Study On Extracting Greenhouse From Remotely Sensed Imagery Using Object-Based Image Analysis

Greenhouse is a kind of major protective measures for overwintering vegetable plantation and seedling growing in cold North China. Greenhouse plays an important role in increasing the off-season vegetable supply and promoting the development of rural economy, and increasing farmers’income. Therefore, it is crucial to understand its spatial patten. As a rapid method for surveying the surface resource in large scale, remote sensing technology is of very strong pertinence to extract greenhouse information.Traditional pixel-based image analysis methods extract remotely sensed information mainly based on single pixel by using supervised or unsupervised classification. It is hard to satisfy current demands due to many limitations. An obvious tendency appears recently to study new extracting methods for resources surveying. In this paper, a method named as object-based image analysis (OBIA) is used to extract greenhouse efficiently from remotely sensed imagery.In this paper, Zhongwei city and its contagious region, Ningxia Autonomous Region, are chosen as research area. QuickBird、ALOS and Landsat TM image with different spatial resolution are used as source data to study the extraction methods of different greenhouse remotely sensed information (individual greenhouse、contiguous greenhouse) using OBIA. Firstly, the scale problems in OBIA were studied. These problems include:connotation and extension of scale、the performance of methods which are used for up-scaling and the selection of optimal or appropriate spatial resolution. Then the appropriate spatial resolution images were selected for different greenhouse information extraction target and the optimal combination of parameter values were selected for different spatial resolution images. Secondly, the image object characteristics to distinguish greenhouse and other surface features were studied, and the greenhouse information extraction rule-sets were developed for different spatial resolution images. Then the greenhouse information was extracted from different resolution images. Finally, this paper has discussed the methods used for accuracy assessment of remotely sensed information extraction. Then the greenhouse information extraction accuracy from different resolution images was assessed, and the methods used for extracting greenhouse information from different resolution images were comparatively analyzed.The main conclusions of this paper are as follows:(1) In the object-based image analysis, scale is indicated by means of image spatial resolution and segmentation threshold. For a specific surface feature, there is a maximum spatial resolution, the remotely sensed information of this surface feature is possible extracted from images whose spatial resolution less than the maximum spatial resolution. There is also an appropriate spatial resolution image from which the information extraction of this surface feature can get the highest accuracy.(2) To the pixel-based image up-scaling, the local averaging method outperforms the nearest neighbors, bilinear interpolation and cubic convolution. Traditional local variance method and Statistical separability method have the ability to select the maximum image spatial resolution, but the former tends to overestimate it. This paper thought that the maximum spatial resolution for individual greenhouse information extraction is lm, while20m for contiguous greenhouse information extraction.(3) In the Object-Based Image Analysis, the optimal segmentation threshold is substantially a combination of parameter values (scale factor、shape factor、 compactness factor). The optimal combination of parameter values selected using the method (PSE、NSR、ED) proposed by Liu. et al. are the most precise. So this method is a better method for selecting optimal combination of parameter values.(4) To contiguous greenhouse, the segmentation quality of15m resolution image is the highest; the separability between greenhouse and other surface features of6m resolution image is the best; but the greenhouse information extraction accuracy of9m resolution image is the highest. To individual greenhouse, the segmentation quality of lm resolution image is the highest; but the separability between greenhouse and other surface features and the greenhouse information extraction accuracy of0.61m resolution image is the best. This shows the information extraction accuracy is influenced by image segmentation quality and the separability between different surface features simultaneously. (5) The spectral characteristics of image objects are the most main characteristics for extracting both contiguous greenhouse information and individual greenhouse information. To extracting contiguous greenhouse information, the most main spectral characteristics can be used are the first-order and the second-order statistical characteristics of pixels’ gray value. And To extracting individual greenhouse information, the most main spectral characteristics can be used are the first-order statistical characteristics. The shape characteristics can be used to distinguish surface features which have spatial shape only, for example, greenhouses and roads. The texture characteristics can cover the shortage of the spectral characteristics and improve the extraction accuracy of contiguous greenhouse information, but haven’t effect for extracting individual greenhouse information.