IEEE Geoscience and Remote Sensing Magazine - December 2017 - 27

recent years, several data sets have been made public that
can be used to train deep NNs. The following is a nonexhaustive list.
SCENE CLASSIFICATION (ONE IMAGE IS
CLASSIFIED INTO A SINGLE LABEL)
◗◗ UC Merced data set [177]: This data set is a collection of
aerial images (256 × 256 pixels in RGB space) depicting
21 land use classes. Each class is made up of 100 images.
Because every image comes with a single label, the data
set can be used only for image-classification purposes,
i.e., to classify the whole image into a single land use
class. The data set can be downloaded from [191].
◗◗ Aerial Image data set (AID) [74]: This data set is a collection of 10,000 annotated aerial images distributed in 30
land use scene classes and can be used for image-classification purposes. In comparison with the UC Merced
data set, the AID contains many more images and covers
a wider range of scene categories. Thus, it is in line with
the data requirements of modern deep learning. The
data set can be downloaded from [192].
◗◗ Northwestern Polytechnical University-Remote Sensing Image Scene Classification 45 data set [178]: This data set
contains 31,500 aerial images spread over 45 scene
classes. So far, it is the largest data set for land use scene
classification in terms of both total number of images
and number of scene classes. The data set can be obtained from [193].
IMAGE CLASSIFICATION (EACH PIXEL
OF AN IMAGE IS CLASSIFIED INTO A LABEL)
◗◗ Zurich Summer data set [179]: This data set is a collection
of 20 image chips from a single large QuickBird image
acquired over Zurich, Switzerland, in 2002. Each image chip is pansharpened to 0.6-m resolution, and eight
land use classes are presented. All images are released,
along with their ground truths. The data set can be obtained from [194].
◗◗ Zeebruges, or the Data Fusion Contest 2015 data set [127]:
In 2015, the Image Analysis and Data Fusion Technical
Committee of the IEEE GRSS organized a data processing competition aimed at 5-cm resolution land mapping.
To do so, the organizers provided both an RGB aerial image and a dense (65-points/m2) lidar point cloud over the
harbor of Zeebruges (Belgium). The data are organized
on seven 10,000 × 10,000 pixels tiles. All the tiles have
been labeled densely in eight land classes, including
land use (building, roads) and objects (vehicles, boats)
[126]. The data can be obtained from the Data and Algorithm Standard Evaluation (DASE) website, http://dase
.ticinumaerospace.com/. On DASE, users can download
the seven tiles and labels for five tiles. To assess models
on the two remaining tiles, users can upload the classified
maps on the DASE server.
◗◗ ISPRS 2-D semantic labeling challenge: The working group
II/4 of the ISPRS 3-D Scene Reconstruction and Analysis
DECEMBER 2017

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

provided a subdecimeter resolution data set over the two
cities of Vaihingen and Potsdam (Germany). The data are
similar to those of the Zeebruges data, with the difference
that the height information is provided as a DSM at the
same resolution of the image data. Moreover, images are
provided with an infrared channel. The data set is also
fully labeled into six classes, including land classes
(roads, meadows) and obWITH THE GROWING
jects (cars). It also comes
ATTENTION ON VHR SAR
with a clutter class gatherDATA, THE FUSION OF
ing all unknown objects.
OPTICAL AND SAR IMAGES
The Vaihingen dat a set
comes with 33 tiles having
IN DENSE URBAN AREAS
an average size of 2,000 ×
HAS BECOME AN EMERGING
3,000 pixels. Half the tiles
AND TIMELY TOPIC.
come with labels. The other 17 tiles come with no labels, and participants must
upload classification maps for evaluation. The Potsdam
data set comes with 24 labeled tiles (6,000 × 6,000 pixels)
and 14 unlabeled ones. Both data sets can be obtained
from [189].
REGISTRATION/MATCHING
◗◗ SARptical data set [180]. With the growing attention on
VHR SAR data, the fusion of optical and SAR images in
dense urban areas has become an emerging and timely
topic. At the core of such a fusion topic is the challenging task of coregistering SAR and optical images. Two
such images are acquired with intrinsically different
imaging geometries and thus are nearly impossible to
be coregistered without a precise 3-D model of the imaged scene. SARptical is a unique data set for SAR and
optical image matching in dense urban areas. It consists
of 10,000 pairs of corresponding SAR and optical image patches in central Berlin, with the center pixels of
each patch pair precisely coregistered. They are generated based on coregistered 3-D interferometric SAR point
clouds (which are reconstructed by SAR tomography using tens of TerraSAR-X high-resolution spotlight images)
and 3-D optical point clouds (which are reconstructed
by structure from motion, followed by dense stereo
matching using several UltraCam images with a ground
spacing of 20 cm). This data set can be downloaded from
https://www.sipeo.bgu.tum.de/downloads.
SHOWCASING
Starting to work with CNNs from scratch might seem a titanic task. The number of models available is large, and setting up an architecture from zero is challenging. In this section, we point to three showcasing example that have been
recently provided by remote-sensing researchers. All these
examples are offered with open licenses, and the corresponding papers must be acknowledged when using those
codes. The rules on the respective websites apply. Please
27

https://www.sipeo.bgu.tum.de/downloads http://dase http://www.ticinumaerospace.com/

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2017