Land Use/land Cover Information Extraction From SPOT6 Imagery With Object-oriented And Random Forest Methods In The Huangshui River Basin

Accurate land use/land cover data has important guiding significance for regional land planning,urbanization and ecological environment protection.This paper takes SOPT6 imagery as research data,based on the random forest classification method of one of the object-oriented classification method and the machine learning method,extracts the land use/land cover information of Huangshui Basin respectively,and evaluates the classification accuracy of the two methods Compared with comparative analysis,it aims to determine a more suitable method for extracting land use/land cover information for the study area.The main research conclusions are as follows:(1)In this paper,Object-oriented classification uses multi-scale segmentation and multi-level rule-based classification.The optimal segmentation scale is between40-150,when the color factor weight is higher than the shape factor weight,and the compactness factor weight is higher than the smoothness factor weight,the segmentation results are more in line with the characteristics of the local objects in the study area.The three division layers are determined,the Level1 division scale is determined between 110-150,used to distinguish vegetation from non-vegetation;the Level2 division scale is between 60-90,used to distinguish the first-level feature type;Level3 division scale is 40-60 is used to distinguish the type of secondary features.This segmentation process is in line with the actual situation,and a good classification result is obtained accordingly.(2)Object features in object-oriented classification become an important basis for image analysis.The aspect ratio and shape index are used to distinguish rivers and reservoirs;urban and rural,industrial and mining,residential land in the band average and texture characteristics are significantly different from other features;cultivated land and forest land use near infrared band standard deviation,red band mean,Standard deviation and synergy,variance,mean,standard deviation and other featuresin texture features are combined to distinguish;forested land and sparse forest land in the study area are mainly distinguished by DEM data.It shows that the complementarity of multi-object features is very necessary to improve the accuracy of image classification.(3)In the random forest classification,the importance evaluation and screening of the 24 classification features of the three partitions are performed.The results of the importance evaluation of the feature parameters of the three partitions show that in the spectral features,except for the original four multispectral bands of the image,both NDVI and NDWI It is more important.The altitude and slope are more important in the terrain factor.The mean,homogeneity,dissimilarity and entropy in the texture features are more important.(4)In the assessment of classification accuracy,the object-oriented multi-level classification rule method for the entire river basin has an overall classification accuracy of 90.93% and a Kappa coefficient of 0.90.The overall classification accuracy of the random forest classification method based on machine learning is90.63%,and the Kappa coefficient is 0.90.The classification accuracy of the two classification methods is higher than 90%,Object-oriented classification accuracy and classification efficiency are better than random forest classification.(5)In the classification results of the random forest method,a more serious phenomenon of "salt and pepper" appears,which affects the classification accuracy to a certain extent;and the object boundaries in the object-oriented classification results are clear and regular.The object-oriented classification results are more in line with the actual feature distribution,and are more suitable for the study of high-resolution image information extraction in complex terrain areas than the random forest classification method.