Method to map human and infrastructure vulnerability using CNN land cover: Case study of floating tank explosion at petrochemical plants of LaemChabang, Thailand

Abstract

Industrial storage tanks, used to store flammable materials in the petrochemical industry, can induce potential fire and explosion under specific conditions. Therefore, it is necessary to map the population and environment vulnerability, and, to develop procedures for emergency responses in order to reduce potential casualties. In order to achieve this, Convolutional Neural Networks (CNN) are used in this study using 6 classes: floating tank, forest, house, road, wasteland and water. Datasets are built for a total of approximately 1.4 million tiles with a resolution of 0.33m/pixel and their size are optimized in function of the class. The 6 associated CNN models are built and optimized to classify each class. The validation of the models shows that, with the exception of road and wasteland where the precision is only 73% and 89% respectively, the other 4 classes have a value higher than 95%. Post-processing is performed on each prediction before aggregating these results to obtain the land cover. For the floating tank class, a 5 step post-processing is used based on a Density-Based Spatial Clustering of Applications with Noise algorithm (DBCAN) after which blast simulation is applied and effects on people, buildings and trees are obtained through 4 steps. Finally, the petrochemical site of LaemChabang in Thailand is used as study case. Except for the road class that is difficult to detect, land cover is well performed. Human casualties and surface of damaged buildings are finally estimated demonstrating the usefulness of the tool to be used for the emergency planning of industrial disasters.