Research On Automated Morphological Landform Types Classification Methods Based On The 1:1,000,000 DEM

The landform have the extensive and profound influence on the geographical environment,and it is the dominant factor of the natural region synthesizes.The morphological types and the genetic types of landforms is a unit of causality.Specifically,the morphology is an important basis for the research of the genesis and the results of it.Therefore,it has the vital significance to the research of the geomorphologic pattern analysis and the geomorphological regionalization that classified the morphological types of landforms efficiently and accurately.The pixel-based traditional method of landform classification derived from the remote sensing image analysis field usually results in generating the "salt and pepper effect" in the final results and has a low precision which makes this kind of method difficult to meet the actual demand.In order to maintain the integrity of the classification results and to improve the accuracy,this paper research proposed to automatic classify the morphological types of landforms by using the object-based method and the Random Forests method based on the 1:1,000,000 DEM and the terrain factors derived from the DEM considering that the object-based classification can prevent the "salt and pepper effect" of the classification results and the Random Forests have the advantages of high accuracy,high efficiency and avoiding over fitting.This study obtained the best combination of terrain factors for landform classification by correlation analysis and Sheffield's entropy method.Then the morphological landforms classification is implemented by using the object-based method and the Random Forests method respectively based on the eCognition software and the statistical toolbox of Matlab.The experiment results show that both of the overall accuracy of the two classification methods are high.Specifically,object-based classification results have good integrity but its overall accuracy is a little low.By the contrast,the Random Forests classification results have high overall accuracy but the local patches of the results are fragmentized.For overcoming these shortcomings,two kinds of methods combined with object-based and Random Forests classification are proposed.First method put forward to treat the classification results based on the Random Forests as thematic map and add it to the multi-resolution segmentation procedure which can make the outlines of the segmented image objects better fit with the actual landform.Thus can improve the classification accuracy.The second method aim to put the segmented image objects as the research units of the Random Forests.Comparing all of the classification methods of this paper research,the results by adopting the Random Forests based on the object-based classification have the best integrity and the highest accuracy.