Road Extraction From Satellite Images Using Attention-Assisted UNet

These days, extracting information from remote sensing data has a great impact on various aspects of our lives, such as infrastructure and urban planning, transportation and traffic management, forecasting and tracking natural disasters, searching for mineral resources, monitoring environmental changes, and numerous other fields. One crucial application is extracting accurate road information from aerial images, which has many practical applications ranging from our daily lives to long-term planning for transportation systems to autonomous vehicles. Deep learning models have shown great promise in image-processing tasks, specifically in accurately detecting and extracting roads from aerial images. In this study, various techniques were employed to achieve the desired performance. The model is a UNet assisted with attention blocks in the decoder part and trained with a patched, rotated, and augmented dataset that has been extracted from the DeepGlobe dataset. The preprocessing of the dataset included image and mask patching, rotation, exclusion of background-only images, and excluding images with very little road surface. Both patching and background exclusion in preprocessing as hard attention and attention blocks in the model as soft attention were deployed in order to tackle the inherently biased nature of the dataset. This combination of different techniques empowers the proposed model for superior remote sensing image segmentation performance with an accuracy level of 98.33%. In addition to achieve better performance by the model, another objective is to find the issues that cause the model's performance degradation on some image samples. Therefore, a comprehensive analysis of metrics, with a focus on precision and recall as proper metrics for biased dataset analysis, was conducted to identify potential shortcomings in the model or the dataset, and based on the result, several proposals for future work and further investigations were formulated.