Analyzing Extraction Rules Of Paddy Rice In Complex Terrain Region Based On TM Data Using Object-oriented Method

In this paper, the author chose the Landsat TM-5 data received in June 6,2009 based on the rice phenological calendar in Yongtai county and tried to explore the paddy extraction rules in complex terrain regions based on medium-resolution data using object-oriented method. First, the optimal segment of paddy was achieved by giving priority to the multiresolution segmentation combined with the spectral difference segmentation. The optimal segmentation parameter of paddy was acquired by visual judgment and quantitative analysis based on the differences of local variance of adjacent scale segment. In addition, DEM was used as segmentation data considering the terrain factor of study area. Then, the decorrelation analysis and the inter-class J-M distance were combined to select the optimum classification features from the relevant features of paddy extraction. Finally, the multi-level paddy extraction rule set was obtained based on the improved SEaTH algorithm, simply together with conventional object extraction rules, and the rice planting area was extracted and mapped. Based on the rice extraction results, the confusion matrix was calculated to evaluate the precision and test the portability of the method.The results indicated:(1) The method of combining the differences of local variance of adjacent scale segment with visual judgment could get precise segmentation parameter and good segmentation result, and it improved the defects of trial-and-error approach which were subjective, unchecked, and time-consuming. For complex terrain regions, DEM data could improve the segmentation result. (2) The combination of decorrelation analysis and the inter-class J-M distance could efficiently select the optimal classification features; it was an efficient object-oriented feature selection method. (3) Information extraction was mainly based on spectral features, especially custom characteristic indexes. For medium-resolution remote sensing images, texture features and shape features were not extremely effective in distinguishing surface features. In regions with a complicated distribution of paddy (paddy was both distributed in low and high elevation areas), the separation effect of DEM was not significant. (4) EVI2 and SAVI were the best features in distinguishing paddy and woodland, while MSAVI2 and Mean_TM4 had the strongest separation ability between paddy and water. (5) Multi-level automated classification method, based on object-oriented notion together with the improved SEaTH algorithm and conventional object extraction rules, could effectively improve the paddy extraction accuracy. The overall classification accuracy and Kappa coefficient were more than 90% and 0.8 respectively. The proposed method has a high degree of automation and is worthy of popularization. It is of great significance to paddy extraction based on medium-resolution data in complex terrain regions.