Predictive modelling and alert system for rainfall induced landslides

Abstract

Natural disasters are among the leading causes of death worldwide. Statistics indicate that many people lose their lives in different disaster incidents. Damages to properties and infrastructure from this type of hazard are worth millions of dollars, and much more money is spent on disaster recovery. In many countries, including the East African region, landslides and floods are the most common natural disasters causing fatalities. In the least developed or developing countries, there are no efficient mechanisms for prediction or early warning of rainfall-induced disasters for rescuing people before the occurrence. In Rwanda, the common strategies used to reduce the risk of landslides are to move people from high-risk zones to low-risk areas and cover the land with vegetation (forestation, grass). The first method takes a long time as it has a financial implication and citizens have to be responsible since the support from the government is limited. Today, technological solutions are available, such as wireless sensor networks. The last plays a key role in solving many problems by monitoring environmental parameters and providing alerts to the public. With the help of machine learning techniques, the prediction of disaster occurrence can be done by using historical rainfall data and landslide incidence records in the past. This research aimed to reduce the risks of landslides by identifying and analysing the internal and external landslide-causing factors, the correlation between disasters’ occurrences and the causing factors, and then designing and developing an early warning system for predicting rainfall-induced landslides. The system uses wireless sensors to collect hydrological data that is used to predict landslide incidence and alert the public before the occurrence of the incident. Rainfall, topographical, and geological data were collected, and machine learning techniques have been used to predict landslide occurrences. The wireless sensor network was designed and developed to collect real-time data, send it to the cloud where it is processed, and predict landslide incidence. This research was conducted in 3 systematic phases: The prediction of landslide incidence using machine learning models; experimental study to determine thresholds to be used in the prototype implementation; design, development, and testing of the IoT prototype. In the first phase, Random Forest (RF) and Logistic Regression (LR) are the two machine learning models that were used for the prediction of landslides using historical data. The models’ performance was evaluated using false negative rate (FNR) and the receiver operating characteristics, area under the curve (ROC-AUC). The prediction results revealed that the antecedent rainfall has a significant impact on the occurrence of landslides. The AUC for RF was 0.995 and 0.997 for LR, whereas FNR was 4.80% and 3.84% for RF and LR, respectively. The comparative analysis showed that LR performed better than RF. The correlation between rainfall, soil moisture, and landslide xiv incidence was identified in the second phase of this research. The results from this phase revealed the amount of rainfall and soil moisture content inducing landslides. The results of the experiments showed that for a particular site, the minimum time required to cause slope failure was 8h41, with an intensity of rainfall of 8 mm/hour and soil moisture levels exceeding 90% for the sensors placed more than 100 cm deep in the ground. Those thresholds were used for the early warning system prototype, and the delivery of the warning message is based on threshold values. The system prototype was successfully tested at the selected sample sites at a rate of 71.4%. The study area of this research is the Ngororero and Gakenke districts in Rwanda.