Remote Sensing Classification In Forest Wetland Compatible With The Optical,Radar And Terrain Aided Data

Forest Wetland is an important part of wetland,the spatial distribution of ithas important value in many studies including wetland functioning assessment,greenhouse gas flux estimation,and wildlife suitable habitat identification.In addition,it isa difficult and hot spot in the field of wetland ecology.Researchers have demonstrated that forest wetlands and their adjacent land cover types can be identified by image analysis techniques of multi-source remote sensing data,however,there isn't still to reach a consensus on the optimal approaches of mapping forested wetlands.Therefore,we take Dazhanhe National Forest Wetland as study area,and a comparative study of the Object-based classification using the RF algorithm and per-pixel RF algorithms,make object-based KNN,using predictors derived from Landsat-8 imagery,Radarsat-2 advanced synthetic aperture radar(SAR),and topographical indices.Combined with the field measured data,and made precision evaluation of the algorithm,and analyzed the remote sensing classification error.The main conclusions are as follows:(1)The Landsat8 remote sensing data was segmented by object-oriented multi-scale segmentation method,the feature values of optical,texture,radar and topography data are extracted based on the results of segmentation,and the mean value of each object characteristic variable is used as the prediction variable to be taken into the RF algorithm for the interpretation of the patches in order to construct a classification method based on object-based classification and the RF algorithm,what's more,it provides a method reference for the forest wetlandinformation extraction in the high latitude cold temperate zone.(2)The Object-based classification using the RF algorithm results show that the classification of land cover is better,the overall accuracy and Kappa coefficient of object-based classification using RF algorithm is 80.65% and 0.78,and the producer accuracy for the forested wetland,coniferous forest,broad-leaved forest were 83.33%,83.78%,81.39%,respectively.The results also show that the omissions for the forested wetland and coniferous broad-leaved forest can be improved by incorporating the flooding range into the categorical variable.(3)Object-based classification using RF algorithm performed better than the classification results based the object-based classification using the KNN algorithm and the pixel classification using RF algorithm,and it had the highest overall accuracy and Kappa coefficient.It is proved that forest wetland information extraction is more accurate by object-based classification using RF algorithm.However,there were obvious omissions for the forested wetland and herbaceous swamp,and a lot of noise in the classification results produced by the per-pixel RF algorithm.Meanwhile,object-based classification using the KNN algorithm had the lowest classification accuracy,it is mainly because of the small sample size of the land extraction process prone to error by it.