Object oriented land use classification of Dongjiang River Basin based on GF-1 image
-
-
Abstract
Abstract: We are aimed at the problem of the low classification accuracy of the land surface of the Dongjiang River Basin due to the patch breaking of the Dongjiang River Basin and the numerous lakes and rivers which are found everywhere in the river basin. At present, the data for the Dongjiang River Basin are mostly Landsat TM/ETM images, and the spatial resolution is low, so the multi-view image splicing is needed. The Dongjiang River Basin is in the south China where it is cloudy and rainy. The Landsat TM/ETM image with less cloud is insufficient, and it is difficult to ensure the time consistency, which influences the information extraction effect; and the cost of high spatial resolution image is high, so it is difficult to be applied in the whole basin range. The GF-1 remote sensing image is used as the data source in this paper, and we try our best to use object oriented classification method combined with fuzzy classification and CART(classification and regression trees) decision tree classification method to obtain land use classification information of Dongjiang River Basin by doing the experiments. We spare no effort to attempt a method that can make the classification of Dongjiang River Basin accurately. We read a lot of relevant taxonomy and refer to many experiments, and then launch a classification scheme. We use the software of eCognition 9.0 to complete the process of fuzzy classification and CART decision tree classification, and also we use several kinds of software such as ArcGIS 10.1, ENVI 5.3, SP1, and Qmosaic 6.0 to help do this work. Also we will combine true color remote sensing images of GF-1 to read the classification process visually. First of all, we choose the fuzzy range from 480 to 2 200, which is based on the mean of the near-infrared band, and combined with the true color remote sensing images of GF-1. When the mean value of near-infrared band is less than 480 by the experiment, it is identified as water body, and when the mean value of near-infrared band is more than 2 200, it is identified as non-water. We choose fuzzy less than membership function to distinguish between water and non-water. Then in water category, the length-width ratio index fuzzy range from 1.53 to 4.32 is used to distinguish the river from the reservoir by using fuzzy greater than function. Similarly, we observe the true color remote sensing images of GF-1. When the index of length-width ratio is less than 1.53, it is identified as a reservoir, and when the index of length-width ratio is more than 4.32, it is identified as a river. In the non-water category, we try to use the fuzzy range of the normalized vegetation index NDVI (normalized difference vegetation index) characteristic value (i.e. from 0.21 to 0.62) to distinguish the vegetation and non-vegetation. When the NDVI index is less than 0.21, it is identified as non-vegetation; when the NDVI index is more than 0.62, it is identified as vegetation. Finally, we use CART decision tree classification method based on samples to distinguish river, reservoir, garden plot, woodland, farmland, grassland, unused land and construction land. Compared with the maximum likelihood classification method and the unsupervised classification method applied in GF-1 remote sensing images, the object oriented CART decision tree classification based on sample method has the best classification effect, whose overall classification accuracy is up to 93.27%, and the Kappa coefficient is up to 0.92. This method can be used as an effective method to obtain higher land use information in Dongjiang River Basin, and it also can provide more accurate data for the study of ecological environment changes in the watershed.
-
-