Application Of Spatial Information In Object-based Classification

Remote sensed imagery with very high resolution (VHR) contains substantial spatial information. Much valuable spatial information such as location, shape, and texture can be visually interpreted clearly in the VHR imagery. One a specific entity in median or low resolution imagery is usually characterized by pixels with heterogeneous spectral reflectance, which lead to high heterogeneous degree within a real entity. The increase of heterogeneity calls for new classification method geared towards VHR imagery. Traditional pixel-based classification approach can not be applicable for VHR imagery any more, because assigning individual pixels using pixel-based approach depends solely on the spectral information which may result in the significant "salt and pepper effect" due to the heterogeneous spectral reflectance in a specific entity. Object-based classification approach provides another means of integrating spatial information for classification, which is suitable for VHR imagery processing.Different vegetation categories have very similar spectral reflectance. Therefore, additional spatial information has to be taken into consideration in the classification process, in order to improve the differentiation of various vegetation types on the imagery. This research aims to detect the incorporation of spatial information, in particular texture and landscape pattern, in the object-based classification, via a case study on delineating Torreya using IKONOS imagery.In this research, the method of lacunarity texture characterizing the spatial pattern in the vegetation area was proposed; object-based approach and functions of GIS spatial analysis were combined to carry out the scale transformation of the landscape level metrics, by taking advantages of multi-scale segmentation technology and GIS further development; A landscape metric which could characterize the degree of fragmentation in vegetation area was proposed by optimizing an algorithm, and had been effectively used in object-based classification. The main contents and results are as follows:1. Local indicators of spatial association (LISA), originally developed for point data, was used to measure the spatial autocorrelation of pixels within an image to obtain texture in this research. LISA was used as the ancillary spatial information to be integrated into the object-based classification method for obtaining the better segmentation results and improving the accuracy of the identification of Torreya. Representative subset polygons have been selected as samples to derive semi-variograms representing each vegetation structure. Semi-variogram analyses on four spectral bands (B, G, R, NIR), NDVI and RVI were carried out for subset samples of all the vegetation categories. Semi-variance analysis proved to be useful for assessing the separability of vegetation structure and suitability of spatial scales for texture calculation. For each subset sample of a vegetation structural class, three LISA textures (Moran's I, Getis-Ord G, Geary's c) were calculated for spectral bands, NDVI and RVI based on the corresponding ranges derived from semi-variogram analyses. A Z-test was used to identify the textural band that provides the largest statistically significant differentiation (the highest Z value) between any two vegetation classes. The LISA textural bands with the highest Z-value provided the best discrimination between two vegetation categories, and thus were selected to be incorporated in the object-based segmentation and classification. The results indicated that the addition of LISA textural bands results in better segmentation results. If we rely solely on the spectral for segmentation, the shape and size of objects were highly sensitive to even a fraction of heterogeneous spectral reflectance values in a vegetation object of a given type. Highly fragmentized objects, in most cases, enhanced the risk of wrongly assigning objects belonging to the same land cover type to different categories. The incorporation of LISA texture was effective in integrating pixels of the same class into one object as texture inclusion decreased the high sensitivity of objects to spectral reflectance. Therefore, more "meaningful" objects were yielded which coincide with real entities. In the case of the classification considering the LISA texture, the accuracies of all the vegetation categories were increased, in comparison with the classification relying solely on the spectral information. The result of test for significance through Kappa analysis showed that the classification considering LISA texture was significantly better than that relying solely on spectral bands, which suggested that LISA texture provided important spatial information that could improve object-based classification.2. Normalized Difference Vegetation Index (NDVI) is a very useful feature for differentiating vegetation and non-vegetation in the remote sensed imagery. In the light of the function of NDVI and the spatial patterns of the vegetation landscapes, we proposed NDVI-based lacunarity texture to characterize the spatial patterns of vegetation landscapes concerning the "gappiness" or "emptiness" characteristics. NDVI-based lacunarity texture was used as the feature sources and used in object-based classification. A histogram thresholding method was utilized for the differentiation of vegetation and non-vegetation based on NDVI. The plot of the lacunarity value vs. gliding box was obtained to ascertain the most appropriate window sizes. Based on the most appropriate window sizes, the lacunarity texture imageries were generated and were used as the feature sources. A knowledge base of rules selected by C5.0 decision tree indicated that lacunarity texture exert an important function in the identification of vegetation categories, for the features associated with lacunarity were selected in every vegetation category. The result of test for significance through Kappa analysis showed that the accuracy of classification considering lacunarity texture was significantly better than that relying solely on spectral bands. We draw a conclusion that NDVI-based lacunarity texture can characterize the spatial patterns of vegetation landscapes concerning the "gappiness" or "emptiness" characteristics, and can be used in the object-based classification for improving the identification of vegetation categories.3. Landscape patterns can be clearly shown in VHR imagery. Landscape level metrics have important ecological meanings. However, these metrics were not designed for the classification phase of image processing and therefore can't be used in classification directly. This research detected how to carry out the scale transformation of the landscape level metrics in order to make it possible to be calculated for an extensive mosaic of patches. A multi-scale segmentation algorithm was used to construct the hierarchical system to produce the layers of "fragmentation geometry" and "spatial unit" respectively. A tool was developed using Visual Basic and ArcObjects programming library to quantify the landscape characteristics in the spatial units, and effectively combine the GIS spatial analysis and classification phase of image processing.4. In order to make landscape information of different spatial units comparable and used it in object-based classification, a landscape feature called "ratio of effective mesh size (Meffratio)" is proposed by optimizing the "effective mesh size (Meff)".Some landscape metrics may have some limitations after scale transformation to be calculated in an extensive mosaic of spatial units, because they required strict condition about the number patches and the number of types. Compared with other landscape metrics, Meffratio had an overwhelming advantage in characterizing the landscape structure for spatial units with respect to the algorithm. This is because Meffratio can be calculated for every spatial unit, regardless of whether the spatial unit contains patches of a specific type or not and the number of the patches present within it. Meffratio is consistent with the real characteristic and thus provide important ecological meaning. Furthermore, fragmentation geometry and spatial unit can be defined easily and flexibly enough. "What is the fragmentation geometry" and "which scale is the spatial unit" are totally defined according to individual purposes and specific application needs in different researches. The conclusion is drawn that Meffratio is an effective and feasible landscape metric, which can be flexibly used and illustrates a good prospect of application and extension in object-based classification.