties by means of a supervised ensemble graph neural network) on the graph representing the road network connecting the formal settlements. As a result, we derive for each formal settlement a probability of being indirectly affected by mass movements (e.g., the probability to be isolated as a result of mass movements affecting their surroundings) by graph spectral clustering.

We tested this architecture (named Intergraph) on the Norwegian territory, taking advantage of over 68,000 incidents of reported mass movements since 1957. Our approach achieved an overall performance of 86.25% with the 2020 Gjerdrum quick clay incident as a demonstrated case study. With the intensifying effects of climate change, our study has opened an opportunity to develop solutions for adaptation and mitigation through a new holistic graphical perspective to assess various large-scale geospatial datasets of risk elements such as exposure, vulnerability, and hazard.