Professional Documents
Culture Documents
Hybridmodeltooptimizeobjectbasedlandcoverclassificationbymetaheuristicalgorithmanexampleforsupportingurbanmanagementin Ha Noi Viet
Hybridmodeltooptimizeobjectbasedlandcoverclassificationbymetaheuristicalgorithmanexampleforsupportingurbanmanagementin Ha Noi Viet
net/publication/328638063
CITATIONS READS
0 237
2 authors, including:
SEE PROFILE
All content following this page was uploaded by Manh Van Pham on 31 October 2018.
Quang-Thanh Bui, Manh Pham Van, Nguyen Thi Thuy Hang, Quoc-Huy
Nguyen, Nguyen Xuan Linh, Pham Minh Hai, Tran Anh Tuan & Pham Van Cu
To cite this article: Quang-Thanh Bui, Manh Pham Van, Nguyen Thi Thuy Hang, Quoc-Huy
Nguyen, Nguyen Xuan Linh, Pham Minh Hai, Tran Anh Tuan & Pham Van Cu (2018): Hybrid model
to optimize object-based land cover classification by meta-heuristic algorithm: an example for
supporting urban management in Ha Noi, Viet Nam, International Journal of Digital Earth, DOI:
10.1080/17538947.2018.1542039
1. Introduction
Urban management activities require an accurate spatial/temporal database and an effective manage-
ment strategy to minimize adverse effects (Herold and Sawada 2012; Jonkman and Dawson 2012). In
these database, land cover maps are very important as they provide preliminary information about
the spatial variation of surface types. Therefore, adequate detection of urban landscape plays a con-
siderable role in the proper handling of certain types of land covers. It also helps the decision-making
process in coping various environmental hazards, such as excessive surface runoff and urban flood-
ing. The identification of land cover types can significantly contribute to the efficient management
and reduction of damages to the settlement, agriculture, and livelihood by avoiding construction and
developments in hazard-prone areas.
Methods for land cover classification methods can be divided into two branches, which are pixel-
based and object-based classification. The latter has significantly grown in the last several years with
CONTACT Quang-Thanh Bui thanhbq@vnu.edu.vn; qthanh.bui@gmail.com VNU University of Science, 334 Nguyen Trai
Road, Ha Noi, Viet Nam
© 2018 Informa UK Limited, trading as Taylor & Francis Group
2 Q.-T. BUI ET AL.
the rapid emergence of high spatial resolution, i.e. (smaller than 4 m). Applications of this approach
can be found in many studies that have shown the superior performance of object-based classification
over the traditional pixel-based approach (Duro, Franklin, and Dubé 2012; Myint et al. 2011). In fact,
object-based classification is sensitive to landscape morphology and is found to be a robust method in
an urban study where the human-made structure is present at a high density (Qian et al. 2015).
Typically, the conventional object-based classification process comprises of image segmentation and
classification of segments based on predefined class properties. Depending on specific locations, three
typical parameters, which are scale, compactness, and shape are be defined, mainly by user’s experi-
ences (Qian et al. 2015). The definition of those parameters controls how pixels or existing objects
are grouped and ultimately impacts the overall accuracy of classification tasks. In addition to built-
in algorithms (rule-based method) in popular image analysis packages (W. Zhou and Troy 2008),
machine learning has been increasingly used to fine tune the parameters of classifiers. This approach
could be found in (Hamedianfar et al. 2014) with the C4.5 algorithm in urban mapping from hyper-
spectral images, or the applications of Support Vector Machine (SVM) with high resolution images in
review papers of (Mountrakis, Im, and Ogole 2011) and (Petropoulos, Kalaitzidis, and Prasad Vadrevu
2012). Another common technique could also be found in (Belgiu and Drăguţ 2016; Gong, Im, and
Mountrakis 2011) where random forest and variances of neural network (including deep learning) sig-
nificantly improved classification accuracies of the regular object-based method. Some methods also
introduced and evaluated ensemble techniques to strengthen weak classifiers as in (Chan and Paelinckx
2008). From the literature, it could be seen that few studies employed the power of metaheuristic algor-
ithms in optimizing image classification methods, except for some recent works of (Dou et al. 2015).
Recently, researchers have enriched optimization libraries with fast-developed algorithms, mainly by
nature-inspired algorithms. In fact, no algorithm is capable of solving all optimization problems and
the examinations of new algorithms to specific applications are therefore necessary.
This study posed several questions on how meta-heuristic algorithms could be used for fine tun-
ing classifiers to improve classification performance and how the variation in sizes of training data
could impact classification overall accuracy and individual accuracies for each of four surface type
extractions. To partly answer these questions, this study proposed a novel hybrid model, named
GMNN that employs Grasshopper Optimization Algorithm (GOA) to optimize parameters of the
multiple-class neural network (MNN) for urban pattern detection. GOA was chosen because despite
being successfully tested against benchmarked theoretical functions, the algorithm has yet been
examined in a real application.
Hanoi, the capital of Vietnam was selected as the subject of a case study because of its complex
surface morphology and patterns of land cover types. GMNN was run with different training and
validation data sizes, and the classified outputs were compared to several benchmarked classifiers
such as Support Vector Machine (SVM), Decision Tree (DT) (Qian et al. 2015), and MNN optimized
by two other new algorithms such as Grey Wolf Optimization (GWO), Biogeography-based Optim-
ization (BBO) through conventional statistical indicators. Throughout the experiments, PCI Geo-
matics (Trial mode) was used for image segmentation and feature extraction; Matlab was used for
modeling; QGIS was used for map visualization. Detail description of the procedure and results
are in the following sections, including Data and method in the second part, Results, and discussion
in the third and Conclusion in the final section.
it comprises of (1) a central area including ancient and French-style districts; (2) a new urban area
with high-rise building in the western part where impervious surfaces are dominant with mixed
patches of open space; (3) the remaining areas are a mixture of urban-rural styles in outskirt of
the city, in which bare soil and agricultural are commonly found (Figure 1).
where S is the synthesized image, N is the total number of MS bands, vi is a scalar weight, MSi is the
up-scaled i-th MS band. vi is determined by minimizing the least square error of ||P-S|| (Yuri 2002).
Figure 2. Image segmentation at Scale: 30, Shape: 0.75, and Compactness: 0.5.
attributes associated with image segments, such as default Mean value, brightness, Mean-difference-
to-super-object, etc. in eCognition or with custom measurements in OpenCV (Wieland and Pittore
2014). Sometimes, these attributes were further processed to filter redundant variables and to reduce
dimensionality by using embedded algorithms in commercial image processing software or by using
machine learning techniques (Dou et al. 2015; Hamedianfar et al. 2014). In this study, we used all
attributes that were generated from the image processing procedure using PCI Geomatics. Detail
description of variables was presented in (Table 2)
values between [0,1] to define random behaviors of grasshopper; (6) function s defines the social
forces; (7) ub, lb are lower and upper bounds of the variables; (8) c is a decreasing coefficient to
shrink the comfort zone and is defined as c = Cmax − l(Cmax − Cmin)/L with Cmax, Cmin are
two predefined parameters, l is current iteration and L is maximum number of iterations; (9) T d
is the best location and it is updated after each iteration. Detail description of X, S,A and s-function
can be found in (Saremi, Mirjalili, and Lewis 2017). Si , Gi and Ai are further defined and updated
after each iteration. In short, the procedure of GOA for optimizing urban pattern classification in this
study is showed as follows:
. Initializing the swarm Xid in d-dimensional space and normalizing the distances between grass-
hoppers, in which i = 1 … N; N is number of grasshoppers or population size; The population
size was determined based on trial-error test of RMSE
. Initializing model parameters including Cmax, Cmin and a maximum number of iteration. The
fitness of search agent was calculated by using Equation (3). Preliminary best agent T was defined.
. Updating the position of each grasshopper by Equation (5); Recalculate RMSE and compare it to
the previous value. Update T if a better solution was found.
. Checking the current position of grasshoppers if they were out of the predefined boundaries
. The iteration continued until reaching max iteration or RMSE reaches predefined value. The final
position of the swarm was the optimal solution. These dimensional values would be used in mul-
tiple class neural network to classify land cover pattern in the study area.
and the random value is 0.5. The model performs well if the AUC is larger than 0.5 and by versus. In
fact, the higher the AUC is, the better the model performs.
Predicted value
Yes No
Actual value Yes True Positive (TP) False Negative (FN)
No False Positive (FP) True Negative (TN)
Also, Root Mean Square Error (RMSE), which measures the square root of the difference between
observed and predicted values, Mean Absolute Error (MAE), which measures the difference between
observed and predicted values, and Kappa index, which measures classification accuracy could give
some different views for model comparison.
Table 4. Accuracy assessment for 4000 training objects (best values in shaded cells).
Classifier/Base classifier Optimization algorithm RMSE MAE Kappa Overall Accuracy
MNN GWO 0.2477 0.0902 0.7685 83.2381
MNN BBO 0.2335 0.1129 0.7843 84.4762
MNN GOA 0.2225 0.1086 0.8284 87.1905
DT 0.2345 0.1244 0.7848 84.4771
SVM 0.2564 0.1221 0.7503 82.0012
INTERNATIONAL JOURNAL OF DIGITAL EARTH 11
An additional test was carried out to determine whether the performance of models was statisti-
cally different. Non-parametric Wilcoxon signed-rank test was chosen as the dependent variable was
measured in nominal scale. Each pair of two models was tested by using AUC values with Null
hypothesis stating that there was no difference between the predictive capability of the two selected
model. The results in (Table 7) show that all p values are smaller than 0.05 that the null hypothesis
would be rejected and the differences were statistically significant. The differences in performance
between classification methods could be caused by either the predictive capability of the techniques
or the variation within the dataset. Nonetheless, in this study, several training datasets randomly
selected across heterogenous urban morphology and then classified by our proposed model and
benchmark classifiers show that, from all trials, GMNN significantly improved both the overall accu-
racy and detection accuracy of each land surface. Therefore, with its robustness, this method can be
Table 5. Accuracy assessment for 2000 training objects (best values in shaded cells).
Classifier/Base classifier Optimization algorithm RMSE MAE Kappa Overall Accuracy
MNN GWO 0.2337 0.0846 0.798 0.853333
MNN BBO 0.2241 0.112 0.799 0.854667
MNN GOA 0.2217 0.0866 0.8136 0.869300
DT 0.2375 0.1309 0.7759 0.838667
SVM 0.258 0.1238 0.7348 0.809333
Table 6. Accuracy assessment for 400 training objects (best values in shaded cells).
Classifier/Base classifier Optimization algorithm RMSE MAE Kappa Overall Accuracy
MNN GWO 0.2826 0.1033 0.7242 0.802
MNN BBO 0.2447 0.1043 0.777 0.840
MNN GOA 0.2373 0.1029 0.7836 0.844
DT 0.2414 0.1299 0.7715 0.836
SVM 0.267 0.1316 0.7353 0.806
used in urban applications to partially satisfy the requirements of planners in detecting detailed
change patterns in a metropolitan area.
5. Map generation
To visually display the classification capabilities of GMNN, we produced three maps using the
trained models with three different training and validation datasets. (Table 8) shows the sample out-
put of the model, as each object was assigned four probability values of being assigned to one of the
four classes. For the entire study area, more than 54.000 objects were classified based on probability
values. The results are shown in (Figure 9). It could be seen that three maps were almost identical in
water class. The misclassified objects were mainly between impervious and bare soil because of the
mixture between two types. Although overall accuracies of three cases were quite similar, the spatial
placement of misclassified objects resulted in more fragmented patterns in maps associated with
smaller training datasets.
Figure 9. Subsets of classified maps by GMNN with different training and validation sizes. From left to right: 4000 objects, 2000
objects, 400 objects.
around 87% for a spatially varied study area is the affirmative reasoning for its applicability in other
urban regions. Though this value was measured by multiplying extraction accuracies of four classes,
the extraction accuracy of any single class was much higher, and that would be very useful for specific
applications of certain types of urban land cover.
The consistency of classification algorithms is normally subject to variation of training sample
sizes. However, it was not the case in this study as GMNN seemed to be robust in all three training
and validation datasets. The success of this model was dependent on the proper selection of model
parameters as they were determined through trial-error processes, and these values were found suit-
able for the dataset of the study area. Therefore, settings might need to be changed when this method
is used with another study area with diverse surface morphology. The performance of this hybrid
model indicates a vast potential for the implementation of meta-heuristic algorithms in object-
based classification. In this case, GOA was more powerful than GWO and BBO, but it might be
different if another dataset was used.
Finally, this model was applied for high spatial resolution data but might be applicable for a var-
iety of mid-resolution satellite data such as Landsat or Sentinel. Therefore, more effort should be
spent on examining those kinds data and automating the image processing workflow to derive
land cover types, particularly in larger urban areas.
Disclosure statement
No potential conflict of interest was reported by the authors.
Funding
This research is funded by Vietnam National Foundation for Science and Technology Development (NAFOSTED)
under Grant Number [105.99-2016.05].
ORCID
Quang-Thanh Bui http://orcid.org/0000-0002-5059-9731
References
Abhishek G Neve, G. M. K. a. O. K. 2017. “Application of Grasshopper Optimization Algorithm for Constrained and
Unconstrained Test Functions.” International Journal of Swarm Intelligence and Evolutionary Computation 6 (3).
doi:10.4172/2090-4908.1000165.
INTERNATIONAL JOURNAL OF DIGITAL EARTH 15
Belgiu, M., and L. Drăguţ. 2016. “Random Forest in Remote Sensing: A Review of Applications and Future Directions.”
ISPRS Journal of Photogrammetry and Remote Sensing 114 (Supplement C): 24–31. https://doi.org/10.1016/j.
isprsjprs.2016.01.011.
Chan, J. C.-W., and D. Paelinckx. 2008. “Evaluation of Random Forest and Adaboost Tree-Based Ensemble
Classification and Spectral Band Selection for Ecotope Mapping Using Airborne Hyperspectral Imagery.”
Remote Sensing of Environment 112 (6): 2999–3011. https://doi.org/10.1016/j.rse.2008.02.011.
Dou, J., K.-T. Chang, S. Chen, P. A. Yunus, J.-K. Liu, H. Xia, and Z. Zhu. 2015. “Automatic Case-Based Reasoning
Approach for Landslide Detection: Integration of Object-Oriented Image Analysis and a Genetic Algorithm.”
Remote Sensing 7 (4). doi:10.3390/rs70404318.
Duro, D. C., S. E. Franklin, and M. G. Dubé. 2012. “A Comparison of Pixel-Based and Object-Based Image Analysis
with Selected Machine Learning Algorithms for the Classification of Agricultural Landscapes Using SPOT-5 HRG
Imagery.” Remote Sensing of Environment 118 (Supplement C): 259–272. https://doi.org/10.1016/j.rse.2011.11.020.
Gong, B., J. Im, and G. Mountrakis. 2011. “An Artificial Immune Network Approach to Multi-Sensor Land use/Land
Cover Classification.” Remote Sensing of Environment 115 (2): 600–614. https://doi.org/10.1016/j.rse.2010.10.005.
Hamedianfar, A., H. Z. M. Shafri, S. Mansor, and N. Ahmad. 2014. “Combining Data Mining Algorithm and Object-
Based Image Analysis for Detailed Urban Mapping of Hyperspectral Images.” Journal of Applied Remote Sensing 8
(1): 085091. doi:10.1117/1.JRS.8.085091.
Herold, S., and M. C. Sawada. 2012. “A Review of Geospatial Information Technology for Natural Disaster
Management in Developing Countries.” International Journal of Applied Geospatial Research 3 (2): 24–62.
doi:10.4018/jagr.2012040103.
Huang, X., C. Xie, X. Fang, and L. Zhang. 2015. “Combining Pixel- and Object-Based Machine Learning for Identification of
Water-Body Types From Urban High-Resolution Remote-Sensing Imagery.” IEEE Journal of Selected Topics in Applied
Earth Observations and Remote Sensing 8 (5): 2097–2110. doi:10.1109/JSTARS.2015.2420713.
Im, J., Z. Lu, J. Rhee, and L. J. Quackenbush. 2012. “Impervious Surface Quantification Using a Synthesis of Artificial
Immune Networks and Decision/Regression Trees From Multi-Sensor Data.” Remote Sensing of Environment 117
(Supplement C): 102–113. https://doi.org/10.1016/j.rse.2011.06.024.
Jonkman, S., and R. Dawson. 2012. “Issues and Challenges in Flood Risk Management – Editorial for the Special Issue
on Flood Risk Management.” Water 4 (4): 785–792.
Lary, D. J., A. H. Alavi, A. H. Gandomi, and A. L. Walker. 2016. “Machine Learning in Geosciences and Remote
Sensing.” Geoscience Frontiers 7 (1): 3–10. https://doi.org/10.1016/j.gsf.2015.07.003.
Mirjalili, S., S. M. Mirjalili, and A. Lewis. 2014. “Grey Wolf Optimizer.” Advances in Engineering Software 69: 46–61.
http://doi.org/10.1016/j.advengsoft.2013.12.007.
Mountrakis, G., J. Im, and C. Ogole. 2011. “Support Vector Machines in Remote Sensing: A Review.” ISPRS Journal of
Photogrammetry and Remote Sensing 66 (3): 247–259. https://doi.org/10.1016/j.isprsjprs.2010.11.001.
Myint, S. W., P. Gober, A. Brazel, S. Grossman-Clarke, and Q. Weng. 2011. “Per-pixel vs. Object-Based Classification
of Urban Land Cover Extraction Using High Spatial Resolution Imagery.” Remote Sensing of Environment 115 (5):
1145–1161. https://doi.org/10.1016/j.rse.2010.12.017.
Nguyen, T. H. T., V. T. Tran, Q. T. Bui, Q. H. Man, and T. d. V. Walter. 2016. “Socio-economic Effects of Agricultural
Land Conversion for Urban Development: Case Study of Hanoi, Vietnam.” Land Use Policy 54: 583–592. http://doi.
org/10.1016/j.landusepol.2016.02.032.
Petropoulos, G. P., C. Kalaitzidis, and K. Prasad Vadrevu. 2012. “Support Vector Machines and Object-Based
Classification for Obtaining Land-use/Cover Cartography From Hyperion Hyperspectral Imagery.” Computers &
Geosciences 41 (Supplement C): 99–107. https://doi.org/10.1016/j.cageo.2011.08.019.
Qian, Y., W. Zhou, J. Yan, W. Li, and L. Han. 2015. “Comparing Machine Learning Classifiers for Object-Based Land
Cover Classification Using Very High Resolution Imagery.” Remote Sensing 7 (1). doi:10.3390/rs70100153.
Saremi, M., S. Mirjalili, and A. Lewis. 2017. “Grasshopper Optimisation Algorithm: Theory and Application.”
Advances in Engineering Software 105 (Supplement C): 30–47. https://doi.org/10.1016/j.advengsoft.2017.01.004.
Simon, D. 2008. “Biogeography-Based Optimization.” IEEE Transactions on Evolutionary Computation 12 (6): 702–
713. doi:10.1109/TEVC.2008.919004.
Wieland, M., and M. Pittore. 2014. “Performance Evaluation of Machine Learning Algorithms for Urban Pattern
Recognition From Multi-Spectral Satellite Images.” Remote Sensing 6 (4): 2912–2939. doi:10.3390/rs6042912.
Yuri, Z. (2002, 24-28 June 2002). A new Automatic Approach for Effectively Fusing Landsat 7 as Well as IKONOS
Images. Paper presented at the IEEE international geoscience and remote sensing symposium.
Zhou, W., and A. Troy. 2008. “An Object-Oriented Approach for Analysing and Characterizing Urban Landscape at
the Parcel Level.” International Journal of Remote Sensing 29 (11): 3119–3135. doi:10.1080/01431160701469065.
Zhou, W., and A. Troy. 2009. “Development of an Object-Based Framework for Classifying and Inventorying Human-
Dominated Forest Ecosystems.” International Journal of Remote Sensing 30 (23): 6343–6360. doi:10.1080/
01431160902849503.