I would recommend working with the large-scale labeled datasets such as DOTA ( https://captain-whu.github.io/DOTA/dataset.html ), VisDrone, UAVDet, UAV123 which have been made available for object detection and visual tracking. Recently, DOTA dataset has been extended to provide instance level segmentation as well for the objects in a new dataset name iSAID ( https://captain-whu.github.io/iSAID/).
You could use the mask-RCNN code to improve the results in aerial scene in this new dataset.
Semantic segmentation in UAV images also has widespread applications in agriculture, industries and commercial market space. ( https://arxiv.org/pdf/1910.10026.pdf , https://arxiv.org/pdf/1810.10438.pdf)
Also, you could collect your own UAV data to solve specific problems such as leak detection in pipelines, fault detection in electric grids, etc....