See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/268378551

Satellite imagery classification with LIDAR data

Article · May 2012

CITATIONS READS

17 1,965

2 authors, including:

María C. Alonso
University of Alcalá
69 PUBLICATIONS 1,348 CITATIONS

SEE PROFILE

All content following this page was uploaded by María C. Alonso on 26 May 2015.

The user has requested enhancement of the downloaded file.

 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Science, Volume XXXVIII, Part 8, Kyoto Japan 2010

SATELLITE IMAGERY CLASSIFICATION WITH LIDAR DATA

M. C. Alonso *, J. A. Malpica

School of Geodesy and Cartography, University of Alcalá, Ap. Correos 20, 28871 Alcalá de Henares, Madrid, Spain –
(mconcepcion.alonso, josea.malpica)@uah.es

Commission VIII, WG 8

KEY WORDS: LIDAR, Satellite Imagery, Classification, Support Vector Machine, Feature Extraction, SPOT5

ABSTRACT:

This paper shows the potential of LIDAR for extracting buildings and other objects from medium resolution satellite imagery. To
that end, the study integrated multispectral and LIDAR elevation data in a single imagery file and then classified it using the Support
Vector Machine. To determine the method’s potential, the study used a SPOT5 satellite from an area situated southeast of Madrid,
Spain. First, with the four multispectral bands and the panchromatic band of the SPOT5 image, a multispectral four bands
pansharpening was performed with Principal Component Analysis. Once integrated, these four pansharpening images and LIDAR
data, were treated as independent multiple band imagery to perform the classification.
Using five classes, a sample of ground truth pixels was taken for training and testing. The study used 10% of the ground truth for
training and the entire ground truth for testing the robustness of the classification with and without LIDAR data. To assess and
compare the classification results numerically, confusion matrices and Receiver Operating Characteristic (ROC) were calculated for
the five classes, for both classifications, with and without LIDAR.
Generally, when using only multispectral imagery, some confusion among classes occurs; for instance, buildings with flat asphalt
roofs represent a separate problem in classification, since they are extremely difficult to discern from roads. This is mostly solved
when integrating LIDAR data to the multispectral imagery. In general, when adding LIDAR, the classification results show a more
realistic and homogeneous distribution of geographic features than those obtained when using multispectral SPOT5 alone.

1. INTRODUCTION facilitates the identification of the type of object. In this paper,

we show the benefits of synergistically merging sensors for
The automatic extraction of buildings and vegetation in urban mapping buildings. The automatic extraction of buildings and
areas can be used when updating urban inventories to monitor vegetation in urban areas can aid in the important application of
post-disaster emergencies and to assess building or vegetation updating urban inventories to monitor urban vegetation and to
changes. Regardless of the hazard’s origin, post-disaster assess buildings or vegetation changes.
emergency agencies need to assess the geographic extent of the
damage caused by the disaster and prioritize areas that need Several authors have combined multispectral and LIDAR data.
urgent assistance. Remote sensing techniques can perform some Syed et al. (2005) consider object-oriented classification as
damage assessment tasks; the data obtained from remote superior to maximum likelihood in terms of reducing “salt and
sensing can be integrated into a Geographic Information System pepper.” Alia et al. (2008) describe an automated procedure for
and consequently enhance the effectiveness and efficiency of identifying forest species at the tree level from high-resolution
emergency management. imagery and LIDAR data. The results show an improvement in
the ability to discriminate between tree classes. Zeng et al.
Currently, operators manually extract cartographic features, (2002) showed an improvement in the classification of Ikonos
such as buildings, roads, and trees, by visual interpretation of imagery when integrated with LIDAR. This integration of
satellite imagery and aerial photography. Semi-automatic LIDAR with multispectral data has proven beneficial for: the
algorithms to assist cartographic technicians would improve the detection of buildings (Rottensteiner et al., 2003); the
process. Common classification algorithms for low-resolution facilitation of accurate delineation of impervious surfaces
satellite imagery are too limited to deal with complex high- (Hung and Germaine, 2008); change detection and quality
resolution satellite data and require new algorithms. Several control in urban areas (Walter, 2005); the extraction of roads
authors have studied classification methods for satellite images (Hu et al., 2004); and the classification of coastal areas (Lee
(Bernardini et al., 2008; Alonso et al., 2007). and Shan, 2003). The integration of multispectral imagery and
multi-return LIDAR for estimating attributes of trees was
The combination of LIDAR (Light Detection and Ranging) reported in Collins et al. (2004). Our work combines LIDAR
with multispectral imagery can significantly improve elevation data and SPOT5 multispectral data for the
classification, both in terms of accuracy and automation. classification of urban areas. We used the Support Vector
LIDAR provides the altitude necessary for discriminating Machine (SVM), and our results confirm those of the above
between certain classes but is blind to the type of object it works by providing additional data and classification algorithm.
measures, while the multispectral data provides intensity and

* Corresponding author.

730
 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Science, Volume XXXVIII, Part 8, Kyoto Japan 2010

2. MATERIALS AND STUDY AREA and Mathur, 2004; Melgani and Bruzzone, 2004; Theodoridis
and Koutroumbas, 2003; Perkins et al., 2001). These works led
SPOT5 satellite imagery has been used to determine the us to apply SVM rather than other methods for classification.
method’s potential. Launched in 2002, this satellite captures
panchromatic images with a 2.5 m resolution and multispectral 3. CLASSIFICATION OF SPOT5 WITH LIDAR
images with a 10 m resolution (for the bands R, G, and NIR),
and 20 m (for the MIR band). This experiment captured the With the four multispectral bands and the panchromatic image,
scene at 11:20 a.m. on 12 August 2006; it represents an urban we carried out a multispectral four bands pansharpening
area southeast of Madrid, Spain, on mostly flat terrain. Figure 1 performed with PCA (Figure 2).
(a) shows the study area in a false color SPOT image generated
with bands 1, 2, and 3 for the RGB monitor channels. The SPOT5 multispectral pansharpening images and the
LIDAR data are integrated and treated as independent multiple
band imagery to carry out the classification.

Input

Panchromatic
2.5 m LIDAR 1 m

R, G, NIR, MIR
10, 10, 10, 20 m
Resampling and
Registration

Pansharpening

R, G, NIR, MIR,
R, G, NIR, MIR LIDAR, 2.5 m
(a) 2.5 m

Training sample

Supervised Classification by Support Vector Machine

Test sample

Classified image, Classified image,

accuracy 80.34% accuracy 96.30%
Output

Figure 2. Flow chart for the classification

This study tested several provisional numbers of classes and

(b) training samples for the SVM classifier and performed quality
Figure 1. False color SPOT image (© SPOT Image Copyright assessment following the methodology described in Alonso et
2004) (a) and LIDAR elevation data (b) al. (2008). Eventually, we established an arrangement of five
classes: high vegetation (trees, A), low vegetation (shrubs and
An aerial flight for the same area described for the SPOT5 lawns, B), natural grounds (C), artificial grounds (D), and
scene took place also in August 2006 for the acquisition of buildings (E). With the help of aerial photos and ground visits,
LIDAR data. We obtained the density of the LIDAR with an we selected samples for training and evaluating the classifier
average of three points per square meter, subtracted the Digital from the SPOT–LIDAR merged data set. The study area
Terrain Model from the Digital Surface Model, and created the corresponded to the city of Alcalá de Henares where the authors
elevation model shown in Figure 1 (b). As mentioned earlier, reside and allowed for many visits and opportunities for
this shows the LIDAR elevation image of the same area as gathering samples. Finally, a ground truth of 11020 pixels was
Figure 1 (a). selected. Table 1 and Figure 3 show the distribution of the
ground truth.
The algorithm used for the classification was the Support
Vector Machine (SVM); a detailed description of this algorithm For training we took a 10% random subset of the ground truth
can be found in Vapmik (1995). The idea for SVM initially sample. However, we carried out the test with the whole ground
appeared in an article by Boser et al. (1992), in which they truth.
applied it to optical character recognition problems. The authors
Ground truth sample
demonstrated the superior generalization of SVM as compared
with other learning algorithms. SVM maximizes the margin Pixels Percent
between the training patterns and the decision boundary. Trees (A) 2106 19,11
Shrubs and lawns (B) 982 8,91
Several authors have applied SVM to images; for instance,
Azimi-Sadjadi and Zekavat (2000) used SVM to classify 10 Natural grounds (C) 2310 20,96
different cloud and cloudless areas. The SVM has been Artificial grounds (D) 2284 20,73
compared to other classification methods for remote sensing
Buildings. (E) 3338 30,29
imagery such as Neural Networks, Nearest Neighbor,
Maximum Likelihood and Decision Tree classifiers, surpassing Total 11020 100
them all in robustness and accuracy (Huang et al., 2002; Foody Table 1. Class Distribution Summary

731
 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Science, Volume XXXVIII, Part 8, Kyoto Japan 2010

it has been confused with the buildings class. Obviously, the

LIDAR data has discriminated the motorway as ground by the
SVM when classifying with multispectral imagery adding
LIDAR. On the contrary, in the central part of Figure 5 (a) we
can see several buildings detected in Figure 5 (b) while in
Figure 5 (c), these same buildings have been wrongly taken for
artificial ground. Furthermore, the buildings detected in Figure
5 (c) are fuzzier than those in Figure 5 (b); again, the LIDAR
data has allowed sharper discrimination by considering height.
Visits to the terrain prove that the classification in Figure 5 (b)
distinguishes the trees class from the shrubs and lawn class
better than the classification in Figure 5 (c). For instance, note
the red diagonal of Figure 5 (c) and the cross form at the middle
left side of the image to the middle upper side. See Figure 5 (b)
Figure 3. Distribution of the ground truth and the roundabout at the upper right corner. Figure 5 (b)
correctly detects shrubs and lawn (with two trees taken also
correctly by two of the red points), while 5 (c) has wrongly
4. RESULTS AND DISCUSSION mistaken the objects as mostly trees.

Figure 4 (a) shows the results for the classification of the

multispectral image plus LIDAR data, while Figure 4 (b) shows
the classification without LIDAR. A simple visual analysis of
both images reveals the superior delineation of cartographic
features in Figure 4 (a) over those of Figure 4 (b).

(a) (b) (c)

Figure 5. Aerial image for reference (a), results of SVM

classifications with LIDAR (b), and without LIDAR (c)

The upper part of Figure 6 (a) shows several semi-attached

buildings correctly detected in Figure 6 (b), while they are
(a) mistaken as a unique building in Figure 6 (c). Also, Figure 6 (b)
has correctly classified the parking lots as artificial ground
(blue color), while Figure 6 (c) has mistaken them as buildings.

(b)

Figure 4. Results of the SVM classifications with LIDAR (a) (a) (b) (c)
and without LIDAR (b)
Figure 6. Aerial image for reference (a), results of SVM
Figures 5 (a), 6 (a), and 7 (a) show three aerial images classifications with LIDAR (b), and without LIDAR (c)
representing three details from Figure 4 (a) and Figure 4 (b),
represented with rectangles in Figure 4. Figures 5 (b), 6 (b), and Figure 7 (a) reveals a bullfighting arena (the round object
7 (b) show the classifications of the three details using toward the top of the figure). Figure 7 (b) correctly classifies it
multispectral bands plus LIDAR. Figures 5 (c), 6 (c), and 7 (c) as a building, while it has been mistaken as artificial ground in
represent the classifications of multispectral bands without Figure 7 (c). The surroundings of the bullfighting arena
adding LIDAR. corresponding to parking lots have been detected correctly as
artificial ground in Figure 7 (b) and wrongly classified as
The left upper corner of Figure 5 (a) shows a motorway buildings in Figure 7 (c).
correctly detected in Figure 5 (b) but not in Figure 5 (c), where

732
 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Science, Volume XXXVIII, Part 8, Kyoto Japan 2010

With LIDAR Without LIDAR

Per cent
OmissionCommission Omission Commission
Trees (A) 0.71 0.71 4.56 4.87
Shrubs (B) 2.44 0.62 10.59 8.45
Natur.(C) 0.00 1.79 3.68 11.46
Artif. (D) 7.40 5.79 53.72 32.59
Build. (E) 3.27 3.70 19.59 30.59

Table 4. Classification errors for with and without LIDAR

(a) (b) (c) Table 4 shows the buildings class with an omission error of
19.59% and a commission error of 30.59% without LIDAR,
Figure 7. Aerial image for reference (a), results of SVM whereas the addition of LIDAR reduces these errors to 3.27%
classifications with LIDAR (b), and without LIDAR (c) and 3.70%, respectively. This is a consequence of introducing
elevation data, which allows building extraction to take place
To assess and compare the classification results numerically, we more accurately, as seen in the comparison between the
calculated confusion matrices for both classifications, with and classifications presented in Figure 4 (a and b). To a lesser
without LIDAR. Table 2 shows the confusion matrix for the degree than with the buildings, we can observe an improvement
classification of Figure 4 (a), while Table 3 shows the confusion in the classification of vegetated areas between high vegetation
matrix corresponding to Figure 4 (b). (trees) and low vegetation (lawns, grass, or shrubs).
Ground Truth Spectral signatures of some urban materials are similar in
Class (Pixels)
A B C D E Total
composition. For instance, the roofs of some buildings and
Trees (A) 2091 0 0 15 0 2106 roads and parking lots are made of asphalt, therefore the
Shrubs (B) 0 958 0 6 0 964 multispectral classifier (without LIDAR) assigns them to the
Natur. g(C) 0 3 2310 39 0 2352 same class; however, when we consider LIDAR elevation data,
Artif. g (D) 0 21 0 2115 109 2245 the classifier discriminates between buildings and artificial
Build. (E) 15 0 0 109 3229 3353 grounds, and these errors disappear.
Total 2106 982 2310 2284 3338 11020
As a measure of performance of supervised classification, we
Table 2. Confusion matrix for classification with the LIDAR use the receiver operating characteristic (ROC) curve (see
band. Figure 8). See also Metz (1978), for additional information.

Overall Accuracy = (10703/11020) ~ 97.1234% and Kappa Many applications use the area under the ROC curve (AUROC)
Coefficient = 0.9630 as a measure of accuracy (Hanley and McNeil, 1982). For the
five classes considered in this work, the AUROCs are bigger
Ground Truth using LIDAR than without it (Table 5).
Class (Pixels)
A B C D E Total
Trees (A) 2010 78 0 23 2 2113 LIDAR
Area under the ROC curve
Shrubs (B) 79 878 0 2 0 959 With Without
Natur. g(C) 17 1 2225 104 166 2513 Trees (A) 0.97784 0.94885
Artif. g (D) 0 25 0 1057 486 1568
Shrubs and lawns (B) 0.99475 0.99405
Build. (E) 0 0 85 1098 2684 3867
Total 2106 982 2310 2284 3338 11020
Natural grounds (C) 0.93395 0.90910
Artificial grounds (D) 0.82263 0.71577
Table 3. Confusion matrix for classification without LIDAR Buildings. (E) 0.90492 0.80629
Overall Accuracy = (8854/11020) ~ 80.3448% and Kappa
Table 5. Area under the ROC curve for each class with LIDAR
Coefficient = 0.7454
and without LIDAR classification
The overall accuracy and the Kappa coefficients are 97.12%
and 0.96 (see Table 2) for the SVM classification with LIDAR
data, shown in Figure 4 (a), while overall accuracy reaches
80.34% and 0.7454 (see Table 3) for the SVM classification
without LIDAR data, as shown in Figure 4 (b). Therefore, a
gain of 16.77% is obtained in merging the LIDAR data with the
multispectral bands for this data set. The confusion matrices
allow us to examine the differences of the classifications in
detail. For all the classes considered, the omission and
commission errors are improved to some degree by the
integration of LIDAR (see Table 4).

733
 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Science, Volume XXXVIII, Part 8, Kyoto Japan 2010

REFERENCES

Alia, S.S., Dareb, P., Jonesc, S.D., 2008. Fusion of remotely

sensed multispectral imagery and Lidar data for forest structure
TP TP assessment at the tree level. In: The International Archives of
the Photogrammetry, Remote Sensing and Spatial Information
Sciences, Beijing, Vol. XXXVII. Part B7, pp 1089-1094.

Alonso, M.C. and Malpica, J.A, 2008. Classification of

Multispectral High-Resolution Satellite Imagery Using LIDAR
Trees (A) Shrubs and Elevation Data. In Bebis G. et al. (Eds.): Advances in Visual
lawns (B) Computing ISVC08, Part II, Lecture Notes in Computer Science
5359, pp. 85–94. Springer-Verlag.
FP FP Alonso, M.C., Sanz, A. and Malpica, J.A., 2007. Classification
of high resolution satellite images using texture. In Bebis G. et
TP
al. (Eds.): Advances in Visual Computing ISVC07. Part II LNCS
4842, pp. 499–508. Springer-Verlag.

Azimi-Sadjadi, M.R. and Zekavat, S.A., 2000. Cloud

Natural classification using support vector machines. In: Proceedings of
grounds (C) the 2000 IEEE Geoscience and Remote Sensing Symposium
IGARS, Vol. 2, pp. 669 – 671.

Bernardini, A., Malinverni, E. S., Zingaretti, P. and Mancini,

FP A., 2008. Automatic classification methods of high-resolution
satellite images: the principal component analysis applied to the
TP TP sample training set. In: The International Archives of the
Photogrammetry, Remote Sensing and Spatial Information
Sciences, Beijing, Vol. XXXVII. Part B7, pp 701-706.

Boser, B.E., Guyon, I. and Vapnik, V.N., 1992. A Training

Artificial Algorithm for Optimal Margin Classifiers. In: D. Haussler,
Buildings. (E) editor, Proceedings of the Fifth Annual Workshop on
grounds (D)
Computational Learning Theory pp. 144-152, Pittsburgh. ACM
Press.
FP FP
Collins, C.A., Parker, R.C. and Evans, D.L., 2004. Using
Figure 8. ROC curves in the SVM classifications with (straight multispectral imagery and multi-return LIDAR to estimate trees
line) and without (dotted line) LIDAR. TP and FP are true and stand attributes in a southern bottomland Hardwood forest.
positive and false positive In: ASPRS Annual Conference Proceedings, Denver, Colorado
(on CD-ROM).

5. CONCLUSIONS Foody, G.M. and Mathur, A., 2004. A Relative Evaluation of

Multiclass Image Classification by Support Vector Machines.
This study investigated the impact of LIDAR data on the In: IEEE Transactions On Geoscience And Remote Sensing,
classification of multispectral SPOT5 imagery using SVM over Vol. 42(6), pp. 1335-1343.
an urban area near Madrid, Spain. A visual evaluation shows an
improvement of the classification when using LIDAR data. Hanley, J.A. and McNeil, B.J., 1982. The Meaning and Use of
Two numerical methods—confusion matrices and ROC the Area under a Receiver Operating Characteristic (ROC)
curves—complemented this evaluation of the results. After Curve. Radiology, Vol. 143, pp. 29-36.
defining ground truth samples, we used 10% randomly for
Hu, X., Tao C.V. and Hu, Y, 2004. Automatic road extraction
training for both classifications (with and without LIDAR),
from dense urban area by integrated processing of high
while the evaluation was made with the total ground truth
resolution imagery and lidar data. In: International Society for
samples. We obtained a gain of 16.78% in overall accuracy for
Photogrammetry and Remote Sensing (ISPRS), Istanbul,
the imagery in this experiment. Furthermore, ROC curves for
Turkey.
the classification with LIDAR reveal a superior performance for
all classes compared to the classifications without it. Huang, C., Davis, L.S. and Townshed, J.R.G., 2002. An
assessment of support vector machines for land cover
The SVM classifier with LIDAR shows a more realistic and classification. International Journal of Remote Sensing, Vol.
homogeneous distribution of geographic features than that 23, pp. 725–749.
obtained without LIDAR. This is specifically so for buildings
and artificial ground, as shown by the ROC curves. Hung, M.C. and Germaine, K., 2008. Delineating Impervious
Surfaces Utilizing High Spatial Resolution Multispectral

734
 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Science, Volume XXXVIII, Part 8, Kyoto Japan 2010

Imagery and Lidar Data. In: ASPRS 2008 Annual Conference,

Portland, Oregon.

Lee, D.S. and Shan, J., 2003. Combining Lidar Elevation Data
and IKONOS Multispectral Imagery for Coastal Classification
Mapping. Marine Geodesy, an International Journal of Ocean
Surveys, Mapping and Remote Sensing, Vol. 26(1-2), pp.117-
127.

Melgani, F. and Bruzzone, L., 2004. Classification of

Hyperspectral Remote Sensing Images With Support Vector
Machines. In: IEEE Transactions On Geoscience And Remote
Sensing, Vol. 42(8), pp. 1778–1790.

Metz, C.E., 1978. Basic principles of ROC analysis. Seminars

Nuclear Medicine, Vol. 8(4), pp. 283-298.

Perkins, S., Harvey, N.R., Brumby, S.P. and Lacker, K. 2001.

Support Vector Machines for broad area feature classification in
remotely sensed images. In: Proceeding of Algorithms for
multispectral, hyperstral, and ultraspectral imagery, Vol. 4381,
pp. 286-295, Orlando USA.

Rottensteiner, F., Trinder, J., Clode S. and Kubic, K., 2003.

Building dectection using LIDAR data and multispectral
images. In: Proceedings of DICTA, Sydney, Australia, pp. 673–
682.

Syed, S., Dare, P. and Jones, S., 2005. Automatic classification

of land cover features with high resolution imagery and Lidar
data: an objectoriented approach. Proceedings of SSC2005
Spatial Intelligence, Innovation and Praxis: The national
biennial Conference of the Spatial Sciences Institute,
Melbourne, Australia, Sep.

Theodoridis, S. and Koutroumbas, K., 2003. Pattern

Recognition. Second Edition. Elsevier Academic Press.

Vapmik, V., 1995. The Nature of Statistical Learning Theory.

Springer Verlag. New York.

Walter, V., 2005. Object-based classification of integrated

multispectral and LIDAR data for change detection and quality
control in urban areas. In: Proceedings of the ISPRS WG
VII/1Human Settlements and Impact Analysis.

Zeng, Y., Zhang, J., Wang, G. and Lin, Z., 2002. Urban landuse
classification using integrated airborn laser scanning data and
high resolution multispectral imagery, Pecora 15/Land Satellite
Information IV/ISPRS Commssion I/FIEOS.

ACKNOWLEDGEMENTS

The authors wish to thank the Spanish National Mapping

Agency (Instituto Geográfico Nacional) for providing the
images and LIDAR data for this research. The authors are also
grateful to the Spanish MICINN for financial support to present
this paper in ISPRS; project number CGL2009-13768.

735

View publication stats