High resolution landslide susceptibility mapping using ensemble machine learning and geospatial big data

43 Landslide susceptibility represents the potential of slope failure for given geo-environmental 44 conditions. The existing landslide susceptibility maps suffer from several limitations, such as 45 being based on limited data, heuristic methodologies, low spatial resolution, and small areas of 46 interest. In this study, we overcome all these limitations by developing a probabilistic 47 framework that combines imbalance handling and ensemble machine learning for landslide 48 susceptibility mapping. We employ a combination of One -Sided Selection and Support Vector 49 Machine Synthetic Minority Oversampling Technique (SVMSMOTE) to eliminate class 50 imbalance and develop smaller representative data from big data for model training. A blending 51 ensemble approach using hyperparameter tuned Artificial Neural Networks, Random Forests, 52 and Support Vector Machine, is employed to reduce the uncertainty associated with a single 53 model. The methodology provides the landslide susceptibility probability and a landslide 54 susceptibility class. A thorough evaluation of the framework is performed using receiver 55 operating characteristic curves, confusion matrices, and the derivatives of confusion matrices. 56 This framework is used to develop India's first national-scale machine learning based landslide 57 susceptibility map. The landslide database is carefully curated from global and local 58 inventories, and the landslide conditioning factors are selected from a multitude of geophysical 59 and climatological variables. The Indian Landslide Susceptibility Map (ILSM) is developed at 60 a resolution of 0.001 o (~100 m) and is classified into five classes: very low, low, medium, high, 61 and very high. We report an accuracy of 95.73%, sensitivity of 97.08%, and matthews 62 correlation coefficient (MCC) of 0.915 on test data, demonstrating the accuracy, robustness, 63 and generalizability of the framework for landslide identification. The model classified 4.75% 64 area in India as very highly susceptible to landslides and detected new landslide susceptible 65 zones in the Eastern Ghats, hitherto unreported in the government landslide records. The ILSM