Structural crack classification and grading after disaster: a supervised learning approach

Abstract

A structure can be cracked by various disastrous events (for example: flood, cyclone, volcanic eruption, earthquake, fire outbreak). When a disastrous event occurs in an area, a lot of structures get damaged within a very short time. To reduce the death toll, it’s essential to analyse and classify the structural damage or cracks quickly after the impact. Especially, in dense cities like Dhaka, where millions of structures have been built without following any organized plan, the damage can be unimaginable. Keeping this in mind, the first step after the event should be locating the severely damaged areas quickly and starting rescue operations by prioritizing the level of damage. Researchers conducted much research to identify the damage scale and classify structural cracks using machine learning algorithms. In most of their research, they considered a visual representation of the structure, structural parameters (for example: age, materials quality, strength, etc.), soil quality, magnitude, and so on. Considering these parameters is very important for accurate and precise prediction. However, collecting these types of data is a lengthy process. For that, their methods fail to provide a quick assessment. Therefore, this paper aims to classify the structural cracks and provide a quick assessment by considering only the visual representation of the damaged structure. Additionally, it implements various machine learning (for example: SVM, Decision Tree, KNN, RF etc) and deep learning algorithms (for example: VGG16, VGG19, ViT, ADA-ViT, D-ViT, etc). It also analyses and compares the performance of those models. Finally, this study proposes an architecture that can bring the highest accuracy (98.1%) among all the models that were implemented. Furthermore, in this architecture we have introduced a new approach which is we have considered both initial and damaged visual representation of a structure while analysing the damage grade. For annotating the dataset, this study follows EMS-98 (European Macro-Seismic Scale -98) standard.