Improving the Detection of Object-Oriented Changes in High-Resolution Images based on Random Forest Method in Optimal Features Space

Extended Abstract
Land use/cover (LULC) change detection is one of the most important applications in the remote sensing field, providing insights that inform management, policy, and science. In the recent decade, development of remote sensing systems and accessibility to high spatial resolution images has associated with the improvement of digital image processing. The advantage of high spatial resolution remote sensing imagery further supports opportunities to apply change detection with object-based image analysis, i.e. object-based change detection – OBCD.
OBCD analysis in comparison with pixel-based techniques provides a more effective way, especially in high spatial resolution imagery to incorporate spatial, spectral, textural and geometry feature that can identify the LULC change in comparison with pixel-based technique. OBCD approach is classified into for categories: (i) image-object, (ii) class-object, (iii) multi- temporal object, and (iv) hybrid change detection. Different algorithms and features can be employed in the process of image classification for OBCD. Therefore, the choice of algorithm and optimization features are major challenges in OBCD. This paper has introduced an object- based change detection method based on the machine learning algorithm, which can overcome the traditional change detection method limitation and find the interested changed objects. In this paper, multi-temporal object approach is utilized and high spatial resolution imagery, GeoEye-1 and Quick Bird-1 satellite images were acquired during 2002 and 2015, covering a region of the Geshm Island which were used to detect the meaningful detailed change in the study area. As an essential preprocessing for change detection, multi-temporal image registration with the accuracy of less than one second of a pixel is applied. Also, radiometric correction is performed using histogram matching algorithm in ENVI Software. In the Next step, a number of texture features of images such as mean, variance, entropy, homogeneity, momentum and such are extracted from two images. To reduce the input features space, PCA algorithm is employed and the result of this process is used in the segmentation process. The two images are incorporated with PCA output and are used as inputs feature to segmentation. Segmentation is the first step in OBCD. It divides the image into larger numbers of small image objects by grouping pixels. The segmentation algorithm is a region-merging technique. It begins by considering each pixel as a separate object. Subsequently, adjacent pairs of image objects are merged to form bigger segments. The merging decision is based on local homogeneity criterion, describing the similarity between adjacent image objects. Correct image segmentation is a prerequisite to successful image classification. At the same time, this task requires explicit knowledge representation. Furthermore, optimal segmentation results are depended on not only the choice of segmentation algorithm or procedure, but are also often influenced by the choice of user-defined parameter combinations which are required inputs for many segmentation programs. The segmentation has been done using multi resolution segmentation algorithm which involves knowledge-free extraction of image objects. Multi-resolution segmentation begins with single pixel objects and employs a region-growing algorithm to merge pixels into larger objects; pixels are merged based on whether they meet user-defined homogeneity criteria. Each multi-resolution segmentation task must be parameterized by the user and involves settings of three parameters: Scale, Color-versus-Shape, and Compactness-versus-Smoothness. In this paper the process of segmentation is performed in four different levels using Ecognition software and finally, the level with better output with scale of 100 is selected to provide the change map. The scale values were determined through an iterative method. The color/shape was set to 0.6/0.4 and compactness/sharpness was set to 0.5/0.5 for the selected level. Color and shape weightage are inter-connected to each other. If color has a high value, which means it has a high influence on segmentation; Shape must have a low value with less influence. If both parameters are equal, then each will have roughly equal amount of influence on segmentation outcome. In addition, texture, spatial and geometrical features from the segmented image are extracted. Feature space Optimization (FSO) tool available in Ecognition software have been used to calculate optimum feature combination based on class samples in four classes including: ”barren to road”, ”barren to building”, barren to vegetation” and “barren with no change. It evaluates the Euclidean distance in feature space between the samples of all classes and selects a feature combination resulting in best class separation distance. In this study, the performance of the proposed RF-based OBCD method is compared with the conventional methods such as support vector machine (SVM) and KNN. The commonly used accuracy assessment elements include overall accuracy, producer’s accuracy, user’s accuracy and the Kappa coefficient. The overall accuracy of the change map produced by the RF method was 86.57%, with Kappa statistic of 0.79, whereas the overall accuracy and Kappa coefficient of that by the SVM and NN methods were 83.76%, 0.75 and 75%, 0.63, respectively. Experimental results show that overall accuracy and kappa coefficient obtained from the proposed RF-based OBCD method improve 3% and 18%, 2% and 10% respectively compared with SVM and KNN improved. The results indicated that object base change detection method can be performed more accurately and reliably in the high-density region if it uses image with high spatial resolution. Also, selection of classification algorithm has very impressive effect on the providing change map.