ROAD NETWORK EXTRACTION VIA DEEP LEARNING AND LINE INTEGRAL
CONVOLUTION
Peikang Li
1
, Yu Zang
1
,Cheng Wang
1
, Jonathan Li
1,2
, Ming Cheng
1
, Lun Luo
3
, Yao Yu
1
1Fujian Key Laboratory of Sensing and Computing for Smart City, School of Information Science and E
ngineering, Xiamen University, Xiamen, Fujian 361005, China
2GeoSTARS Lab, Department of Geography and Environmental Management, University of Waterloo,
Waterloo, Ontario N2L 3G1, Canada
3China Transport Telecommunications&Information Center, Beijing 100011, China
ABSTRACT
In this paper, we propose a learning-based road network
extraction scheme from high resolution satellite. First, the
convolutional neural network(CNN), which is able to
capture large context of local structures, are applied to
predict the probability of a pixel belonging to road regions,
and assign labels to each pixel to describe whether it is road.
Then, a line integral convolution based algorithm is
developed to smooth the rough map to connect small gaps.
Finally, by combining with some common image processing
operators, road centerlines are able to be acquired. Attribute
to the learning capacity of CNN, and the line integral
convolution based connection scheme, the proposed road
extraction method is able to provide high quality results
comparing to current state-of-art road extraction methods.
Index Terms—Convolutional neural network, line
integral convolution, road network extraction, satellite
images
1. INTRODUCTION
Road network extraction from satellite images is a crucial
and challenging remote sensing problem. The extraction
system may lead to automated map generation and updating
which is significant for providing many important services.
For example, accurate road maps could route vehicles and
plan path automatically for unmanned aerial vehicles. Also,
the system could provide evaluable prior knowledge for the
detection and recognition of vehicles, buildings or other
objects. Up to now, road maps are mainly constructed and
updated manually. So automated road network extraction
will save great labor and has gained a lot of attention in
remote sensing area.
Much of the approach of road network extraction relies on
spectral or structure features [1-5]. For example, the
appearance of some roads is high-contrast regions with low
curvature and constant width. These traditional method may
include a typical detection strategy involving edge detection,
followed by edge grouping and pruning. But sometimes the
feature based assumptions are limited. For instance, the
spectral behaviors of data from different sensors appear
differently, as shown in Fig. 1 (a) and (b),which are two
remote sensing images from Pleiades-1A and Geoeye
satellite, the spectral intensity of road regions is bright for (a)
and dark for (b). This may lead to potential errors of road
detections on data. The occlusions by trees, buildings, etc.
may also lead to small gaps and crack on the extracted map.
Another kind of extraction method are learning-based.
Learning-based approaches have attempted at predicting
whether a given pixel is road or not given features extracted
from some context around it since last decade [6-10]. But
most of them failed to scale up to large challenging datasets
because their networks are less powerful.
In this paper we propose a deep learning based approach and
a LIC based road connection method to address these
challenges. Learning approaches are particularly well-suited
to the road detection task because it is a rare example of a
problem where expert-labelled data is abundant. It is easy to
obtain hundreds of square kilometers of high-resolution
satellite images and aligned road maps. LIC based method is
applied to enhance the road network structure. Finally,
combined with several image processing operators, a refined
road network is acquired.
2. METHODOLOGY
Firstly, convolutional neural networks have been applied to
train enough training dataset, which contains thousands of
images with the size of 32*32. Then we define a 32*32
sliding window to slide input remote sensing image pixel by
pixel. The trained CNNs model has been applied to predict
if a pixel of our slicing image belongs to the road, and to
form a rough map. So that we can form the rough structure
map. At the last, a LIC based post processing are used to
connect the gaps and cracks and refine the rough road map,
which is caused by occlusions and shadows.
2.1. Problem description
Let be a satellite image and let be the corresponding
road network image. So if the pixel on the image
1599978-1-5090-3332-4/16/$31.00 ©2016 IEEE IGARSS 2016