Salient Target Shape Adaptive Neighbor for High Resolution RS Imagery Classification
ISBN 978-85-88783-11-9
Authors
1Yan, L.; 2Yufang, D.; 3Guobin, C.
1SPATIAL INFORMATION RESEARCH CENTER,SOUTH CHINA NORMAL UNIVE Email: yanli@scnu.edu.cn
2SCHOOL OF COMPUTER, SOUTH CHINA NORMAL UNIVERSITY Email: 756261051@qq.com
3SPATIAL INFORMATION RESEARCH CENTER, SOUTH CHINA NORMAL UNIV Email: chigb@scnu.edu.cn
Abstract
The wide applications of high spatial resolution remotely sensed images (RSI) are calling for more and more accurately classified imagery. Accordingly, researchers have been trying very hard to improve the effectiveness and efficiency of the high resolution RSI classification. However, feature extraction, as a significant processing step in classification procedures, fails to fully extract the spatial features from high-resolution imagery, and that causes inaccuracy in various applications. Our research group first proposed a novel feature extracting approach based on the shape adaptive neighborhood (SAN), and present a scientific analysis of the approach. In SAN approach, a certain number of pixels in the SAN are of the same type of terrain object as the “central pixel” and the heterogeneity, the judgment will be correct. Nevertheless, as the limitation of the “central pixel” and the heterogeneity, it is not good representation of the texture feature of the objects and improving the accuracy of the classification for high resolution RSI. On the visual attention mechanism (VAM) point of view, SAN is one of the most substantial method of human visual mechanism when processing external visual information but still not enough. Thus, not only the concept of SAN is proposed to model visual perception and is applied to extract spatial features from high-resolution imagery but also VAM is introduced to the SAN approach. Hence this paper emphasized on salient target shaper adaptive neighbor (ST-SAN) for RSI classification method which is a novel method derived from SAN and Itti salient model. At first, salient target computing model based on human visual attention mechanism are analyzed and compared such as the classical Itti salient target model and GBVS (Graphic Based Visual Saliency) model to determine the parameters. And then, we combined the Itti model description method and SAN model, which aims to create the new ST-SAN high resolution RSI classification method and to solve the problem of the object' s texture feature lost in SAN. It improved the heterogeneity expression of the pixels, SAN generation and description methods for high resolution RSI. The steps of ST-SAN approach as follows: 1) Obtaining a series of saliency maps like RG, BY, I channels first; 2) Using the improved Activation filtering model based on different threshold in the current window to generate the SAN (shape adaptive neighborhood); and then, 3) Extracting the shape and texture feature values (texture and shape feature etc) in the corresponding neighborhood; After Image feature extraction, 4) Taking these features as different bands of the image to make classification of the study area by using Support Vector Machine, SVM method and further to verify the efficiency and benefit of ST-SAN feature extraction approach. Finally, experiments on a color composed image of WV(0.5 meter) panchromatic image and WV multispectral image (2 meter) using the proposed approach were conducted, to perform classification for a Land Use / Land Cover (LULC) application. In the experiment, the SAN approach, the classification accuracy in water, building two objects have greater improvement. Respectively improved 12.58% and 6.26%, and the Kappa coefficient is 0.67, the overall classification accuracy of 74.8%.. In contrast, the use of ST-SAN classification approach had a greater increased in the classification results which Can determine the classification results as "very good" and the Kappa coefficient is 0.80, the overall precision is 83.1%. Especially in the classification of water, forest and residential area, its precision is up to 92.5%, 85.2%and 88% . From the human eye observation point of view, the results of ST-SAN approach to show more"clean" feature types than other tow results, a terrain object classification results are less influenced by finely divided smaller interference error. And indicates the promising applications in the auto-interpretation of RSI.