Search

This study presents an experimentation result to observe the impact of spatial scales and imbalanced data treatment on the landslide susceptibility mapping using machine learning algorithm namely Random Forest (RF). Three districts in the Central Java Province, Indonesia have been used as the case study with regards to the data availability. The presence of data scarcity of landslide occurrence causes data imbalance at various pixel sizes. To treat imbalanced data, Synthetic Minority Oversampling Technique (SMOTE) resampling technique was employed. The RF algorithm was implemented in two modes namely trained using raw imbalanced dataset and trained using SMOTE-balanced dataset on all designated pixel sizes. Five frequently used metrics were chosen to test the model skills. The results show that balancing the data with SMOTE has notably improved the model performances. The performance of models trained using raw data have no trend of improvement with spatial scale density. On the other hand, the model performance trained using SMOTE-balanced data reveal an enhancement of mapping accuracy along with the increase of spatial resolution. Hence, it is prudent to treat imbalanced data using a suitable method like SMOTE to produce a reliable machine learning-based landslide susceptibility mapping that is valuable for predictive disaster mitigation planning and management. [ABSTRACT FROM AUTHOR]