Study On Urban Impervious Surface Extraction Based On Multi-features And Random Forest

The acceleration of urbanization leads to the rapid expansion of urban built-up areas and the land use types in urban areas,such as vegetation and water body,are gradually replaced by a large amount of impervious surface.The rapid expansion of impervious surface areas makes cities suffer from urban waterlogging,and intensified heat island effect.It is very important to extract impervious surface information with high efficiency and precision for urban spatial planning and ecological environment improvement.Remote sensing technology has been widely used in the research of impervious surface extraction due to its wide monitoring range and strong real-time performance.When extracting the impervious surface in pixel scale,the spectral confusion has always been the major problems affecting extraction accuracy.In order to reduce this spectral confusion,the main urban areas of Zhengzhou and Hangzhou were taken as the research areas and the urban impervious surface extraction was studied based on multi-features and random forest.Firstly,Sentinel-2 images,Sentinel-1 images and Luojia 1-01 images were used to extract multi-features.Secondly,a feature selection method with null importance was proposed to select the features.Finally,the improved random forest algorithm was used to classify the images,and the high-precision impervious surface information was obtained.The main research results are as follows:(1)The applicability analysis of commonly used impervious surface extraction algorithms.The maximum likelihood method,Support Vector Machine(SVM),Artificial Neural Network(ANN)and Random Forest(RF)algorithm were used to classify Sentinel-2 multispectral images,and obtain the impervious surface.The results showed that the accuracy of random forest algorithm was 2.9%,4.95% and 3.61%higher than the other three algorithms in Zhengzhou,and 5.41%,4.78% and 3.97%higher than the other three algorithms in Hangzhou.Based on these results,the random forest algorithm was used as a classification algorithm to study the influence of multifeature information on impervious surface extraction results.(2)The multi-features extraction and feature selection method were established to enhance the information of impervious surface.Sentinel-2 images,Sentinel-1 images and Luojia 1-01 images were used to extract spectral features,texture features and temporal features,and their effects on the accuracy of classification results were studied.A feature selection method based on null importance was designed to correct the feature importance measurement deviation,and feature importance analysis was carried out.The results showed that compared with using original bands,the multi-feature fusion can significantly reduce the confusion between impervious surface and bare land or vegetation.For Zhengzhou and Hangzhou,the overall classification accuracy was improved by 8.90% and 6.10%,respectively.The feature selection method based on null importance can eliminate redundant and ineffective features.Compared with the multi-feature fusion method,its overall classification accuracy of the two study regions was improved by 0.30% and 0.80%,respectively.(3)An improved Random Forest classification algorithm considering spatial neighborhood information was proposed to obtain the high precision impervious surface information.The PLR method was introduced into RF algorithm,and its category probability was modified to obtain the improved random forest algorithm,and the improved random forest algorithm was obtained.The impervious surface was extracted based on the optimal feature combination and the improved random forest algorithm.The experimental analysis showed that the improved Random Forest could significantly improve the pepper and salt phenomenon of classification results,and improve the accuracy of impervious surface extraction results.The overall accuracy of classification result in Zhengzhou was 95.60%,and that in Hangzhou was 96.80%.