Landuse Classification of Multispectral Satellite Images using Particle Swarm Optimization Lizhier B. Coralde

Vladimir Y. Mariano

Institute of Computer Science University of the Philippines Los Banos, College, Laguna, 4031

Institute of Computer Science University of the Philippines Los Banos, College, Laguna, 4031

[email protected]

[email protected]

ABSTRACT Among the multiple benefits and uses of remotely sensed satellite images, one of the most important has been its use in landuse classification. In this paper, Particle Swarm Optimization (PSO), an alternative and evolutionary approach was used to classify land use of multispectral satellite images. A medium resolution LANDSAT TM 7 data was used in the experiment. Experimental results show that the proposed algorithm has high classification precision and is promising in classifying regions/pixels according to landuse.

Keywords Landuse classification, Geographical Information Systems, remote sensing, multispectral satellite images, Landsat, Particle Swarm Optimization.

1. INTRODUCTION Landuse can be regarded as one of the vital elements in planning and management in the physical and natural sciences. At present, its dynamic and substantial transformation and change has caused alarm to most scientists (i.e. geographers, climatologists, hydrologists, foresters etc.) due to its potential impact on biosphere-atmosphere and hydrologic cycles [11]. Urbanization and rapid deforestation which was scientifically proven to contribute to severe weather anomalies and phenomena like ENSO (El Nino Southern Oscillation), La Nina and Climate Change can be readily detected by effective classification procedures utilizing appropriate satellite images. Several years ago, acquisition of multi-spectral satellite images from international sources was difficult and expensive. However, in the advent of advancement in hardware and software technology, multi and hyper spectral satellite images of varying temporal and spatial resolution can be freely requested from designated government institutions and/or downloaded from several web sites. In the country, PAGASA (Philippine Atmospheric Geophysical and Astronomical Services Administration) has been acquiring multi-spectral images from international sources since their establishment for forecasting use. At present, the establishment of new ground receiving stations as part of their modernization program facilitates the acquisition of high spatial resolution multi-spectral images gathered by recently launched satellite imaging sensors such as EarthSat and EnviSat. These satellite images generally consist of hundreds of spectral bands and only a few were used for weather and flood

forecasting purposes. The unused bands which may contain spectral properties and signatures significant to landuse and land cover classification remains unutilized. Land use classification from multi-spectral satellite images is often a difficult and complicated task. As the amount of data increases and as the characteristics of the images become more complex, the more intricate the classification process would become. Through time, efforts were done to improve the classification process. Several free and commercial image processing software were developed. In addition, several algorithms and methodologies were also applied for pixel based cluster analysis and classification. The classification problem can also be seen as a clustering problem, which in this study pertains to the grouping of pixels in the multispectral space. Pixels that belong in the same cluster therefore have similar spectral values and to quantify the relationship between them, similarity measure is necessary. The most obvious measure of similarity between two pixels is the distance between them [6]. In this paper, a new technique called Particle Swarm Optimization (PSO) is presented to classify land use of multispectral satellite images. PSO is an alternative populationbased evolutionary computation technique which was inspired by the social behavior of bird flocking or fish schooling. In PSO, the population consists of potential solutions, called particles, similar to birds in a flock. PSO is initialized with a group of random particles (solutions) and then searches for the optimal solution by updating generations. All particles have fitness elite values which are evaluated by the fitness function to be optimized, and have velocities which direct the flying of particles. While flying, every particle updates its velocity and position based on its own best experience and that of the entire population. The updating policy will cause the particle swarm to move toward a region with higher object value [16].

2. PARTICLE SWARM OPTIMIZATION Particle swarm optimization (PSO) is a population-based stochastic optimization technique inspired by social behavior of bird flocking or fish schooling. PSO has been successfully applied in many areas such as function optimization, artificial nueral network training, fuzzy system control, and other areas. In PSO, each single solution that is represented by a bird in the search space is called the “particle”. All particles have fitness

values which are evaluated by the fitness function to be optimized, and have velocities which direct the flying. The particles fly through the problem space by following the current optimum particles. Normally, a group of random particles (solutions) initializes the PSO and then searches for optima by updating generations. In every iteration, each particle is updated by following two “best” values, i.e. pbest and gbest. Pbest is the best solution the particle has achieved so far while gbest is the global best value achieved by the whole population. Lbest or local best is another best value that could replace gbest. This is obtained when only a subset of the population is taken by the particle as its topological neighbor.

band) were then plotted in a graph (Figure 2). This is to examine their natural pattern and determine the maximum number of classification that it can effectively identify. Those classes that share the same spectral values were combined in one class. After assessing the properties of the area and the limited resolution of the satellite images, five major landuse classes were decided namely bare soil, vegetation/forest, urban, inland water and sea.

After finding the two best values, the particle updates its velocity values and positions with the following equation (a) and (b). v[] = v[] + c1 * rand() * (pbest[]-present[]) + c2 * rand() * gbest[]-present[])

(a)

present[] = present[] + v[]

(b)

where v[] is the particle velocity, present[] is the current particle (solution), pbest[] and gbest[] are defined as stated before, rand() is a random number between (0, 1), and c1 and c2 are learning factors.

Figure 1. LANDSAT 7 Image of Southern Luzon

3. MULTISPECTRAL LANDUSE CLASSIFICATION USING PSO In this study, an alternative method to solve the landuse classification of multispectral satellite images using Particle Swarm Optimization was presented. First, the training samples (pixels) were collected from the image. These samples were then used to train the PSO-based classifier to find the optimum centroid of the class. The centroid of a class is a point in the multispectral space which has the least distance to all the data. The results of the training were then used to classify the test data.

(a) bare soil

3.1 Data Collection The image used in the experiment is a LANDSAT 7 multispectral image of Southern Luzon as shown in figure 1. It has dimensions of 8740 by 7728 pixels and an area approximately 34,000 square kilometers. Bands 1, 2, 3, 4, 5 and 7 were used in the experiment since these are the only bands useful in landuse classification. The training and testing data were selected using ENVI 4.3 Image Processing Software. Before collecting the data, a reference image was first created to aid in the validation of the true classification. The ENVI 4.3’s built-in ISODATA (Iterative elf-Organizing Data Analysis Technique) classification algorithm was used. The minimum and maximum numbers of classes specified were 3 and 6 with maximum number of iterations and threshold equal to 10 and 0.95, respectively. The resulting reference image plus manual analysis were used as guide in collecting the ground truth data. After collecting the sample data for each class, their spectral values (values in each

(b) inland water

(c) sea

Table 2. PSO parameters Individualistic factor

C1 = 0 - 1.426939

Socialistic factor

C2 = 0 - 1.426939

Inertia Weight

0.689343

Number of Swarm Particles

20

Maximum number of iterations

10000 Best fitness = no of samples * dimension

(d) urban

End condition

3.3 Classification Classification of data is performed by comparing its spectral values with the centroids generated by the PSO-based program for each class. It is assigned to the class which has the least distance.

4. RESULTS AND DISCUSSION Figure 2. Spectral values of the five classes The number of pixels collected for each class are shown in table 1. Table 1. Number pixels used for training and testing

Bare Soil

Training Sample 420

Test Data 200

Vegetation/Forest

974

200

824

Class

Total 620 1174

Urban

624

200

Inland Water

896

200

1096

Sea

875

200

1075

3.2 PSO Training The training samples of a class are then fed into a standard PSO program to find its optimum centroid c = {c1, c2,…,cn} or clustering center in the n-dimensional space. The fitness of a particle x = {x1, x2,…, xn} is computed by the following formula: m n

F(x) =∑ ∑ | dij- xj | i=1 j=1

where m is the total number of sample data d with n dimensions. The main goal of the optimization is to find the particle x (the centroid) which has the least distance to all the samples. The parameters used in the training are shown in table 2. New positions and velocities are calculated using formula (a) and (b). A local neighborhood was also adapted. Update of positions and velocities are repeated until the specified amount of fitness is satisfied or the maximum number of iteration was already reached. The training of each class was run 10 times and the best values are then noted. This procedure is done for each class.

It was observed that the inclusion of all spectral values in training the classes is not as effective as when only a subset of the bands is used. When all the spectral values were used, the PSO was not able to converge or was not able to find a solution. Bands labeled in the figure as 2, 3 and 4 are the best features used since they were able to discriminate the 5 classes from each other. The confusion matrix generated by the PSO-based classifier is tabulated in table 3. It can be observed that the classifier has a very high accuracy with an overall accuracy of 92.7%. Table 3. Confusion Matrix of the PSO-based Classification PSO-based Classification Results

uthTrndouGr

(e) vegetation/forest

Veg

Land

Urban

Water

Sea

Veg Land

100 0

0 93.5

0 6.5

0 0

0 0

Urban Water Sea

0 0 0

0 0 0

92.5 0 0

5.5 100 0

2 0 100

5. CONCLUSION This paper has presented a new PSO-based technique to classify landuse in medium resolution multispectral satellite images. Experimental results show that the proposed algorithm has high classification precision and is promising in classifying landuse. The high precision of the classifier is attributed to the knowledge of the natural pattern and limitation of the data. Future work will explore new datasets including higher resolution images to further test the efficiency and fully explore the capacity of the PSO-based classifier.

6. ACKNOWLEDGEMENTS The authors thank the College of Arts and Sciences and the Institute of Computer Science UPLB for their financial support through CAS-TF #8217300 and ICS-GF #2326103, respectively.

7. REFERENCES [1] ANGELINE, P. 1998. Evolutionary Optimization versus Particle Swarm Optimization: Philosophy and Performance Differences. Seventh Annual Conference on Evolutionary Programming, 1998, pp. 601-611. [2] CIVCO, D.L., 1993. Artficial nueral networks for landcover classification and mapping. International Journal of Geographical Information Systems, 7(2), pp. 173-186. [3] CHANDRAMOULI, K. and IZQUIERDO, E. 2006. Image Classification Using Chaotic Particle Swarm Optimization. 2006 IEEE International Conference on Image Processsing, 8-11 October 2006, pp. 3001-3004. [4] DAS, S., ABRAHAM, A. and SARKAR, S K. A Hybrid Rough Set – Particle Swarm Algorithm for Image Pixel Classification. Sixth International Conference on Hybrid Intelligent Systems, December 2006. pp. 26-30 [5] DAS, S., ABRAHAM, A and KONAR A. Spatial Information Based Image Segmentation Using a Modified Particle Swarm Optimization Algorithm. Proceedings of the Sixth International Conference on Intelligent Systems Design and Applications, 2006. [6] DUDA, T., M. CANTY and D. KLAUS. 1999. Unsupervised Land-Use Classification of Multispectral Satellite Images. A comparison of Conventional and Fuzzy-Logic Based Clustering Algorithms. [7] EBERHART, R. and J. KENNEDY. 1995. A New Optimizer Using Particle Swarm Theory. Sixth International Symposium on Micro Machine and Human Science, 1995, pp. 39-43. [8] EBERHART, R. and Y. Shi, 1998. Comparison between Genetic Algorithms and Particle Swarm Optimization, Seventh Annual Conference on Evolutionary Programming Proceeding 1998, pp 611-619. [9]

GARNIER, S., G. JACQUES and G. THERAULUZ. 2007. The Biological principles of swarm intelligence. Springer Science + Business Media, LLC 2007. Published online:17 July 2007.

[10] GAREY, M.R. And DS JOHNSON. 1979. Computers and Intractability – a guide to NP-Completeness. San Francisco: W.H. Freeman and Company. [11] KENNEDY, J.and R. EBERHART. 1995. Particle Swarm Optimization. IEEE International Conference on Nueral Networks 1995, pp. 1942-1948. [12 PERREIRA, R.A. 2004. Monitoring Heat Island Phenomenon in Metro Manila: A Remote Sensing- GIS Approach. Unpublished M.Sc. Thesis, College of Engineering, University of the Philippines Diliman, Quezon City. [13] POLI, R. 2008. Analysis of Publications on Particle Swarm Optimization Applications. Journal of Artificial Evolution and Applications. Volume 2008, Issue 1. [14] POLI R., J. Kennedy and T. Blackwell. 2007. Particle swarm optimization: An overview. Springer Science + Business Media, LLC 2007. Published online: 1 August 2007. [15] WREN, A.1996. Scheduling, Timetabling and Rostering – A special relationship?. Practice and Theory of Automated Timetabling, Lecture Notes in Computer Science. ISSN 1611-3349. Volume 1996, pp.46-75 [16] Shu-Chuan C., C. Yi-Tin ; H. Jiun-Huei. 2006. Timetable Scheduling Using Particle Swarm Optimization Innovative Computing, Information and Control, 2006. ICICIC apos;06. First International Conference on Volume 3, Issue , 30-01 Aug. 2006, 324 – 327. [17] XU, X. and Zhang, A. 2006. An unsupervised Particle Swarm Optimization Classifier for SAR Image. 2006 International Conference on Computational Intelligence and Security, 3-6, November 2006, pp.1630-1634. [18] ZHANG, L., H. YU and S. HU. 2004. Optimal Choice of Parameters for Particle Swarm Optimization. 2004. Journal of Zhejiang University Science. ISSN 10093095.