Characterizing Land Subsidence Using Random Forest Algorithm in Taiwan

In light of the increasing frequency of severe droughts linked to climate change in central and southern Taiwan, the Choshui delta in central Taiwan has witnessed a notable surge in land subsidence. Consequently, the pace of land subsidence has peaked at 7.8 cm per year as of 2021 in the Choshui delta, Taiwan. The process of soil compaction in the Choshui delta primarily unfolds as a time-dependent geological phenomenon. A comprehensive understanding and characterization of this process are essential for the formulation of effective mitigation and adaptation strategies. This study presents a pioneering approach to characterize land subsidence in the Choshui delta, utilizing a random forest algorithm (RFA).

The random forest is an ensemble machine learning algorithm known for its versatility and effectiveness in various predictive modeling tasks. It is widely used due to its ability to handle complex relationships for tasks such as classification, regression, and feature selection in land subsidence. By leveraging this advanced modeling technique, we aim to identify and analyze the underlying causes of land subsidence, providing valuable insights into the dynamic interplay of environmental factors. This research contributes to a more comprehensive understanding of land subsidence patterns by incorporating a multi-factorial perspective in the face of changing climatic conditions. Using the RFA, we may identify and analyze the dominant factors affecting land subsidence. Subsequently, a land subsidence prediction model is established based on the RFA, considering the multi-factorial perspective, include cumulative compaction, groundwater levels, electricity consumption for pumping, and rainfall.

The RFA is utilized to identify the temporal patterns from historical time-series data spanning from 2008 to 2021, which is specifically associated with the land subsidence site in the study area. To validate the proposed model, we compare its predictions to the historical time-series data, utilizing metrics such as root mean square error, correlation coefficient, and coefficient of determination. Results demonstrate that, the optimal RMSE, R, and R² values during the prediction phase. The RFA in the context of land subsidence prediction performs exceptionally well, exhibiting high accuracy in predicting land subsidence patterns in the rapidly subsiding areas of the Choshui delta.

Keywords: land subsidence; climate change; random forest algorithm; groundwater; Choshui delta.