Enhancing infrastructure detection using drone imagery: a comparative analysis using machine learning models with a custom Bangladesh-based dataset for improved urban planning

Abstract

The growth of cities in different developing nations including Bangladesh is posing several challenges in the making of sustainable cities. Growth of infrastructure whilst maintaining the sustainability of urban sprawl is a major challenge in these areas experiencing rapid development. This research aims to address the issues of achieving the efficient and effective technology for an accurate and timely monitoring and cost-effective urban infrastructure through the utilization of airborne remote sensing technologies. This paper describes and compares several machine learning models, including YOLOv5, YOLOv7, YOLOv8 and a custom model developed by us that enhances the performance of object detection of infrastructure conditions on images captured by drones. Our study is based on a unique data set specifically developed for Bangladesh encompassing multiple aspects of urban features as well as the context of infrastructure expansion across different regions. This dataset is critical during the training and testing stage since the combination of features that works well for our models is tailored for optimal use in urban environments in Bangladesh. In the developed approach, the created machine learning models will be used to identify and recognize the exact building, the roads, and other urban infrastructure features in the images that are used for training. Although our comparison and experimental tests are devoted to the measurement of the detection accuracy and calculation speed of each model, we also dedicate a section for the assessment of detection appropriateness in varied urban scenarios as well as in various illumination conditions. There is a significant improvement in many of the conclusion and result as shown for YOLOv8 and our own custom model in terms of the ability of the model to perform more accurate prediction and the ability to handle the spam case at various conditions of spatial and temporal dimensions of urbanization. The above findings suggest that using drone imagery can indeed be effective in utilizing deep learning tools for urban growth in cities enhancing urban planning. It also demonstrates the possibility of using the approach presented in our work in cities with various populations and levels of development for the most efficient management of these dynamic urban processes. The use of temporal data also have enhanced our knowledge in that it is now possible to monitor temporal changes that occur in infrastructural systems and also in making urban design prediction.The former approach means avoiding the necessary conceptual development for more effective uses of such models in real-world development of cities and communities. In conclusion, this thesis is able to show that hybrid of the latest machine learning models for infrastructure detection shows a great potential for the application to urban areas within Bangladesh and similar developing areas with a high population density in a short time. Through surveys it provides a comparison not only of the strengths and weaknesses of each model, but also the framework for future development of more responsive and sustainable urban planning applications.