Search

Landslides triggered by heavy rains are increasing in number and creating severe losses in hilly regions across the world. Rainfall thresholds on regional and local-scales are being used for forecasting such events, for efficient early warning. Empirical and probabilistic approaches for defining rainfall thresholds are traditional tools which are being used as part of the forecasting system for rainfall induced landslides. Such methods are easy-to-use and are based on statistical analyses. They can be derived without looking into the complex hydro-geological processes involved in slope failures, but are often associated with the disadvantage of higher false alarms, limiting their applications in a regional landslide early warning system (LEWS). This study is an attempt to improve the performance of conventional meteorological thresholds by considering the effect of soil moisture, using a probabilistic approach. Idukki district in southern part of India is highly susceptible to landslides and has witnessed major socio-economical setbacks in the recent disasters happened in 2018 and 2019. This tourist hub is now in need of a landslide forecasting system, which can help in landslide risk reduction. This study attempts to understand the effect of averaged soil moisture estimates derived from passive microwave remote sensing data, for improving the performance of conventional empirical and probabilistic thresholds. For defining empirical thresholds, an algorithm-based approach such as Calculation of Thresholds for Rainfall-induced Landslides Tool (CTRL-T) has been used. Probabilistic thresholds were defined using a Bayesian approach, finding the posterior probability of occurrence using the marginal and conditional probabilities of the control parameters along with the prior probability of occurrence of landslide. The derived rainfall thresholds were quantitatively compared with the Bayesian probabilistic threshold derived using rainfall severity and soil wetness using an area under the curve (AUC) based on receiver operating characteristics (ROC) curve method. The results show that when the antecedent moisture content in soil is less, only severe rainfall events can trigger landslides in the study area; while less severe rainfall events can also trigger landslides when the soil is wet. The role of soil wetness in the initiation is used to improve the performance of the conventional methods, and a ROC approach was used for the statistical comparison of different models. Further, the results indicated that the probabilistic threshold using rainfall severity and soil wetness outperformed the conventional approaches with AUC of 0.96, being the most sensitive and specific among the models considered. This result opens new promising perspectives for the development of an operational LEWS in the Idukki district based on a combination of rainfall and soil moisture data. Moreover, this work contributes to strengthen the advancing trend of hydro-meteorological thresholds based on soil moisture, which is gaining a growing attention in landslide studies and that, to date, was lacking evidences in monsoon regions.

The objective of this work is twofold: (i) automatically setting up a landslide inventory using a state-of-the art semantic engine based on data mining on online news and (ii) evaluating if the automatically generated inventory can be used to validate a regional scale landslide warning system based on rainfall-thresholds. The semantic engine scanned internet news in real time in a 50 months test period. At the end of the process, an inventory of approximately 900 landslides was automatically set up for the Tuscany region (23,000 km 2 , Italy). Using a completely automated procedure, the inventory was compared with the outputs of the regional landslide early warning system and a good correspondence was found, e.g. 84% of the events reported in the news is correctly identified by the warning system. On the basis of the obtained results, we conclude that automatic validation of landslide models using geolocalized landslide events feedback is possible. The source of data for validation can be obtained directly from the Internet channel using an appropriate semantic engine dedicated to perform a monitoring of the Google News aggregator. Moreover, validation statistics can be used to evaluate the effectiveness of the predictive model and, if deemed necessary, an update of the rainfall thresholds could be performed to obtain an improvement of the forecasting effectiveness of the warning system. In the near future, the proposed procedure could operate in continuous time and could allow for a periodic update of landslide hazard models and landslide early warning systems with minimum or none human intervention.