In this paper, a Deep learning architecture for semantic segmentation of aerial images is proposed. One of the problems with convolutional neural network is the loss of spatial data after using convolution layers which can be retrieved using up-sampling layers. The proposed network uses skip connection with residual learning within an encoder-decoder architecture to reduce ambiguities in the up-sampling layers. We designed a FCN that uses Infra-Red-Green (IRRG), normalized Digital Surface Model (nDSM) and Normalized Difference Vegetation Index (NDVI) information from ISPRS Vaihingen dataset 2D semantic labeling and reached more than 87% as our best total accuracy.
Paper
Full text
Semantic Segmentation of Aerial Images using FCN-based Network
Semantic Scholar · Computer Science · 2019
Abstract
In this paper, a Deep learning architecture for semantic segmentation of aerial images is proposed. One of the problems with convolutional neural network is the loss of spatial data after using convolution layers which can be retrieved using up-sampling layers. The proposed network uses skip connection with residual learning within an encoder-decoder architecture to reduce ambiguities in the up-sampling layers. We designed a FCN that uses Infra-Red-Green (IRRG), normalized Digital Surface Model (nDSM) and Normalized Difference Vegetation Index (NDVI) information from ISPRS Vaihingen dataset 2D semantic labeling and reached more than 87% as our best total accuracy.