Accurate Road Detection from Satellite Images Using Modified U-net
Journal
2018 IEEE Asia Pacific Conference on Circuits and Systems, APCCAS 2018
Pages
423-426
Date Issued
2019
Author(s)
Abstract
In this paper, we present an accurate neural network algorithm to detect roads in satellite images. Based on convolutional neural networks, from a 6-channel image, this model is able to transfer the road structure to the output using both the U-net and the atrous convolution architecture. To train this model, we introduce a new combination of existing loss functions including the binary cross-entropy and the Jaccard distance to avoid false positive detection and increase binary classification accuracy. In terms of precision, recall, the F-score and accuracy, experiments carried out using the Massachusetts roads dataset, provide better results than state-of-The-Art road extraction models. © 2018 IEEE.
Subjects
convolutional neural network (CNN); Deep learning; road extraction; satellite images
SDGs
Other Subjects
Convolution; Deep learning; Extraction; Neural networks; Roads and streets; Satellites; Binary classification; Convolutional neural network; False positive detection; Jaccard distance; Neural network algorithm; Road extraction; Satellite images; State of the art; Feature extraction
Type
conference paper