Goodness Criterion For Selecting Optimal Combination Of Parameter Values In Remote Sensing Image Segmentation

In object-based image analysis (OBIA),the size, shape of segments is controlled by combination of parameters. Combination of parameters selected has a decisive effect on the accuracy of image classification. So combination of parameters selected is a very critical procedure. In automatic selection of parameters combination, there is a kind of methods called goodness methods. With goodness methods, the optimal parameter combination can be selected objectively and automatically. In this research, goodness methods for selecting optimal combination of parameter values in image segmentation are discussed. The main conclusions are as follows:(1) The goodness methods can be divided into Curve Methods and Function Methods based on their characteristics. Curve methods’ processing is simple and higher efficiency. But the optimal combination of parameter values can not be selected by curve methods. With curve methods can only select the optimal scale parameter. And with lower degree of automation, curve methods require human visually interpretate curve. Function methods’ processing is complex but the optimal combination of parameter values can be selected by function methods. Besides that, the degree of automation of function methods is higher.(2) When to estimate the bottom-up segmentation process (multiresolution segmentation) which is widely used, the intra-region homogeneity should not be evaluated by variance or standard deviation. Because multiresolution segmentation define standard deviation as spectral heterogeneity. Multiresolution segmentation judge whether a pair of adjacent image objects should be merged based on this spectral heterogeneity. If evaluate intra-region homogeneity by variance or standard deviation to select optimal parameter combination, weight for shape parameter can not be selected correctly. Because according to the principle of the segmentation algorithm, a small weight for shape parameter must generate small standard deviation within objects. Therefore, the weights for shape parameters which are selected by variance or standard deviation are generally small and incorrect.(3) GLCM Entropy and NSR are defined as intra-region homogeneity metric and inter-region heterogeneity metric. It is effective to use both of the metrics to select optimal scale parameter. But intra-region homogeneity metric and inter-region heterogeneity metric can not be used to select weights for both the shape and compactness parameters. For function methods can select the correct shape and compactness parameters, shape metric SI need to be considered. Shape metric is used to evaluate whether objects’ shape is correct. With shape metric, function method can get the optimal parameter combination.(4) Reference datasets should be used in order to select optimal combination of parameter values for different land-cover categories. From each original segment dataset, a corresponding segment dataset can be extracted by reference dataset. Function methods are only used to evaluate corresponding segments and then to select optimal combination of parameter values for different land-cover categories becomes feasible.(5) This research proposes a goodness function (GV) to evaluate combinations of parameter values used in image segmentation. The classification accuracy show that the optimal combination of parameter values selected by GV achieves better quality of image segmentation than the optimal combination of parameter values selected by F(v,I). GV can improve accuracies than F(v,I), in terms of both producer’s and user’s accuracies.