Regional landslide susceptibility mapping of earthflows using deep neural networks: A case study in the southern Apennines, Italy

Mapping earthflow phenomena is of paramount importance for effective environmental and land management policies. Earthflows are landslides characterized by variable extensions and impacts, which can severely affect areas where they originate and develop. The Central Apennines, in southern Italy, exhibit a high number of these phenomena, along with other types of landslides. There, earthflows diffusively impact structures and infrastructures, causing damage and conditioning the economic and social development of these internal and often poor areas of Italy. The spatially diffuse availability of medium-high resolution geological, geomorphological and topographic data is boosting the use of artificial intelligence approaches. These can learn from those data and from the available earthflow maps, to automatically classify those areas where earthflows occur or may potentially happen. Here, Deep-Learners (DL), specifically Deep Neural Networks (DNN), are used to map earthflows in a large area of the southern Apennines, mainly characterized by clay and flysch terrains. The spatial data used are lithology, aspect, curvature, land use and slope, while earthflows are mapped as if they exist or not. Deep Learners will return a spatial map of points, representing the presence or not of a landslide phenomenon. The implemented DL architecture, featuring two hidden layers of 100 and 25 neurons, achieved a quantitative performance on the test set: Accuracy of 0.65, Precision of 0.55, Recall of 0.45, and an F1-Score of 0.5. The results showed a good ability in predicting areas subject to earthflows offering a strategic advantage for risk mitigation in vulnerable regions with limited high-resolution data. • DNN maps earthflow susceptibility using simple geomorphic predictors. • Pixel-based approach shows high spatial coherence in Southern Italy. • Curvature and slope angle are the main drivers for earthflow prediction. • Model detects unmapped landslides, helping update regional inventories.