Study On Land Cover Information Extraction In Southern Hilly Area Based On Remote Sensing Data And Machine Learning

Land-Use and Land-Cover Change(LULC)is an important part of global environmental change.Timely and accurate grasp of national land use information is a solid foundation for land use structure adjustment,rational development of land resources and dynamic monitoring of land use.In hilly and mountainous areas of southern China,due to the complex terrain,broken distribution of ground objects,and affected by cloudy and rainy weather,there are few high-quality optical remote sensing data and lack of effective and accurate land use information extraction methods.In recent years,machine learning methods have achieved good performance in the field of land use / cover information extraction and have been widely concerned.Aiming at the problem of low classification accuracy caused by terrain fragmentation and mountain shadow in southern hilly and mountainous areas,this paper takes the source area of Dongjiang River as an example,combines with the characteristics of southern hilly and mountainous areas,uses Sentinel-1,Sentinel-2A and DEM multi-source remote sensing data to extract 31 indicators,constructs 6 feature variable sets,and designs 6 experimental schemes.At the same time,combined with random forest bag outside error(OOB)and feature recursive elimination method,the feature variable selection and feature importance sorting are carried out,and the best classification combination for southern hilly information extraction is screened.In this paper,four machine learning algorithms,namely K-Nearest Neighbor(KNN),Support Vector Machine(SVM),Random Forest(RF)and Extreme Gradient Boost(XGBoost),are constructed and compared to explore the performance of adding red edge features,radar features and terrain features to the extraction of land use classification information in southern hilly and mountainous areas.The main conclusions are as follows:(1)In the process of feature variable selection and importance ranking,the variable combination after feature selection can reduce all feature variables from 31 to 15 under the condition of ensuring the optimal accuracy of each classifier,which improves the classification efficiency and performance.The scores of feature importance from high to low are: B3,B12,B2,B5,MNDWI,B11,B6,B8,DEM,NDVI,PODU,GNDVI,RVI,CIre and MSAVI,among which the highest score is the spectral characteristics of the image;(2)In the extraction of land use information from the source of Dongjiang River,the spectral characteristics,non-red edge vegetation index and water body index of Sentinel-2A remote sensing image data are used only,and the classification accuracy of each classifier is improved by introducing red edge features,radar features and terrain features.The terrain features have the greatest impact on the classification accuracy.The overall accuracy of KNN,SVM,RF and XGBoost is increased by 1.61 %,3.07 %,3.8 % and 3.65 %,respectively,and the Kappa coefficient is increased by1.87 %,3.62 %,4.51 % and 4.35 %.In the optimal variable combination,the classification performance of XGBoost is better than that of random forest algorithm(RF),support vector machine(SVM)and K-nearest neighbor algorithm(KNN)under the same feature,and the extraction effect of land use information in Dongjiangyuan is the best.(3)For the types of cultivated land,reservoir,garden land,bare land and river features in the source area of Dongjiang River,the addition of terrain features can effectively improve the classification accuracy,and the addition of radar features helps to extract urban land and roads.It shows that the machine learning algorithm based on multi-source data can provide technical support and theoretical reference for the extraction of land use information in southern hilly and mountainous areas with complex terrain.