Handling data imbalance in machine learning based landslide susceptibility mapping: a case study of Mandakini River Basin, North-Western Himalayas

Machine learning methods require a vast amount of data to train a model. The data necessary for landslide susceptibility mapping is a collection of landslide causative factors as predictors and landslide inventory as a response variable; however, landslides do not occur everywhere, and the occurrence of landslides is limited in an area. This geophysical phenomenon leads to severely skewed class distribution, wherein the number of landslide samples (minority class) is significantly less than non-landslide locations (majority class). The imbalance in landslide data hampers the predictive ability of learning algorithms, and hence, the final models show poor performance in the class with fewer samples. This work uses two undersampling techniques, namely, EasyEnsemble (EE) and BalanceCascade (BC), for reducing the effect of imbalance in data. The landslides that occurred between 2004 and 2013 are randomly divided into two groups, i.e., 70% of the samples for training and 30% for testing, whereas the landslides that occurred between 2014 and 2017 have been used for validation. The balanced data is used with the support vector machine (SVM) and artificial neural network (ANN), thereby making four new approaches, i.e., EESVM, EEANN, BCSVM, and BCANN, for susceptibility mapping. We used several metrics, such as recall, geometric mean, precision, accuracy, and Heidke skill score, to evaluate the performance of landslide susceptibility maps. The AUC for imbalanced data with SVM and ANN is 0.50, which shows that the model cannot discriminate between landslide and non-landslide locations. This misclassification is due to a small number of landslide samples and serious class biases. The balanced data using EE and BC methods gives promising results and shows significant improvements, wherein the AUC of EESVM, EEANN, BCSVM, and BCANN is 0.869, 0.918, 0.881, and 0.923, respectively. Among all the methods, the recall and G-mean values were highest for EEANN, which represents the best separation performance of EEANN on landslide samples. Furthermore, we have used the standard error (SE) of AUC and 95% confidence interval to test the significance of various combinations of classification and undersampling schemes. The SE is highest for EESVM and BCSVM among all methods. Based on several accuracy metrics, we conclude that EEANN performs better than all the other methods. The BC-based method does not perform well for landslide susceptibility mapping and provides the highest misclassification of landslide samples. The study shows that the susceptibility maps prepared over balanced data using SVM and ANN show remarkable improvements in accuracy over imbalanced data.