Object-oriented Classification Of Sandy Lands Using GF-1 Satellite Imageries

At present, the situation for desertificaiton combating is still very grim in China. Due to irrational human activities and fragile eco-environment interactions, decline in land productivity, loss of land resources and land degradation are happening in arid, semi-arid and semi-humid areas. China is one of the countries which are suffering the most serious desertification. Therefore real understanding and knowledge of the distribution of sandy lands have a very prominent role for the analysis of dynamic change and effectiveness of the ecological engineering in sandy land. The main method is ground-based survey in traditional monitoring sandy land nationwide, however it needs more manpower, material and financial resources. Also, this based on ground survey method is not to better meet sandy land monitoring and evaluation timely in the large area. Compared with other land cover types, sandy land has complex feature and it is not only related to the coverage of vegetation but also affected by soil type. Therefore, it is imminent to research and develop a sandy land remote sensing classification method in the multi-scale.In this paper, the study area is in the Otingdag sandy lands and the data sources are GF-1 satellite data, GF-2 satellite data and TM series data. The field survey plot information is as the evaluation data. First, two methods about remote sensing images segmentation were compared and the improved methods were proposed. Second, the optimal segmentation scales were determined for the different categories of sandy land in multi-scale remote sensing image segmentation. The paper presented an optimal image feature selection method for as many as 60 kinds of image features and sandy land classification flow chart was formed. Finally, on the basis of these studies, the scale effects because of the spatial resolution in remote sensing images were researched and changes in the Otingdag sandy lands were analyzed. The main findings are as follows:(1) The Fractal Net Evolution Approach(FNEA) has high applicability to achieve higher efficiency, but the effect of a single segmentation scale is not ideal, multi-scale segmentation is needed. The Full Lambda-Schedule algorithm is based on global optimization, so the segmentation result is better, but it has a huge amount of computation, low efficiency and is not suitable for handling a wide range of remote sensing images. After the image is over-segmentation, the Full Lambda-Schedule segmentation algorithm is carried out to reduce the amount of computation and improve the operation efficiency and adaptability. In order to determine the optimal segmentation scale of each category type in multi-scale segmentation, the paper used the Jefries-Matusita(J-M) value to select the optimal segmentation scale which was used to select the optimal characteristics between categories automatically. And the method re-integrated the final classification accuracy was applied to determine the optimal segmentation scale of each category.(2) More than 60 kinds of object features were produced based on the data characteristics of GF-1. The methods of information gain ratio, J48 decision tree, random tree, standard deviation and coefficient of variation were used synthetically to select image features artificially. Based on these findings, the classification flow chart of sandy lands was formed and the sandy land distribution map in the Otingdag sandy lands was gotten. Compared with the traditional pixel-based classification research, the method had the higher accuracy, reaching 85.61%, and the salt and pepper phenomenon are not too obvious. And compared to other object-oriented classification research, the paper has the greater processing range, reaching 200km×200km and has the higher reproducibility. It is possible to classify and identify the sandy land in the wider range based on the GF-1 satellite.(3) Based on the above findings, the method of J-M value is applied to the study on sandy land scale effects because of the spatial resolution which was used to evaluate the separability between categories. There is an optimal spatial resolution for a certain category of sandy land and in this resolution the sandy land has the highest classification accuracy. In addition, on the basis of the evaluation results of sandy land in 2002 and 2013, the paper carried out the change monitoring and dynamic analysis of the sandy land in decade. The sandy land areas of the Otingdag in 2002 and 2013 were 21740 km2 and 20542 km2. The area of sandy land decreased by about 1192 km2, accounting for 5.48% of the original sandy area. The fixed sandy land areas in 2002 and 2013 were 7116 km2 and 5613 km2, while the areas of semi-fixed sandy land were 10861 km2 and 10320 km2. The shifting sandy land areas were 3763 km2 and 4609 km2. It can be seen, The main type of the Otingdag sandy land is semi-fixed sandy land, while the area of shifting sandy land is smallest. The desertification situation are curbed to some extent.