AUTHOR=Mohanty Sharada Prasanna , Czakon Jakub , Kaczmarek Kamil A. , Pyskir Andrzej , Tarasiewicz Piotr , Kunwar Saket , Rohrbach Janick , Luo Dave , Prasad Manjunath , Fleer Sascha , Göpfert Jan Philip , Tandon Akshat , Mollard Guillaume , Rayaprolu Nikhil , Salathe Marcel , Schilling Malte 

TITLE=Deep Learning for Understanding Satellite Imagery: An Experimental Survey

JOURNAL=Frontiers in Artificial Intelligence

VOLUME=Volume 3 - 2020

YEAR=2020

URL=https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2020.534696

DOI=10.3389/frai.2020.534696

ISSN=2624-8212

ABSTRACT=Translating satellite imagery into maps requires in-tensive effort and time, especially leading to no accurate maps ofthe affected regions during disaster and conflict. The combinationof   availability   of   recent   datasets   and   advances   in   computervision   made   through   deep   learning   paved   the   way   towardsautomated satellite image translation. To facilitate research in thisdirection,  we  introduce  the  Satellite  Imagery  Competition  usinga  modified  SpaceNet  dataset.  Participants  had  to  come  up  withdifferent segmentation models to detect positions of buildings onsatellite images. In this work, we present 4 approaches based onimprovements  of  U-net  and  MaskRCNN  models,  coupled  withunique  training  tricks  using  boosting  algorithms,  morphologicalfilter, CRFs and custom losses. The results as high as AP=0.937and  AR=0.959  from  these  models  demonstrate  the  feasibility  ofthe  Deep  Learning  in  automated  satellite  image  annotation.