A multi-task deep learning approach for landslide displacement prediction with applications in early warning systems

Accurate landslide displacement prediction is important for the construction of reliable landslide early warning systems (LEWS). Recently, deep neural networks have become the dominant approach for landslide displacement modeling. However, we show that focusing solely on low prediction residuals is not perfectly aligned with the goals of LEWS, where the emphasis is on precise forecasts near the warning threshold. This can result in poor efficiency of threshold-based warning prediction. We propose a multi-task approach to model training, where auxiliary targets are used to optimize the model towards the performance relevant for LEWS. The methodology is validated using the data from the deep-seated Urbas landslide in north-western Slovenia, which has been monitored by GNSS since 2019. Developing a displacement prediction model for Urbas is a step towards extending the existing wire-based mechanical alarm system. We employ a convolutional neural network for day-ahead displacement prediction using recent landslide activity, hydrometeorological measurements and seismological data. The proposed multi-task model retains a competitive [Formula: see text] score for warning prediction while achieving a significantly lower mean absolute error compared to the reference models. The proposed methodology is generally applicable and has the potential to improve the efficiency of landslide modeling in the context of LEWS.