Comparative study of different spatial resolution DEM and different GIS-based statistical models for landslide susceptibility mapping at Ossey watershed area in Bhutan

Abstract

Landslide is one of the most frequent disasters at the Ossey watershed area in Bhutan causing inconvenience to the local people, financial losses, and claiming the lives of the people every year. This study aim to developing Landslide Susceptibility Mapping (LSM) at the Ossey watershed area in Bhutan and find the magnitude of impact of factors on the landslide by the factors. This study compares the accuracy of the different bivariate statistical models and the different spatial resolution on the accuracy of the LSM using various statistical methods.

The landslide inventory was done using the sentinel-2 imagery data, google earth image and field investigation. A total of 164 landslide locations were identified during landslide inventory of which 70% (115 landslide) were used for training datasets and the remaining 30% (49 locations) for the validation dataset. The LSM was developed using the fifteen factors which are derived from DEMs (ALOS PALSAR and SRTM), geological map of Bhutan, sentinel 2 data, digital topographic map of Bhutan, and rainfall data from Bhutan. All the influencing factors were resampled into three spatial resolutions namely to 12.5m, 30m, and 90m.

Three primary models were used to develop LSM which includes 1) Frequency Ratio (FR), 2) Index of Entropy (IOE), and 3) Weight of Evidence (WOE). The primary models were combined to form hybrid models which includes Frequency Ratio-Index of Entropy (FR-IOE), Index of Entropy, and Weight of Evidence (IOE-WOE), and Weight of Evidence and Frequency Ratio (WOE-FR). The LSM was developed using three different spatial resolutions for individual primary and hybrid models. All the LSM developed using various models were classified into five classes using the natural break classification to check area variation in different landslide zone.

The LSMs was validated using sensitivity, specificity, accuracy, Kappa index, Area Under the Curve (AUC), Root Mean Square Error (RMSE). The sensitivity shows degree of correctly classified landslide pixel, specificity shows degree of correctly classified non-landslide pixel, accuracy shows the proportion of correctly classified landslide and non-landslide pixel, kappa index shows the reliability of the models, AUC shows prediction rate and RMSE shows the relative error between the models. The WOE and its hybrid models shows better accuracy in all the validation parameters. The highest sensitivity (0.8095) corresponds to WOE, IOE-WOE and WOE-FR, highest accuracy for WOE-FR(0.7925), highest Kappa index for WOE-FR(0.5850), highest AUC for 0.8817(WOE and IOE-WOE), and the lowest RMSE for WOE(0.3722). This clearly shows that WOE is best model due to its superior accuracy. Moreover, when WOE is combined with other inferior models, it increases the accuracy.

Regarding the deviation of accuracy using different accuracy parameters, it was observed that finer spatial resolution is much better than the coarse spatial resolution with higher sensitivity, specificity, accuracy, kappa index, AUC and lower RMSE.

The results are expected to help researchers to understand how the accuracy deviates with the change in spatial resolution and to choose the best bivariate statistical analysis. The resultant maps are expected to provide a technical guide for the planners, decision-makers, and engineers for future developmental activities at the Ossey watershed area.

-