Optimizing the Decomposition of Time Series using Evolutionary Algorithms: Soil Moisture Analytics

Soil moisture plays a crucial part in earth science, with impact on agriculture, ecology, hydrology, landslides, and water resources. Extremes in soil moisture, which we denote as peaks and valleys, caused by heavy rainfalls and subsequent dry weather, are very important when predicting future soil moisture or even landslides. Existing methods, like moving averages, have limitations when it comes to smoothing time series data while preserving peaks and valleys. In this work, we propose a novel method, HyperSTL, for extrema-preserving smoothing of soil moisture time series. The method optimizes an existing time series decomposition technique, Seasonal Decomposition of Time Series by Loess (STL). HyperSTL optimizes STL's control parameters, which we call hyperparameters, using an objective function over the decomposed components. We demonstrate in experiments with nine soil moisture datasets that using HyperSTL generally results in improved predictions compared to using other smoothing methods.