Ganesh Kumar T, IJRIT
64IJRIT International Journal of Research in Information Technology, Volume 1, Issue 7, July, 2013, Pg. 64-73
International Journal of Research in Information Technology (IJRIT)
www.ijrit.com
ISSN 2001-5569Supervised Methods for Land Use Classification
1
K Rajalakshmi,
2D Murugan,
3Manish T I,
4Ganesh Kumar T,
5Divya C S
1Assistant Professor, Center for Information Technology & Engineering, Manonmaniam Sundaranar University, Tamilnadu, India
2 Associate Professor, Department of Computer Science & Engineering, Manonmaniam Sundaranar University, Tamilnadu, India
3Assistant Professor, Department of Computer Science & Engineering, MET‟S School of Engineering, Kerala, India
4 Research Scholar, Department of Computer Science & Engineering, Manonmaniam Sundaranar University, Tamilnadu, India
5 PG Student, Center for Information Technology & Engineering, Manonmaniam Sundaranar University, Tamilnadu, India
1 [email protected] , 2[email protected], 3 [email protected],
4 [email protected], 5 [email protected]
Abstract
Remote sensing (RS) technologies were utilized to collect, analyze and extract information from satellite images for the land use classification of Tirunelveli district. This paper concerns the evaluation of the four supervised techniques such as MKNN (Modified K-Nearest Neighbour), Maximum likelihood, Minimum Distance to Mean, Parallel Piped for the classification purpose of the LANDSAT imageries. The images consist of the three visible bands-red, green and infra-red. The training sites were established and the relative performance of the techniques was evaluated. The accuracy of the classified images was validated using a reference data set. The results produced high degree of accuracy.
Keywords : Remote sensing, Supervised Classification, Maximum Likelihood, Minimum Distance to Mean, Parallel piped, Land Use and Land Cover.
1. I
ntroductionRemote Sensing (RS) associate to the science of identifying different earth surface and features assessment using electromagnetic radiation. Remote Sensing technologies can be used to get spatially variable data for a number
Ganesh Kumar T, IJRIT
65 of applications. A number of these technologies can supply data to solve problems, and can often be accomplished at a lower comparative cost than many other established methods. It has also been used to supervise land use changes;this has a significant role in urban advance and the purpose of water quality parameters. Remote sensing is very useful for the construction of land use and land cover information which can be useful to resolve the allotment of land uses in the watershed. The evolution in technology of remote sensing has caused it to become one of the most commonly used techniques in the world. Remote sensing is the acquisition of information about an object or phenomenon, without making physical contact with the object. Land cover mapping is important in many scientific and commercial applications involving resource monitoring. Remote sensing provides an effective tool for analysis of land-use and land-cover changes at a regional level. Land use classification of the LANDSAT imagery helps to identify and analyze the different regions like ocean, hilly regions, human settlements, vegetation etc. in the Tirunelveli district. This paper evaluates four supervised classification methods for automatically obtaining LULC in the Tirunelveli district from LANDSAT TM images.
Remote sensing imagery has many applications in mapping land use and land cover, crop growing, soils mapping, forestry, city development, archaeological investigations, armed inspection, and geomorphologic surveying, among other uses. In this study two different classification methods were used: Unsupervised and supervised classification. Unsupervised classification is the classification of natural groups, or structures, within multispectral data. Supervised classification is the process of using training samples, samples of known identity to classify pixels of unknown identity. A training area is a small example of standardized areas selected by the image analyst prior to arrangement. Each area is determined from maps, ground data, or other information (e.g. land use database). Training sites should be free of anomalies and be large enough to provide good statistical representation.
Training areas should avoid edge pixels containing the combined backscatter of multiple targets (mixed pixels), and inconsistencies within the area such as roadways, power lines, intermittent cover, etc. Once clear training areas are used to produce signature figures for each definite class.
2. Methodology
The following methodologies are study area and data used for feature extraction.
2.1. Data and Study Area
LANDSAT 7 was intended to last for five years, and has the facility to collect and send out up to 532 images per day. It is in a polar, sun-synchronous orbit, meaning it scans across the entire earth's face. With an elevation of 705 kilometers +/- 5 kilometers, it takes 232 orbits, or 16 days, to do so.
2.2. Feature Extraction
Raw intensities are usually not sufficient for successful land use classification. Therefore, for each band, and for each pixel, we evaluated a number of textural features on a sliding window centered on each image pixel.
Texture provides valuable information such as contrast, uniformity, regularity etc. for the identification of objects in the image. A key quality of texture is the replication of a model or patterns over a region. The pattern may be recurring exactly, or as a set of small variations, possibly as a task of position. Texture can also be defined as a set of local statistics or other local properties of an image that are constant, slowly varying or approximately periodic.
The classification of specific textures in an image is achieved primarily by modeling texture as a two- dimensional gray level variation. This two dimensional collection is called as Gray Level Co-occurrence Matrix (GLCM). The elements of this symmetrical matrix, p(i,j), represent the relative frequency by which two pixels with grey levels "i" and "j", that are at a distance “d” in a given direction, are in the image or neighborhood. Using this matrix, several statistical features representing texture properties, like contrast, moment, angular moment, variance and entropy are used. The GLCM is computed in four directions for 0, 45, 90 and 135. Based on the GLCM four statistical parameters energy, entropy and connection are computed. Finally a characteristic vector is computed using the means of all the parameters.
Ganesh Kumar T, IJRIT
66 The steps for extracting texture features of image using GLCM can be given as below.• Merge the R, G and IR bands of the image.
• Convert thus formed image into gray level image.
• Compute GLCM matrix.
• Compute the statistical features Probability, Average, Moment, Entropy, Variance where P (i,j) is probability density that the first pixel has intensity value i and the second j, which is separated by distance d.
3. Classification
The conventional methods of organization mainly follow two approaches: unsupervised and supervised.
3.1 Unsupervised Classification
Unsupervised classifiers do not operate training data as the basis for organization. Rather, this family of classifiers involves algorithms that observe the strange pixels in an image and cumulative them into a number of classes based on the natural groupings or clusters present in the image values. It performs very well in cases where the values within a given cover type are close together in the quantity space, data in different module are comparatively well unconnected.
3.2 Supervised Classification
Supervised categorization can be defined normally as the process of samples of known distinctiveness to classify pixels of unknown identity. Samples of known distinctiveness are those pixels located within training areas.
Pixels positioned within these areas term the training samples used to guide the categorization algorithm to assigning specific spectral values to appropriate informational class.
3.3 MKNN: Modified k-Nearest Neighbor
The main idea of the offered method is transmission the class label of the data according to K validated data points of the train set. In other hand, first, the validity of all data samples in the train set is computed. Then, a weighted KNN is performed on any test samples. In the MKNN algorithm, every sample in train set must be validated at the first step. The validity of each point is computed according to its neighbors. The validation process is performed for all train samples once. After assigning the validity of each train sample, it is used as more information about the points.
Ganesh Kumar T, IJRIT
67 3.4 Maximum likelihoodA statistical conclusion rule that examines the likelihood function of a pixel for each of the classes, and assigns the pixel to the class with the highest probability. The classifier assumes that the training statistics for each class have a normal or 'Gaussian' distribution. However many are not, radar statistics in particular. Training statistics with bi- or tri-modal histograms are not suitable as they specify non-homogeneity within classes and are non- Gaussian. The classifier then uses the training Statistics to compute a probability value of whether it belongs to a particular land cover category class. This allows for within-class spectral variance. The image analyst can use a- prior knowledge to weight the probability Function. MLC usually provides the highest classification accuracies.
Accordingly, it has a high computational requirement because of the large number of calculations needed to classify each pixel. For each feature vector, the distances towards class means are calculated. This includes the calculation of the variance-covariance matrix V for each class i. The formula used in Maximum Likelihood reads:
Where
di distance between feature vector (x) and a
class mean (m
i) based on probabilities
Vi the n x n variance-covariance matrix of class
i, where n is the number of input bands
|Vi| determinant of Vi
Vi-1 the inverse of Vi
Y x - m
i; is the difference vector between
feature vector x and class mean vector m
iyT the transposed of y
For each feature vector x, the shortest distance di to a class mean mi is found if this shortest distance to a class mean is smaller than the user-defined threshold, then this class name is assigned to the output pixel else the undefined value is assigned.
3.5 Minimum Distance to Mean
It determines each pixel's 'distance' from the class means, and assigns them to the neighboring class. If the pixel is further from analyst-defined distance from any group, it remains unclassified or 'unknown'. It does not estimate differing degrees of variance within the class, therefore it has a lower overall correctness than the Gaussian Maximum Likelihood classifier. This classifier is the fastest of the usually used algorithms because it is mathematically simple. For each feature vector, the distances towards class means are calculated.
• The shortest Euclidian distance to a class mean is found.
• If this shortest distance to a class mean is smaller than the user-defined threshold, then this class name is assigned to the output pixel.
• Else the undefined value is assigned.
The minimum Euclidian distance classifier is defined by the following equation:
For J= 1 to d
Ganesh Kumar T, IJRIT
68 Gi(X) is the result for class i on pixel XT indicates transposition of the elements in brackets d is the number of channels in the classification X=(x1... xd) is the (d by 1) pixel vector of grey-levels Ui= (u1,...,ud) is the (d by 1) mean vector for class i j is the subscript of jth element of a vector
SUM [] is the total of elements inside brackets
If for all i not equal j, Gj(X) < Gi(X),then X is classified as j.
3.6 Parallel Piped
The parallel piped classifier uses the threshold of each class signature to resolve if a given pixel falls within the class or not. The threshold specifies the magnitude (in standard deviation units) of each side of a parallelepiped surrounding the mean of the class in feature space. If the pixel falls inside the parallelepiped, it is assigned to the class. However, if the pixel falls within more than one class, it is put in the overlap class (code 255). If the pixel does not fall inside any class, it is assigned to the null class (code 0). The parallelepiped classifier is typically used when speed is required. Unfortunately, in many cases this results in poor accuracy and a large number of pixels classified as ties (or overlap, class 255).
• The steps for parallel piped classifier are follows:
• For each training region determine the range of values observed in each band.
• These ranges from a spectral box (or parallelepiped) which is used to classify this class type.
• Assign new image pixels to the parallelepiped which it fits into best.
• Pixels outside all boxes can be unclassified or assigned to the closest one.
• Problems with classes that exhibit high correlation between bands. This creates long „diagonal‟ data-sets that don‟t fit well into a box.
Advantages of parallel piped classifier are they are mathematically simple, they are computationally efficient and they are sensitive to different degrees of variance in the data. The disadvantages of parallel piped classifier are problems occur in regions of overlap and they does not account for inter-band covariance.
3.7 Euclidean Distance
Euclidean distance simply is the geometric distance in the multidimensional space. This is almost certainly the most usually chosen type of distance. The Euclidean distance between points p and q is the length of the line segment connecting them ().In Cartesian coordinates, if p = (p1, p2,..., pn) and q = (q1, q2,..., qn) are two points in Euclidean n-space, then the distance from p to q, or from q to p is given by
The location of a point in a Euclidean n-space is a Euclidean vector. So, p and q are Euclidean vectors, initial from the origin of the space, and their tips indicate two points. The Euclidean norm, orEuclidean length, or extent of a vector measures the length of the vector The location of a point in a Euclidean n-space is a Euclidean vector. So, p and q are Euclidean vectors, initial from the origin of the space, and their tips indicate two points.
Ganesh Kumar T, IJRIT
69 The Euclidean norm, or Euclidean length, or extent of a vector measures the length of the vectorWhere the last equation involves the dot product
.
4. Results and Analysis
4.1 Performance matrix
Following shows the table of classification rate for the all specified methods.
Supervised methods Classification rate
MKNN 92.5612
Maximum likelihood 91.3245 Minimum Distance to
Mean
88.4892 Parallel Piped 82.6261
Following shows the table of time taken value for the all specified methods.
Supervised methods Time taken value
MKNN 2.8123
Maximum likelihood 2.9356 Minimum Distance to
Mean
3.6483
Parallel Piped 3.9276
Ganesh Kumar T, IJRIT
70 Fig 1: shows the graphical representation of the classification rate.Ganesh Kumar T, IJRIT
71 Fig 2: shows the graphical representation of the time taken value.4.2 Experimental images
In this paper, three bands of LANDSAT data were utilized namely R, G and IR. These bands were covering more than Tirunelveli district. Thus these were cropped in order to concentrate on the specific district. After that, these images were merged to form a true color image which is shown below.
The following images show the outputs produced by the four supervised methods.
Ganesh Kumar T, IJRIT
72(e) (f) (g) (h)
(e) MKNN Classification Image (f) Maximum Likelihood Classification Image (g) Minimum Distance to Mean Classification Image (h) Parallel Piped Classification Image
5. Conclusion & Future Enhancement
In this paper, four supervised methods were used for the land use classification of Tirunelveli district. The methods used were MKNN (Modified K-Nearest Neighbor), Maximum likelihood, Minimum Distance to Mean, Parallel Piped. The remote sensing data were obtained from LANDSAT. First of all, feature extraction was performed for reducing the dimensionality of the input images. Then classification methods were applied to these images by utilizing Euclidean distance. Accuracy assessment methods like classification rate and time taken value were applied and it was found that out of the four supervised methods used, MKNN is best for the classification of the LANDSAT imageries. For future research, data from other remote sensing satellites such as IKONOS, SPOT can be used. Also, other supervised methods such as SVM (Support Vector Machine), SOA(Service Oriented architecture)methods can be utilized. Moreover, instead of Euclidean distance, other measures such as log likelihood, chi-square can be used.
6. References
[1] Anderson, J.R., Hardy, E.E., Roach, J.T. and Witmer, R.E. (1976), A land use and land cover classification system for use with remote sensor data, U.S. Geological Survey Professional Paper, No. 964, USGS, Washington, D.C.
[2] Bazan, J. and H.S. Nguyen, S.H. Nguyen, P. Synak, J. Wroblweski (2000): Rough Set Algorithms in Classification Problem. In L. Polkowski, S. Tsumoto, and T. Lin editors, Rough Set Methods and Applications, Physica-Verlag, Heidelberg New York, p 49-88.
[3] Bazan, J.and M. Szczuka (2000): RSES and RSESlib - A collection of tools for Rough Set Computations.
Lecture Notes in Artificial Intelligence 3066, Berlin, Heidelberg, Springer, p 592-601
[4] Bazan, J. and M. Szczuka (2005): The Rough Set Exploration System. In J. Peter, A. Skowron editors, Transactions on Rough Sets III, Springer, p37-55.
[5] Ehlers, M., Gahler, M. and Janowsky, R., (2003), Automated analysis of ultra high resolution remote sensing data for biotope type mapping: new possibilities and challenges, ISPRS Journal of Photogrammetry & Remote Sensing, 57, 315– 326.
[6] Engman, E. T. and Gurney, R. J. (1991) Remote Sensing in Hydrology, Chapman and Hall, London, 225 p.
[7] ERDAS IMGINE (2002) ERDAS Field Guide, 6th Edition, 686 p. ESRI (2001) ArcGIS version 8, 8 volumes.
[8] Foody, G. M.. (2002), Status of land cover classification accuracy assessment”. Remote Sensing and Environment, 80, 185-201.
Ganesh Kumar T, IJRIT
73 [9] Foody, G.M., Ghoneim, E.M. and Arnell, N.W. (2004) Prediction locations sensitive to flash flooding in an arid environment, Journal of Hydrology, No. 292, pp. 48-58.[10] Franklin, J. T.L. Logan, C.E. Woodcock, and A.H. Strahler. 1986. Coniferous forest classification and inventory using LANDSAT and digital terrain data. IEEE Transactions on Geoscience and Remote Sensing. 26:139- 149.
[11] Gross, M.F., V. Klemas, and J.E. Levasseur. 1988. Remote sensing of biomass of salt marsh vegetation in France. International Journal of Remote Sensing. 9:397-408.
[12] Guerschman, J. P., Paruelo, J. M., Di Bella, C., Giallorenzi, M. C. and Pacin, F., 2003, Land cover classification in the Argentine Pampas using multi-temporal Landsat TM data, International Journal of Remote Sensing, 24(17), 3381–3402.
[13] Haack, B. and Bechdol, M., (2000), Integrating multisensor data and RADAR texture measures for land cover mapping, Computers & Geosciences, 26, 411 – 421.
[14] Hutchinson, C.F. 1982. Techniques for combining LANDSAT and ancillary data for digital classification improvement. Photogrammetric Engineering and Remote Sensing. 48:123- 130.
[15] Johnson, G.R. and W.G. Rohde. 1981. LANDSAT digital analysis techniques required for wild land resource classification. Proc. Arid Land Resource Inventories: Developing Cost- Effective Methods, LaPaz, Mexico, p.204- 213.
[16] Kim, D., “Data classification based on tolerant rough set,” Pattern Recognition, 34, pp. 1613-1624, 2001 [17] Kuncheva L. I., Combining Pattern Classifiers, Methods and Algorithms, New York: Wiley, 2005.
[18] Parvin H., H. Alizadeh, B. Minaei-Bidgoli and M. Analoui, “An Scalable Method for Improving the Performance of Classifiers in MulticlassApplications by Pairwise Classifiers and GA”, In Proc. of the Int. Conf. on Networked Computing and advanced Information Management by IEEE Computer Society, NCM 2008, Korea, Sep.
2008.
[19] Parvin H., H. Alizadeh, B. Minaei-Bidgoli and M. Analoui, “CCHR: Combination of Classifiers using Heuristic Retraining”, In Proc. of the Int. Conf. on Networked Computing and advanced Information Management by IEEE Computer Society, NCM 2008, Korea, Sep. 2008.
[20] Parvin H., H. Alizadeh, M. Moshki, B. Minaei-Bidgoli and N. Mozayani, “Divide & Conquer Classification and Optimization by Genetic Algorithm”, In Proc. of the Int. Conference. on Convergence and hybrid Information Technology by IEEE Computer Society, ICCIT08, Nov. 11-13,2008,Busan,Korea.
