| Wetlands are characterized as good habitats for highly endangered wildlife that are capable of controlling environment and protecting species genes. However, it exists the complex characteristics of wetland habitats due to the wetland locating between land and water bodies, and it is well known that inaccessibility to wetlands often causes a large difficulty to do wetland research. Hence remote sensing plays an important role in the wetland scientific research as a useful tool to generate parameters of ecological and environmental process of wetlands. Especially, it has achieved so much on the ability of high resolution imagery and its application methods recently. In this study, the Honghe National Nature Reserve (HNNR) was selected as the study area, which locates in the Northeast portion of the Sanjiang Plain in China. And HNNR has been listed as a key international wetland within the Ramsar list in2002. The study used two types of remote sensing data, one is acquired by a camera system equipped on unmanned airship with a very high spatial resolution of0.13m, and the other is satellite remote sensing data SPOT-5with different scale spatial resolution images (10m×10m,20m×20m,30m×30m). Different data type using different classification methods, in our study, we used objected-based and pixel-based classification methods applying to high spatial resolution and low spatial resolution images. However, all of the images were classified at a community plant scale in marsh wetlands.And a very detailed classification system of wetland plants was made for the9types of plant communities. Object-based classification method, the approach for classification based on subjects (groups of pixels) rather than each single pixel, was used to delineate and map the different wetland communities as a new methods. For detecting the efficiency of the different classification methods of remote sensing, the authors also attempted another method of supervised maximum likelihood classification for this wetland mapping. The result indicates that:(1) Airship-imagery can fully characterize the detailed plant features such as plant shape and structure, the different vegetation types such as marsh, meadow, various arbors, and shrub, can all be derived from our images at plant community scale with an overall accuracy of91.77%;(2) By comparison between the object-oriented classification method especially for the high-resolution imagery and the traditional maximum likelihood classification method, authors can conclude that the former classification method has a higher accuracy, while the latter result is not so satisfactory. Hence, one conclusion from this research indicates that the selection of classification method is very important for wetland mapping at a community scale by using remote sensing technique;(3) Our wetland mapping result shows that the spatial distribution pattern of wetland plant communities are controlled by both the environmental gradient of wetness and micro-topographies of wetlands, showing a mutual alternative zonal distribution pattern within the HNNR.Due to the high price and difficult acquired, though high spatial resolution imagery had high classification accuracy assessment, our study explored lower spatial resolution images to classify at a plant community scale. Compared to finer resolution imagery, lower resolution images with indecisive boundaries and no more detailed information were classified as pixel-based method. To infer spatial information from a finer to a coarser spatial resolution, we proposed and compared four up-scaling methods (nearest, bilinear, cubic, majority), the widely used Window Averaging (WA) method. We applied and compared these methods in a case study in which SPOT-5images were aggregated from10m×10m to20m×20m, and30m×30m images (Fig2). At each level, data are aggregated directly from the original images (e.g. from10m×10m to30m×30m) instead of a precious aggregation (e.g. from20m X20m to30m×30m). We proposed up-scaling vector-based high accuracy classification as training samples to the supervised classification method applying to the tree scaled SPOT-5images. The result showed that this kind of method provide retailed information about various marsh wetland features in a relatively quick and simple manner and seemed to be well-suited to map marsh wetland cover types. The advantage of this training sample lies in the operator unable to have experience and knowledge. The classification did not depend on bands composition or the time image acquired, if only it is the same area and not much changed. The accuracy assessment based on KAPPA accuracy measures suggested that classification accuracy of four maps appeared to be mostly was over75%. However, owing to the nature of wetland and scene characteristics, the high classification maps vector data as the training samples applied on lower image has been judged to be adequate for this purpose of the study. Although the application of the vector training sample applied on images offers the advantage of simplicity and ability to apply the method over much scale images, the classification maps did not achieve enough high accuracy result. Maybe that is the limitation of pixel-based classification, so it is important to select appropriate classification system scale. |
PDF Full Text Request