Combining active-learning approaches with support vector machines for landslide mapping

Cost-effective spatial landslide models play a critical role in landslide mapping after an event and landslide susceptibility modelling for spatial planning and hazard mitigation. Challenges faced by many researchers in compiling the necessary landslide inventories are the time-consuming instance labelling and imbalanced data when training machine-learning models. Active learning is a practical way of reducing labelling costs by selecting more informative instances for labelling by an expert. Although this method has increasingly been adopted in remote-sensing classification, it is relatively new in the context of landslide mapping. To test the performance and potential benefits of active learning in this context, we combined two common active learning strategies, uncertainty sampling and query by committee with a state-of-the-art machine-learning technique, the support vector machine (SVM). Their utility is illustrated in a case study in the Ecuadorian Andes by comparing their performances to SVMs with simple random sampling of training locations. Based on the mean AUROC (area under the receiver operating characteristic curve) as a performance measure, SVMs with uncertainty sampling tended to perform better than random sampling and query-by-committee strategies. Meanwhile, uncertainty sampling achieved more stable performances according to a lower AUROC standard deviation across repetitions. Taken together, under limited data conditions, active learning with uncertainty sampling is more efficient by selecting more informative instances for SVM training. Therefore, we suggest that this strategy can be incorporated into the workflow of interactive landslide modeling not only in emergency response settings but also to more efficiently generate landslide inventories for event-based landslide susceptibility modeling.