Slope Inspection System: Using image processed by machine learning algorithm to determine risk of slope failure

Abstract

Malaysia is a tropical country that experiences rainy and hot weather throughout the year. The higher rainfall intensity leads to higher landslide occurrences in Malaysia. Landslides that occur nearby human settlements increase the risk and hazard to the public and properties that lead to significant economic losses. There are various methods in surveying the risk and hazard of landslide areas such as terrestrial laser scanning (TLS) and global positioning system (GPS). Most of the past research uses the conventional method which requires an in-situ field survey, lab analysis, and an additional software package to determine the hazard level for a slope. The conventional method is inefficient and time-consuming. In this paper, the potential of a machine-learning algorithm to improve the conventional approach in detecting the hazard in a landslide is discussed. The algorithm assesses the level of risk based on trained supervised images identified by experts in the field. Using the trained model, it was found that Convolution Neural Network (CNN) is able to perform better than Fully Connected Layer within the reduced processing time, with increased accuracy of 22%. However, the accuracy of the CNN on test data in IIUM could still be improved further. Currently, there is an actual prediction accuracy is at 50% for the test data in IIUM, a 17% error compared to CNN prediction. More training data could be added to the CNN to improve the current accuracy.