| The Chinese Loess Plateau is one of the four plateaus in China.It is a typical ecological fragile area,and this area is characterized by gullies,broken surface,and serious soil erosion and water loss.As the main body of geomorphic pattern in the Chinese Loess Plateau,gully landform is not only the main inducement of soil erosion and water and soil loss,but also the product of soil erosion and other disasters to a certain extent.Therefore,it is of great significance to carry out the spatial analysis and characteristic study of gully landform in the Chinese Loess Plateau region for local soil and water conservation,environmental improvement,ecological balance and so on.At present,there are few large-scale studies in the Chinese Loess Plateau gully landform extraction and the extraction accuracy is not high.In order to study the spatial distribution characteristics of gully landform in the Chinese Loess Plateau,explore the relationship between the spatial distribution of gully landform and environmental characteristics,and establish an accurate extraction model of gully landform in a large scale.In this paper,the Yanhe River Basin in the Chinese Loess Plateau is taken as the research area,and the samples of gully and slope in the Chinese Loess Plateau are artificially extracted with the help of Google Earth Pro platform,and the Chinese Loess Plateau gully landform is extracted based on remote sensing image and machine learning,and the corresponding extraction model was established,and the model accuracy and extraction results were compared.The main research results and conclusions are as follows:(1)The prediction and spatial distribution of large-scale gully landform in Yanhe River Basin based on logistic regression model are analyzed.The results show that:The model predicted Brightness,Greenness,Wetness,first principal component(PCA1),third principal component(PCA3)and slope of geomorphic features as the optimal feature combination for gullies extraction.The prediction accuracy is determined by optimal logistic regression model of the spatial distribution of gully landform was 73.73%and the AUC value was 0.805.The gully landform of Yanhe River Basin accounts for about 52.05%of the whole basin area,and its spatial distribution is gradually concentrated from northwest to southeast.(2)Based on random forest,support vector machine and neural network machine learning model,the gully landform spatial distribution extraction models was established and the results shows that:in terms of model characters,the random forest model established by Coastal,Blue,Red,NIR,SWIR,GLCM4,elevation and positive and negative terrain(PNT)characters determined by remote sensing image in December and DEM data has the best gully extraction effect,and its overall accuracy was 88.33%,and the AUC value was 0.952.The three models predicted the gully spatial distribution in the sub-basins,which all showed a gradual concentration from northwest to southeast.(3)Based on 6 verified samples of 1km~2×1km~2 uniformly distributed in the test area,the prediction accuracy and differences of gully landform in different machine learning models are compared,and the results show that:Compared with the other three models,the random forest model has the best prediction result of the spatial distribution of gully in sub-basins,and the consistency between the gully contour and the artificial recognition of the gully landform is the highest.The average overall accuracy of the six samples based on the random forest model was80.48%,higher than the 76.48%of the SVM model,75.87%of the LR model and71.85%of the ANN model. |
