The landslide occurrences in mountainous regions are mainly triggered by the earthquakes, hydro-climatic conditions. Accurate landslide inventory mapping on a regional scale after the main event is still a challenging task. However, landslide inventory maps considered as the main factor for landslide susceptibility assessment and risk analysis. As most of the landslide susceptibility models apply inventory maps to investigate the importance of each landslide conditioning factors. Consequently, analyzing the factors based on their significance in landslide assurances to finding the more likely areas for future hazards. However, happening a landslide in nature cannot be considered as an important event. The importance of these landslides appears when conducted with human made infrastructures and especially settlement areas. In this respect, accurate detection of both landslide and settlement areas is considered as the main factors for the risk analysis and mapping. This study presents an integrated approach of machine-learning (ML) models and Dempster-Shafer theory (DST), using RapidEye optical images and ALOS digital elevation model (DEM) from a case study area in higher Himalayas. The visible bands of RapidEye optical images and the topographic landslide conditioning factors of slope, aspect and plan curvature which derived from the ALOS DEM were considered as the model's input dataset. Support vector machines (SVM) and random forest (RF) which considered as the common ML models for image classification (Mezaal, Pradhan, and Rizeei 2018), were used for detection of landslide and settlement areas. The models were trained based on the training segment of a landslide inventory map which prepared by an intensive field working and improved by manual extraction from RapidEye optical images. In this study, to improve the accuracy of the resulted probabilistic outputs of the applied models, the final decision was performed using DST. The DST is an extension of the Bayesian probability theory can improve the accuracy of the model's probability outputs based on three functions namely: belief assignments (BBAs), belief, and plausibility (Feizizadeh 2018). Three areas of disbelief, uncertain and belief were illustrated in the DST scheme (See figure 1). Since the probability outputs of models are usually uncertain, the DST is a practical approach for reducing the existing uncertainty among such these outputs. In this study, the confusion/error matrices were derived from a comparison between the corresponding inventory map and the outputs of the SVM and RF models. The Kappa coefficient was obtained based on the confusion matrix with the following equation (1): Kappa coefficient (〖=(θ〗_1- θ_2))⁄((1-θ_2 )) (1) where θ_1 refers to the ratio of correctly detected areas. Whereas, θ_2 denotes the proportion of agreement predicted by chance. The confusion/error matrices were used for the fusion process for increasing the accuracy of the resulting final maps. The whole process was summarized in the flowchart of figure 2. Our study illustrated a fusion of ML models of detecting of landslide and settlement areas using the DST method. The integrated approach fused the outputs of the models to prepare a more accurate detection. The more precise maps were derived based on the integrated approach, which was specified by an improvement in landslide and settlement areas detection over SVM and RF models.
|Publikationsstatus||Veröffentlicht - 12 Mai 2019|
|Veranstaltung||ESA Living Planet Symposium 2019 - Milan, Italien|
Dauer: 13 Mai 2019 → 17 Mai 2019
|Konferenz||ESA Living Planet Symposium 2019|
|Zeitraum||13/05/19 → 17/05/19|
Systematik der Wissenschaftszweige 2012
- 105 Geowissenschaften