Land Cover Classification in Combined Elevation
and Optical Images Supported by OSM Data,
Mixed-level Features, and
Non-local Optimization Algorithms
Dimitri Bulatov, Gisela Häufel, Lukas Lucks, and Melanie Pohl
Abstract
Land cover classification from a
challenging task in Remote Sens
able elevation data, shadows an
of appearances are abundant in urban terrain. In this paper,
we propose an approach for supervised land cover classifica-
tion that has three main contributions. Firstly, for the cumber-
some task of training data sampling we propose an algorithm
which combines the freely available OpenStreetMap data with
the actual sensor data and requires only a minimum of user
interaction. The key idea of this algorithm is to rasterize the
vector data using a fast segmentation result. Secondly, pixel-
wise classification may take long and be quite sensitive to the
resolution and quality of input data. Therefore, superpixel
decomposition of images, supported by a general framework
on operations with superpixels, guarantees fast grouping
of pixel-wise features and their assignment to one of four
important classes (building, tree, grass and road). Particularly
for extraction of street canyons lying in the shadowy regions,
high-level features based on stripes are introduced. Finally,
the output of a probabilistic learning algorithm can be post-
processed by a non-local optimization module operating on
Markov Random Fields, thus allowing to correct noisy results
using a smoothness prior. Extensive tests on three datasets of
quite different nature have been performed with two probabi-
listic learners: The well-known Random Forest and by far less
known Import Vector Machine are explored. Thus, this work
provides insights about promising feature sets for both classi-
fiers. The quantitative results for the
ISPRS
benchmark dataset
Vaihingen are promising, achieving up to 94.5% and 87.1%
accuracy on superpixel and on pixel level, respectively, de-
spite the fact that only around 10% of available labeled data
were used. At the same time, the results for two additional
datasets, validated with the autonomously acquired training
data, yielded a significantly lower number of misclassified
superpixels. This confirms that the proposed algorithm on
training data extraction works quite well in reducing errors of
second kind. However, it tends to extract predominantly huge
and easy-to-classify areas, while in complicated, ambiguous
regions, first type errors often occur. For this and other algo-
rithm shortcomings, directions of future research are outlined.
Introduction
Motivation
Land cover classification, especially in urban and semi-urban
environment, is a key step for creating semantic
3D
models
ata. As mentioned e.g. by Bulatov et
es of semantic models are: Higher
level of compression, as well as better
eroperability on the non-expert users’
part. Buildings are modeled on a desired level of detail, trees
and vehicles are placed at positions they have been detected
and are represented by geo-specific models from a library,
etc. The emphasis of that and comparable studies (Haala,
2005; Lafarge and Mallet, 2012) was predominantly laid on
reconstruction of complicated objects, in particular, build-
ing roof types. However, not much effort was invested into
a precise and reliable subdivision of the underlying terrain
into different classes. Therefore, at most, few very discrimina-
tive features, such as elevation and color indexes (like
NDVI
,
normalized difference vegetation index), were considered to
separate buildings from trees, roads from vehicles, water bod-
ies from grass areas.
In order to perform a more systematic preparation of data
for the aforementioned reconstruction task, classification of
the underlying terrain is necessary. In real-case scenarios,
there are plenty of factors hindering a correct assignment
of pixels to classes, which would later result in incorrect
building outlines, wriggled street courses, etc. Examples of
assignment errors are sometimes related to seldom or overlap-
ping classes, such as hills covered by shrubbery and grass,
destroyed buildings, bridges in a non-negligible height over
the ground, etc. Even without taking these anomalies into
account or setting them right at a later point, latest develop-
ments brought about an extremely broad spectrum of air-
borne sensors and their products. Taking into account rather
heterogeneous scenes to be captured, this may provide very
particular patterns of texture, distributions of shadows and
types of objects to be classified. Under these circumstances,
it is not realistic to obtain a good classification result with-
out (1) taking into account examples of the data currently
investigated, (2) computing sometimes sophisticated features,
and (3) applying a classification algorithm supposed to learn
thresholds on features for separation of training data and thus
classify the test data. In other words, classification approaches
have three major ingredients: Training data, feature set, and
learning algorithm.
Contributions
In this work, we will focus on the three components men-
tioned above. Firstly, it is interesting to investigate to what
Fraunhofer Institute of Optronics, System Technologies
and Image Exploitation Gutleuthausstr. 1, 76275 Ettlingen,
Germany (
).
Photogrammetric Engineering & Remote Sensing
Vol. 85, No. 3, March 2019, pp. 179–195.
0099-1112/18/179–195
© 2019 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.85.3.179
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2019
179
