Study On Selecting Optimal Parameters Of Image Segmentation Based Improved Watershed Algorithm For Remote Sensing Imagery With High Spatial-resolution

Since the beginning of21st century, one of remarkable characteristics in the field of satellite remote sensing technology is quickly towards high spatial resolution. The greatest characteristic of high spatial resolution remote sensing imagery is rich detail information of feature space and prominent information of structure and texture. Classical pixel-based image analysis method is mainly based on the spectral characteristics of images, so it is difficult to effectively extract necessary object information from the high spatial resolution images. Object-based image analysis method arises at the historic moment. Spectral features, shape features and texture features of images were comprehensively analyzed, and the smallest unit of the study is no longer a single pixel but one object, and the subsequent analysis is also based on objects.The paper takes Jianye District of Nanjing city and the surrounding area as the study area, and uses IKONOS image as the data source. Firstly several commonly used algorithms of multi-scale image segmentation were discussed. Secondly, two modules selecting optimal image segmentation scale through discrepancy measures and goodness measures were established, then the optimal image segmentation scales obtained by the two modules were compared. Finally, the image was segmented by using the optimal segment scale, then the segmented image was classified and accuracy assessment was obtained.The main conclusions of this paper are as follows:(1) This paper uses the method based on the inconsistency measure proposed by Liu. et al(2012) to select optimal image segmentation scale. Its significant feature is that the geometric and arithmetical difference between reference objects and corresponding segmented objects are both considered. Because under-segmentation will lead to classification error, PSE index is used to better control under-segmentation error; considering over-segmentation, NSR index ensures less over-segmentation. This method can well measure the difference between reference objects and corresponding segmented objects, and it is very effective to select optimal image segmentation scale.(2) This paper proposes a method of goodness measures for selecting optimal image segmentation scale. Its main characteristic is that goodness measures are based on corresponding segmented objects in order to ensure consistency of segmented objects and reference objects. The paper proposes new indexes to measure internal homogeneity in objects and heterogeneity between adjacent objects. The index to measure heterogeneity between adjacent objects uses local Moran’s Ⅰ,and it can reflect spatial relationships between corresponding objects and their adjacent objects in the segmented image well. Compared with the internal homogeneity index, the paper proposes function of SEP to select optimal image segmentation parameter. The experiment has achieved the anticipated effect.(3) The two methods of discrepancy measures and goodness measures for select optimal image segmentation scale have two different angles. Discrepancy measure is based on the angle of discrepancy between segments and actual features, and goodness measure is based on the angle of the formation of segments. Though the angles are different, both can well evaluate image segmentation results, and the optimal segmentation scales are almost consistent.(4) Different features in the image have different scales, so the concerned scale for special aims is distinct. The optimal segmentation scale for one feature is only fit for this one, but not for all others.