Satellite Images Vectorization Based on Clustering and Interpolation Technique Rehab

Dealing with satellites Images, interpretation, improvement and modification area important issues of researchers concerning, because of the great benefits derived from them. GIS on the other hand is making to build disaster planning, crisis management, and alarm systems effective decision. In this paper a method for digitization gray scale satellite image to a digital map will propose. The proposed method depends on slicing the image into multi-layer, and preprocessing steps such as noise removal and contrast adjustment then use approach K-mean clustering unsupervised Classification . At the last step; the edges of segments (objects) are interpolated for softening the sharp edges using cubic spline interpolation, and these parts will be rendered with desired color that gets from original image. Layers results could be considered as one map for roads, buildings, green zones, etc. more than one layer could be merged in new map and may append information on it. The produced digital map will be saved as SVG vector file format.


Introduction
Each pixel of an image data sensed from a distance exemplifies a spot at a particular place of the planet Earth.These images have a large number of purposes which contain meteorology, cartography & military intelligence.The images of a satellite could be one of the following: vapor of water, infrared & visible light as shown in Figure 1 [1].Image is the easiest way to obtain geographic information.The important point here is whether the data which are obtained from satellite image will give required accuracy in GIS or not.Image digitization is like a conversion of the state of signals from alternating into digital by the way of taking specimen and quantization process.On a space of display, any x,y grid of coordinates can be called raster.The raster file is sometimes referred to as a bitmap because it contains information that is directly mapped to the display grid.A raster file is usually difficult to modify without loss of information.Examples of raster image file types are: BMP, TIFF, GIF, and JPEG files [3].A vector graphic is a set of instructions to draw lines, curves, and lettering.The advantage of vector graphics is that, with a relatively small file size, they can be rendered sharply at any size and resolution.As a more abstract & straightforward representation of a data than raster image is supplied by structure of vector data, a conversion from raster to vector, the automatic one of course, is not simple to be done, or what we name "the process of vectorization", yet the vector-raster conversion is rather common.Tracing of line, recognition of shape, pre-processing, creation of topology, attribute assignment and acquisition of image are all contained in the process of rastervector complete conversion.The basic idea of any vectorization method is the classification process, where if a pixel satisfies a certain set of criteria then that pixel is assigned to the class that corresponds to those criteria.The methods of classification of image may be assorted into two major categories depending on two methods, the pixel based and the object based.The methods of pixel based have the ability to assort single pixels and they have not got to rely on the neighborhood or the pixel's spatial information.The methods of object (Region) based can deal with imagery that has a high resolution, and that could lead the process of classification of most of the methods of pixel based to be irritated [6].Classes are going to be particularized with the ground's common features relying on the information kind taken from the main data.image can be considered as a complicated operation, and sure, more than one factor can be effective on it.There have been many non-parametric classifiers which are considered as significant approaches for the classification of multi-source data such as decision tree classifier, knowledge based classification and fuzzy logic.There is another point which is very significant in the operation of vectorization, that is the mathematical process of interpolation.

Literature Review
There are many vectorization approaches developed in various domains for different purposes.In the following, some proposals have been drawn from the scientific literature:  Saleh, M.& Mohammed, K. & Ahmed, M., in 2012 [17]developed a conversion model of vectorization for the images of quad tree satellite to the format of vector, so that model may be utilized in the projects of GIS.Results showed that the model is acceptable for mapping purposes, and this quality leads it to be a well model utilized to get the quad tree vectorization issue settled.And for accuracy purposes, this model must be checked and enhanced.There are two ways of accuracy to be considered, which are location and size.points are waiting to settle, the first step is considered to be finished and an early age of group is accomplished [13].At this end we would have to add up k new centroids again a center of the clusters produced from the preceding step.After obtaining those k new centroids, a new binding has to be achieved be among the points the same data set and the closest new centroid.And there has been a loop.As a result of that one might take notice that the k centroids alter their places slowly till there are no more alterations.That would mean a state of quietness happen to the centroids.Eventually, the object of the said algorithm is minimizing an objective function, a squared error function in such a situation [14].
Where C is centroid of cluster x is pixels in image and j is Euclidian distance.The the following algorithm is in detail:

Interpolation using Cubic Spline
There are many Interpolations method used to estimate data or points from other collection of them, the estimation values have a relation with the other known values.In this paper the cubic Spline used for interpolated data that will be used in vector graphics which is explained in figure 2 [19].

The Proposed method
The proposed method is consists of stages start with pre-processing for satellite image.Each stage has specific functions; the main stages of the proposed method are explained in

Image Clustering
The k-mean clustering approach is used in this proposal for image clustering.In this stage the number of layers requested or clusters should be specified, and the main steps of clustering process will follow the following steps ad explained in figure 6.
Step1: select N number of clusters (layers) Step 2: convert the pre-processed image to one dimension array V (vector) Step 3: Select random N points as centroid of each clusters Step 5: specify a minimum distance from N distance to be member of clusters Step 6: find the mean of each clusters as new centroids Step 7: apply a previous step for a new centroids until there are no changing on new centroids.
Step 8: convert V to 2D array represent the same dimension of image with class labelClassMatrix

Object Extraction
The object extraction depend on the previous step by applying segmentation of each layer (cluster) alone, the segmentation of layers shown in figure 7, and follow the following steps: Step 1: for each layer construct zero matrix ZM1, ZM2, …, ZMn Step 2: for each layer from 1 to N apply step 3 Step 3: if ClassMatrix equal to layer number then set the layer matrix = color in preprocessed image otherwise set layer matrix to zero

Object border Interpolation
Cubic Spline interpolation applied on each layer extracted from the previous step to smooth the borders of objects.An interpolating polynomial is gained by the said method, and that polynomial will be softer and has less errors.The interpolation applied on each layer separately for later rendering it.Figure 12 shows an example for this step.The rendering (filling) with color is a last step in the proposed method that recognize each layer, see figure 13, the selection of color depend on calculating the histogram of each layer in original image and take the maximum frequency in red color, in this a maximum green is found and a maximum blue is also found by the following procedure: Step 1: for each object (layer) obtained from previous apply Step 2: depend on objects segment before split three band of original image Step 3: find the histogram of red band of object Step 4: find the maximum value in histogram that not equal to zero Step 5: by fixing the maximum red find maximum green that not equal to zero Step 6: by fixing the maximum red value with maximum green find maximum blue that not equal to zero Step 7: render the object by the obtained color

Implementation and Testing Results
The Proposed method have been applied on several satellite images, the number of layer selected is 5 layers.The total test images and layers shown in figure 14.The number of layers are variable and selected by the user depend on the approach used it for.

Figure1
Figure1: Satellite Image An individual pixel can symbolize class covers' mix, within class variableness or other patterns of complicated surface cover in remote sensing imagery, which one class cannot depict appropriately.Utilizing this image the demanded specifications are taken from.The varied specifications out of this like used land, forest and unused land and vegetation indices are particularized.The algorithms of classification, right after taken the specifications out of the image, can be applied to obtain varied groups of classification.The classification of an Vol: 13 No:3 , July 2017 DOI : http://dx.doi.org/10.24237/djps.


Sumner, R. & Noris, G ,A Hornung and Gross, M., in 2013[18] proposed -for the purpose of the drawings of clean line which get the topology of drawings analyzed to get by the junction obscurityan algorithm of vectorization.We have two mechanisms contained in the approach, a mechanism assists obscure hard cases, which is called clustering of gradientbased pixel, and another uses the topology of drawing to make educated picks as treating with junctions in order to enhance the existing mechanisms result which supply the topology of drawing, and this one called procedure of reverse drawing. Suleyman, E. & Ahmet, S., in 2015 [11] createdfor transferring an image of a Landsat satellite from some machine to another one throughout web servicesa scenario.It is demonstrated by experimental results that the reduction of the size of file belongs to data of an original raster map is feasible, to achieve the drastically transmitting and to the quality of scalability to be achieved on the application.Anyway, the hardness of algorithms causes time consumption for the operation of simplifying.Vol: 13 No:3 , July 2017 DOI : http://dx.doi.org/10.24237/djps.1303.284BP-ISSN: 2222-8373 E-ISSN: 2518-9255  B. Yuan, S.C. Liew, L.K. Kwoh, in 2010 [21] have used Line-Segment Vectorization on High-Resolution Satellite Images: This paper presents a procedure to extract and vectorize the mentioned line-segments.Scan part utilizes the continuing of edges and correlating of intensity to get line-segments recognized domestically.The incomes evolved by experiment utilizing images of satellite have high-resolution maintain that the suggested procedure of vectorizing line-segment can be both influential and competent for the applications in reality like manual/automatic establishing top extracting from the images of stereo Image classification is a complex process that may be affected by many factors.Effective use of multiple features of remotely sensed data and the selection of a suitable classification method are especially significant for improving classification accuracy.Nonparametric classifiers such as fuzzy logic, neural network, decision tree classifier, and knowledge-based classification have increasingly become important approaches for multisource data classification.More research, however, is needed to identify and reduce uncertainties in the image-processing chain to improve classification accuracy In general, image classification approaches can be grouped as supervised and unsupervised, or parametric and nonparametric, or hard and soft (fuzzy) classification, or per-pixel, subpixel, and per-field.Based on whether training samples used or not used image classification can be of two types (Supervised Classification) and (Unsupervised Classification),in this method use the K-Means Clustering was chosen to classify each pixel in the image input as well as to label each classification.In this way, the layers of the workbook were created according to the work of this algorithm, which depends mainly on the value centroids  The algorithm of K-Means Clustering This algorithm can be considered as one of the algorithms of unsupervised learning to settle the problem of well-established clustering or classification.It obeys a way quite simple and modest to get a given data set classified throughout a given number of clusters [12].Defining k number of centroids, each one should be for a cluster.And those centroids Vol: 13 No:3 , July 2017 DOI : http://dx.doi.org/10.24237/djps.1303.284BP-ISSN: 2222-8373 E-ISSN: 2518-9255 better sited in a cunning way due to the fact that says different location causes different result.For that, arranging them very far from one another would be the ideal solution.Then we ought to take each point of a given data set and link it up to the closest centroids.As all

4 :
N number of clusters (layers) 3: convert the pre-processed image to one dimension array V (vector) Select random N points as centroid of each clusters 5: for each point in V find Euclidean distance from each centroid depend on objective function ( ) 6: specify a minimum distance from N distance to be member of clusters 7: find mean of each clusters as new centroids 8: Apply a previous step for a new centroids until there are no changing on new centroids.V to 2D array represent the same dimension of image with layer label layer Matrix 10: End Algorithm  Algorithm of vectorization Vectorization,( Raster to Vector) or simply written R2V, is a process for converting raster image to a vector image.The history of vectorization process begins from scanned-line art black-white images and continues to evolve to cartoon drawings and finally realistic images.Various approaches have been proposed to solve the same problem.The most of these works focus on line drawings, fonts, texts and sketches which are in fact binary images with sparse set of pixel in the region of interest On the other hand, the Vectorization (raster-to-vector conversion) consists of analyzing a raster image to convert its pixels representation into a vector representation The basic assumption is that such a vector representation is more suitable for further interpretation of the image [20].

Figure3.
Figure3.Vectorization of Satellite images is the goal of this work based on interpolation methods.The basic method can have a proper definition by some relationships of mathematics that are among paths that connects points for image description & the points themselves.Image clustering, object extraction, border interpolation and object filling is the last stage in proposed method.

Figure 5
Figure 5 explains the equalization of gray scale image.

Step 4 :
for each point in V find Euclidean distance from each centroid depend on objective function ( )

Step 4 :VolFigure 8 Figure 8
Figure 7: splitting object Figure 8 shows an example of multi-layer extraction for a single satellite image: