Terrace Information Extraction Form Remote Sensing Image Based On Object-oriented Classification Method

Terrace is a widely used soil and water conservation engineering practice and plays an important role to reduce soil and water loss in the Loess Plateau region of China. Accurate and rapid extraction of terrace information has become one of the core technologies for the dynamic monitoring and evaluation of soil and water conservation at a regional scale. In this paper, a typical watershed of Yan’gou watershed in the Hilly and Gully Region of Loess Plateau was taken as research area to study the object-oriented classification method with the high spatial resolution SPOT5 imageries on the system of ENVI. The image objects were set through the image segmentation firstly. Then, the rules for terrace information extraction from remote sensing datasets were established via the analysis of spectral, texture and spatial feature of image objects. In this case, the automatic extraction of terrace was achieved. Finally, the accuracy of object-oriented classification method extraction results was evaluated comparing with traditional supervised classification, unsupervised classification and manual visual interpretation results. The main results derived from this study are as follows:(1) It is critical to set a reasonable segmentation threshold for image segmentation when using the object-oriented classification method for the extraction of terrace, and the effect of segmentation is directly related to the level of accuracy of terrace extraction. If the segmentation threshold is too small, segmentation will be broken and the integrity of terraces will be destroyed in a certain extent. If the threshold is too large, some of the terraces will merge into woodland and grassland nearby. The two cases are not conducive to the terrace extraction effectively. Based on comparatively analyses from the preview window the feature extraction model of ENVI provided, a reasonable segmentation and merging threshold for terrace extraction were set as 38% and 88%, respectively.(2) Object-oriented classification method can fully exploit the image information, and comprehensive utilize spectral information, texture information and space information of image objects to provide more basis for terrace information extraction. This study applied rule based feature extraction method to extract terrace information. With the premise of the features the features or multi-features combination chosen have the most information, can be able to distinguish features easily and ensure the accuracy, in the principle of choosing the features as little as possible, the mean of NDVI, minimum value and maximum value of spectral properties, range and variance of texture properties, and the area attributes of spatial properties were chosen to identify terrace information on the basis of statistic and analysis of spectral features, texture features and spatial features of image objects, and the rules for terrace information extraction from remote sensing datasets were established. Finally, the automatic extraction of terrace from remote sensing images was achieved effectively.(3) The results derived from supervised classification, unsupervised classification, manual visual interpretation and object-oriented classification were compared and evaluated. The results showed that: 1) The overall accuracy of terrace extraction based on object-oriented method reached to 88.25% with a kappa coefficient of 0.76, which implying a very good quality of image classification; the overall accuracy and kappa coefficient increased by 14.75% and 0.34 respectively compared with maximum likelihood method, which was the relative optimal method among the traditional supervised classifications. 2) The extraction results of object-oriented classification method had high position accuracy and area accuracy reached to 78.38%, which was significantly higher than that of the traditional classification methods. 3) The results of object-oriented method could avoided the phenomenon of “peeper and salt” effctively, and were in accordance with the actual situation with good visual effect. This study indicates that object oriented classification technology for the extraction of terrace information has good applicability in this research area.