Enhancing the performance of data-driven models through pre-processing and feature engineering for the forecasting of landslide displacement

Forecasting the evolution of landslide displacements over time in order to have a plan for early warning and risk management is meant to be searched by a comprehensive look at the past and present. In recent years, machine learning techniques have made remarkable advancements in the investigation of natural hazards, specifically by harnessing data patterns and historical information to enhance prediction accuracy. This innovative approach not only improves the understanding of the natural phenomena but also empowers the efforts to reach informed decisions based on reliable forecasts. In particular, machine learning models have the power to describe sophisticated and nonlinear relationships concerning the complex evolution of phenomena. This study highlights the effectiveness of preprocessing and feature engineering techniques, such as transformations, Fourier series, and temporal lags, when applied to the analysis of the evolution of the displacement patterns of slow landslides with time. It emphasizes that, in some cases, even with straightforward methods, like linear regression and Prophet, reliable results can be achieved. A workflow for modeling time series forecasting has been specifically developed, with the aim of processing large volumes of data, as well as incorporating selected features derived from time indices and external inputs. The results from both models, optimized through careful feature engineering, showed high reliability and performance, especially when bolstered by well-designed regressors, lag structures, and seasonal markers. In terms of accuracy, the Prophet model exhibits higher performance. The study is deemed to show that engineered features significantly decrease prediction errors, and the key takeaway highlights the importance of feature richness over model complexity.

Keyword: Machine learning, Feature engineering, Pre-processing, Landslide, displacement, Time series.