Lightweight Multiscale Attention-Aware Method for Semantic Segmentation of Urban Structural Buildings in Drone Aerial Imagery

Abstract

Every year, the lives of millions of people all over the world are affected by natural disasters such as floods, earthquakes and hurricanes. Urban and under-development communities are often hit the most during disasters. Disaster resilience requires proper planning and timely response. Recently, there has been growing interest in the potentials of computer vision and artificial intelligence to tackle challenges related to disaster resilience and emergency preparedness. Compared to cities, detecting structural buildings in urban environments is a more challenging task. Within this context, this paper presents a novel lightweight multiscale attention-aware method for semantic segmentation of urban structural buildings in drone aerial imagery. Detailed experiments were conducted to evaluate our proposed method using drone aerial images from Zanzibar city, Tanzania. Despite being lightweight, significant performance gains were achieved compared to relevant state-of-the-art methods in the literature.