Importance of multi-parameter approaches in the development of Machine Learning algorithms for landslide displacement forecasting

Landslides are a constant source of danger for people and the built environment. Even with more recent advancements in this field, these natural hazards still represent a serious problem worldwide and the effort to mitigate their impact has driven the development of specific knowledge, technology, and practices. One of the current challenges involves the ability of new sensors and monitoring techniques to provide a considerable amount of information, which should necessarily be addressed with automatic procedures for their elaboration. A possible approach relies on the introduction of algorithms from the field of Machine Learning applied to the interpretation of the landslide behavior. ML algorithms have key advantages in this context: they do not need explicit knowledge or models of the problem, since they learn directly form the examination of data. They scale with data, so that they get more accurate when presented with large amount of information. Moreover, they can adapt to learn from many different sources like numeric measurements, images, text, or sound. The methodology here discussed was designed for landslide monitoring and early warning activities based on hydrogeological information, with the objective to predict the monitored site behavior few days in advance. In particular, input data for this process are displacement measurements, water level, and meteorological conditions. The model relies on an Artificial Neural Network that will read the values of these parameters measured over a predefined number of consecutive days, predicting the displacement of the day following the last observed. While the model is training, the inputs are selected from past data, so that the model’s prediction can be compared with already available measurements. The error in the prediction is used to adjust the algorithm until it starts to make accurate forecasts. Once the model has reached this stage, it can be shown the measurements of the last days, so that in will predict the development of the slope a few days from now. Moreover, the presence of several sensors on the studied site could give the possibility to assess the landslide behavior in sectors where no monitoring tool is present, thanks to spatial interpolation procedures performed on forecasted displacements. For these types of algorithms, the addition of irrelevant information can lead to lower accuracy in the results. Theoretically, with a large enough dataset, the model should become able to distinguish between useful and inconsequential inputs, in practice however it is very hard to have the necessary tools to achieve that. Accounting for that, the training process will be performed on models including different subsets of the available data, in order to identify the informational value of cross-correlations between monitored parameters.