Study On Comparison Of Woodland Information Extraction Methods In Heyuan

Change in forest cover have widespread effects on the provision of ecosystem services, and provide important feedbacks to climate change and biodiversity. Researching the dynamic changes of forest resources contribute to understanding its mechanism and causation. Thereby it will make the quantity, distribution and pattern of forest resources more reasonable by building correlative artificial regularization.moreover, it will be extremely critical if the accuracy of image interpretation can be improved for better understanding the change of forest.Parametric methods such as maximum likelihood classification assume normally distribute remote sensor data,however,texture and evevation information usually are not normally distributed.As a result,some nonparametric methods are applied in this research.A pair of Landsat TM images which acquired around 1991 and 2004 was used in this article. Support vector mechine, Neural net, Decision tree these three methods have been used to extract forest cover change in the study area which highest forest cover in Guangdong Heyuan. It will be help to find the extraction method which better appropriate with the study area.These image enhancement algorithms have been used before image information extraction:Firstly, three of these index have been chosen including vegetation transformations, texture transformations,MNF transformations as well as. Secondly,the first three bands from MNF have been chosen as the input bands for gray-level co-occurrence matrics (GLCM).GLCM has been used to produce texture information ,which four of more widely used measure chosen,then 12 bands with texture informations produced.At last, the elevation band has been used for classification.After comparing and field verifying three different extraction methods, it shows that:1. The overall accurate of decision tree method reached 90.4%, kappa coefficient is 0.88 that is the best performance of these three methods.it turn out a complete classification type in the result.nearly few error happened even before the post classification.but the performance of SVM was a little bit worse than decision tree, at the same time,the overall accurate was still over 85%.The worst performance was neural net,it's kappa coefficient less than 0.8. 2. Every method have its own strongpoint and weekness in this research.The strength of Decision tree is easily to set parametric, but it's hard to use the elevation information in the input . it was quiet convient to train the sample when you use SVM, it was so difficulty to solve the the kernel function problem.The advantages of neural net are able to learn from existing examples adaptively, which makes the classification objective.but the limition is that a neural net work is often accused of being a black box.it is difficult to interpret these weights due to their complex nature.3. From 1991 to 2004, Forest structure change of Heyuan as below: Large increase in woodland, an increase of 26.42%, the increase was part of Zijin County, mainly in southern and eastern Longchuan County; land the biggest decline, a decrease of 27.3%, which is mainly a large area of grassland and dry into woodland. Forest cover increased from 41% to 62%, with the same basic fit statistics.