Research On Land Cover Classification Based On Random Forest Classifier

Global change is an important issue in the scientific community.Land use/cover plays a crucial role in studying global climate change and characterizing land surface processes.To date,most studies have mainly focused on extraction of land cover using low-resolution remote sensing data(300 m-1k m)which was characterized by low spatial resolution,short time span,phenomena of mixed pixels and inconsistent between different products.Moreover,mediumresolution imagery derived land cover information was also limited by visual interpretation and scene-by-scene supervised classification.Based on the study of selection of classification features and classifiers for medium-high resolution satellite remote sensing data,the study proposed an approach of automatic land cover mapping in Wuhan urban agglomeration(WUA)and Daxing gan ling Region.Firstly,the paper analyzed the Feature Selection(FS)and Random Forest classifier and summarized the character of spectral feature,textural feature and Topographical feature.Secondly,a classifier model was built with spectral feature,textural feature and seasonal characteristics of each land cover types in the study area for the development of an regional classification of land cover.Finally,this paper acquired the forest cover change based on existing land cover data such as FCC and FNF,and analized the driving factors for land cover changes.Details are as following:1.Feature Selection of Landsat classification characteristics and classifier: This paper presented a method of involving features in machine learning based on studying the spectral feature,textural feature and other features,which reduced the dimension of the features.2.Land cover mapping based on regional classification in a big area: On the basis of pre-processing of Landsat imagery,the test using RFC for one Landsat surface reflectance image was carried out,then a classifier model RFC was built for regional classification of land cover.The result showed that the overall accuracy of the regional classification was more than 80%.3.Forest Cover mapping based on space-time extensional classification: on the one hand,forest cover mapping based on a feature extension method with time series Landsat imagery showed that the overall accuracy achieved was between 80.18% ~ 90.23% with the DOY(Day of Year)less than 31;on the other hand,forest cover mapping based on a feature extension method with neighbour Landsat imagery showed that the overall accuracy achieved was between 83.96% ~90.43%.4.Forest cover change and driving factor analysis in Dahinggan Region: Based on the forest cover mapping in 2000 to 2015 from FCC or FNF,the paper analyzed the forest cover change and driving factors in Dahinggan Region.The results showed that the forest cover in Dahinggan Region rose from 2000 to 2015,the driving factors of forest cover change in Dahinggan Region included economical,policical and natural factors.