Abstract:
Abstract: Spatial distribution status of construction land is closely related to the regional economic and social development. Therefore, timely monitoring and delivery of data on the dynamics of construction land are far-reaching for policy and decision making processes. Classifying land-use/land-cover and analyzing changes are among the most common applications of remote sensing. One of the most basic and difficult classification tasks is to distinguish the construction land from other land surfaces. Landsat imagery is one of the most widely used sources of data in remote sensing of construction land. Several techniques of construction land extraction using Landsat data are described in some literatures, but their applications are constrained by low accuracy in various situations, and usually using the technique of single index or multi-index. The purpose of this study was to devise a method to improve the accuracy of construction land extraction in the presence of various kinds of environmental noise. Thus we introduce a multi-features decision tree (DT) classification model for improving classification accuracy in the areas that including bare land, shadow and some streams, in which the other classification methods often fail to classify correctly. The model integrates four spectral indexes, the pattern recognition technique and spatial algorithms. The four spectral indexes are the normalized difference three bands index (NDTBI), the normalized difference building index (NDBI), the modified normalized difference water index (MNDWI) and the normalized difference vegetation index (NDVI) respectively. The pattern recognition technique is referred to support vector machine (SVM). And the spatial algorithm is to create buffer zone.The test site was deliberately selected so that it consists of complex surface features, such as bare land, hill shade, and some small streams that are liable to be mixed up with construction land on the Landsat imagery. For that reason, Landsat-8 OLI images (path/row 128/41) were selected in sight of its perfect performance in quality. All the Landsat images used are of product type L1T and were geometrically corrected and converted to top-of-atmosphere reflectance consequently. The subset image of one transition zone between urban and rural region in Bijie city of Guizhou province was selected as the test site, the area of which is 144 km2. Besides, the urban center of Qixingguan district of Bijie city was selected to produce thematic map of construction land for application inspection of classification robustness.Building decision tree (DT) nodes is the key of the methodology. Firstly, a new index called NDTBI was developed to be combined with NDBI to identify all the construction land from background. NDTBI and NDBI were separately arranged into the first and second node of DT. However, after executing the DT of the two nodes NDTBI and NDBI, noise such as bare land, shade, water existed together with all construction land at the same time. Secondly, MNDWI and NDVI were added into DT nodes in order to separate water and shadow from construction land. After executing the DT of the above four nodes, the bare land and streams, which share the same spectral feature with construction land, were the only objects mixed up with construction land. Therefore, the SVM classification, in the light of the optimal performance in land-cover classification, provides the opportunity to suppress the bare land noise. Meanwhile, the spatial algorithm was used to separate the small streams from construction land. To execute SVM classification, the training samples were selected by using principal components (PC) transformation of the multi-spectral images. The spatial algorithm is to create buffer zone for the small stream using vectors from visual interpretation. As a result, a segmentation binary image was obtained and added into DT. Two principles were proposed for the buffer zone creating: on the premise that construction land could not be included in the buffer zone, when some channel segment of stream is closely adjacent to the construction land, the width and shape of buffer zone should be same as that of river.To assess the accuracy, the high resolution fusion image (15 m) was taken as the reference dataset, a total of 564 sampling pixels were randomly drawn from the strata (i.e., construction land and non-construction land), 277 pixels from construction land and 287 from non-construction land. The accuracy assessment for test site reached 97.52%. Besides, for application inspection of the classification robustness, the Landsat-8 multispectral images of the urban center of Qixingguan district was used to produce the thematic map of construction land, which proved to be perfect with the overall accuracy 98.03%.The method developed in the paper is easy for understanding and operating, widely applicable, and capable of separating the disturbing noise from construction land. Furthermore, the model dramatically improved the accuracy of extracting construction land from the moderate resolution images. And the validation results suggest the widely applicable in western region of China, especially for the area with land-use/land-cover diversity.