Image Pre-processing Land Use Classification

37 the pixel values of the fill scene which can be found using the mean and standard deviation of the data. For greater precision and a product which is visibly better looking, the corrective gains and biases can be calculated in a moving window around each pixel in the scene. This is the basis of the localized linear histogram match LLHM Scaramuzza, et al. 2004. In this research, only LANDSAT images 2005 and 2008 have been processed by SLC-OFF Gap Filling in order to fix the SLC-OFF gaps which appear on the images. Figure 11. SLC-OFF Gap Filling. A SLC-OFF Image, B Overlaid Image: SLC-OFF image and the fill scene image, and C SLC-OFF Gap Filled Image Geometric correction aims to correct positional and geometrical errors that may be occurred on the remotely sensed data. Geometric correction uses one reference data that is known has fine positional and geometrical properties regarding to the real world coordinate system. Base Maps of Riau Province acted as reference data in the geometric correction process which means the LANDSAT images and other related spatial data should refer to the Base Maps of Riau Province as spatial references. The geometric correction process used Polynomial Geometric Model which uses polynomial coefficients to map between image spaces. The order of the polynomial may be from one up to five with no enforced A B C 38 upper limit Leica Geosystem 2005. LANDSAT image has standard pixel resolution 30 x 30 meters, and in consequence the Root Means Squared RMS Error for this research was set to below 5 meters that might be suitable to accommodate various spatial data which involved in the whole research. Figure 12. Geometric Correction. A Reference Data, B Image to be Geo- correction, and C Geo-corrected Image Based on the LANDSAT 7 Worldwide Reference System WRS, Siak District is covered by three scenes of LANDSAT image that are pathrow 126059, 126060 and 127059. The SLC-OFF Gap Filling and geometric correction processes were conducted to the LANDSAT images that are used in this research. After two main image pre-processing have been conducted, there were still two processes, which should be done, in order to prepare the LANDSAT images ready to be used in Land Use Classification. The processes were Image Mosaicking and Image Clipping. The Image Mosaicking aims to join the LANDSAT images which have a same acquisition time 2002, 2005, and 2008. Each acquisition time consists of three scenes of LANDSAT images 126059, 126060 and 127059. The Image Mosaicking was conducted before A B C 39 the Land Use Classification in order to make the process of Land Use Classification became more effective and efficient in term of the working time and the processes itself. In order to increase the effectiveness and efficiency of the Land Use Classification, the mosaicked images from each year were clipped by the Siak District boundary which has been expanded with 5 km buffer along the boundary. This expansion has been done in order to avoid no data value which might be caused by the position displacement. Figure 13. LANDSAT 7 ETM+ of Siak District have been processed by SLF- OFF Gap Filling, Geometric correction, Image Mosaicking, and Image Clipping 40

4.1.2 Image Processing

Image processing has been done in order to derive the information of land use categories from LANDSAT images to produce time-series data of land use of Siak District. There are three main processes were conducted in image processing that are Training Data Generation, Supervised Classification, and Post- Classification. Common method to get training data for supervised classification is by creating signaturesampling data using polygon tool that is usually available in various remote sensing software, and uses land use history data as reference data that were collected from field data collection andor land use history maps that have been collected from other institution. However, in this research the training data generation has been done in different technique. The training data generation has been done by using image segmentation technique, that segments an image into regions of pixels and grouping neighboring pixels which have similar feature values brightness, texture, color, etc.. At this step, the training data generation using image segmentation technique acted as like unsupervised classification method, but here the region of pixels which have been resulted, then could be selected as training data for supervised classification. This technique is very comfortable to generate training data which have similar feature properties in short time, and it can minimize the subjectivity which may usually involve in common method when delineating the polygon of training data. The next task to do was taking regions of pixels selections which were related to the training data of land use categories. At this phase, the land use history data and GPS data that were taken from field data collection activity acted as guidance in selecting regions of pixels as training data for supervised classification. The land use categories that would be interpreted in this research were 6 land use categories that are Forest land, Cropland, Grassland, Wetlands, Settlements, and Other lands see Appendix 1. The selections of training data have been conducted for each land use type with minimum number of sample for each type was 30 samples, and each sample carries at least 3 x 3 pixels in order to accommodate the horizontal positional accuracy of GPS point data which is between 5 – 10 meters. 41 The supervised classification was done by using K-nearest neighbor which basically considers the Euclidean distance in n-dimensional space of the target to the elements in the training data. The K-nearest neighbor is simple and intelligible which classifies the image by examining the root of square differences between coordinates of a pair of objects. In term of supervised classification which was done in this research, the spectral properties values of the training data were calculated, such as minimum, maximum, average, and standard deviation from each band, then its properties were used to classify the object using K-nearest neighbor into a known feature type. Figure below illustrate the processes which have been explained above. Figure 14. Land Use Classification Process. A LANDSAT image, B Result of Training Data Generation, C Training Data Selection, and D Result of Supervised Classification The basic difference between the traditional digital image interpretation and the land use classification by using the combination of image segmentation and K-nearest neighbor in this research is the traditional image interpretation working on pixel-based classification, whereas land use classification applied in this research working on region-based classification. The pixel-based classification emphasizes on calculation of the spectral properties from each pixel A B C D