Stadium Detection From Alsat-2 and Google-Earth Multispectral Images using YOLOv5 and Mask R-CNN

Deep Learning (DL) has recently shown promise performance in remote sensing (RS) field. Object detection is one of the hottest research and challenging topic in RS due to the large variant in object distributions, complex object geometry, sun angle, scales, weather conditions, etc. In this paper, stadium detection approach based on YOLOv5 and Mask R-CNN models is proposed and tested on two multispectral datasets; Alsat-2 and Google-Earth imageries in three different scenarios. The proposed framework provides a comparative study of multi-source and single source training, considering the trade-off between the detection accuracy and the generalization capacity where the experimental results show that the average detection accuracy of the proposed technique for the merged training samples is the highest against single training source.