ObjectiveIn order to meet the needs of nature reserve monitoring, improve the precision of remote sensing inventory on vegetation types, an object-oriented classification method and machine learning algorithms were developed in vegetation classification by GF-1 WFV remote sensing data.

MethodThe study site is located in Baishuijiang National Natural Reserve, Gansu Province of northwestern China. The GF-1 WFV multispectral data, Landsat-8 OLI remote sensing data, DEM data and field survey data were employed as the key data sources. Firstly, the multiresolution segmentation of GF-1 WFV data was carried out, and the research area was divided into many polygon objects. Then spectral features, geometric features and texture features from polygon objects were extracted to vegetation classification using CART decision tree. Finally, the accuracy of classification was analyzed by error matrix based on TTA mask.

ResultIn the multiresolution segmentation process, the shape factor and compactness were set to 0.2 and 0.5, respectively, the boundary of the polygon objects was identical with ground objects. When the shape factor and compactness were fixed, the optimal segmentation scale was 40. The accuracy results showed that the overall accuracy and Kappa coefficient were exceed of 83% and 0.80 in three CART decision trees, which was superior to KNN algorithm and SVM algorithm. Overall accuracy and Kappa coefficient of CART decision tree reached 85.18% and 0.832 2 by spectral, geometric and texture features, which was better than CART decision tree by spectral features or spectral combined with geometric features.

ConclusionThe image classification based on CART decision tree algorithm and object-oriented classification method were suitable for vegetation classification in nature reserve by GF-1 WFV image, which could effectively assist the nature reserve monitoring.