Learning to extract buildings from ultra-high-resolution drone images and noisy labels
- 20 May 2020
- journal article
- research article
- Published by Taylor & Francis in International Journal of Remote Sensing
- Vol. 41 (21) , 8216-8237
Abstract
Building maps have a plethora of applications in government, industry and academia. In most cases, large scale maps can be retrieved from OpenStreetMap vector data. However, for certain rapidly changing built and semi-built environments, corresponding maps are not as accurate and contain label noise such as missing, incorrectly present, shifted labels, etc.; mainly because buildings in those regions are constantly being constructed, deconstructed, replaced and altered. One such case is extant in the Rohingya camps of southeastern border region of Bangladesh. Mass refugee influx in late 2017 and following population growth has necessitated the construction of buildings and expansion of camps. Consequently, reliable methods are necessary for detecting and documenting camp buildings. Ultra-high-resolution drone images of Rohingya camps are semantically segmented through fully convolutional U-Net deep learning systems for generating accurate building maps from noisy labels. A wide variety of noises are prevalent in the labels. Deep learning systems provide less noisy predictions compared to the classification tool in the most widely used Geographic Information System (GIS) software, ArcGIS. Data augmentation and regularization allows reliable learning, even in the presence of label noise. During testing, calculation of numeric performance metrics against noisy labels can grossly underestimate true skill and performance of the model. A subset of 22 million pixels of the testing data is relabelled by hand to obtain noise-free labels. Testing our generated maps against noisy and noise-free labels confirms that true performance is higher than otherwise indicated by freely available building maps. Empirical results reveal that utilized pipeline is able to learn from noisy data and produce labels which are more accurate and less noisy. Labels generated by our best performing system provide Intersection-over-Union (IoU) gain of 17.6% and Dice score gain of 13.6% over freely available labels from OpenStreetMap. Finally, spatio-temporal building maps are generated to portray the applicability of this research.Keywords
This publication has 31 references indexed in Scilit:
- Efficient Label Collection for Image Datasets via Hierarchical ClusteringInternational Journal of Computer Vision, 2017
- Combining satellite imagery and machine learning to predict povertyScience, 2016
- U-Net: Convolutional Networks for Biomedical Image Segmentation2015
- Fully convolutional networks for semantic segmentation2015
- Building and road detection from large aerial imagery2015
- Comparison between artificial neural networks and maximum likelihood classification in digital soil mappingRevista Brasileira de Ciência do Solo, 2013
- Comparison of standard maximum likelihood classification and polytomous logistic regression used in remote sensingEuropean Journal of Remote Sensing, 2013
- Learning to Detect Roads in High-Resolution Aerial Images2010
- Land use and land cover change in Greater Dhaka, Bangladesh: Using remote sensing to promote sustainable urbanizationApplied Geography, 2009
- The use of prior probabilities in maximum likelihood classification of remotely sensed dataRemote Sensing of Environment, 1980