Abstract:

Objective: The aim of this study was to compare the effects of land cover classification and dominant tree species identification in the core area of Chongli Winter Olympic Games under different method combinations of domestic Gaofen-2(GF-2) satellite images, and to analyze the influences of spatial resolution, processing unit, feature set and classification algorithm on the overall accuracy and tree species classification accuracy, so as to provide empirical reference for the research on land cover and tree species classification in Chongli district and promote the industrial application of domestic Gaofen series data. Method: Taking the Chongli Winter Olympic core area as the research object and GF-2 image as the data source, the land cover classification results of 50 different method combinations were compared and analyzed from four dimensions: different units(pixel and object), different image spatial resolutions (1 m and 4 m), different feature sets(spectral features, texture features and shape features) and different classification algorithms. At the same time, two dominant species of Betula platyphylla and Larix principis-rupprechtii were identified. The UAV(unmanned aerial vehicle) aerial image and sub-compartment data were used to obtain training and test samples. The overall classification effects under different methods were evaluated through the overall accuracy(OA) and kappa coefficient. The classification accuracy of dominant tree species was evaluated by the harmonic average(F1) calculated by production accuracy(PA) and user accuracy(UA). Result: 1) The land cover classification accuracy of the MLC(maximum likelihood) method at pixel level, 4 m resolution and spectral feature set was the highest, with an OA of 79.65% and Kappa coefficient of 0.722. The highest F1-score of Larix principis-rupprechtii was 0.79, and the corresponding classification method combination was Bayes method at object level, 1 m resolution and spectrum + texture + shape feature set. The highest F1-score of Betula platyphylla was 0.77, and the corresponding classification method combination was Bayes method at object level, 1 m resolution and spectrum + texture feature set. 2) There was no definite response between the classification accuracy, spatial resolution and feature set of different classification algorithms. When other conditions were controlled to be the same as possible, the improvement of spatial resolution and the increase of features didn't necessarily improve the classification accuracy. Under different spatial resolutions or feature sets, the response direction and degree of the same classification algorithm to the change of one other factor were also different. 3) Land cover classification performed better at the pixel level. There was no significant difference between the pixel and object level of Larix principis-rupprechtii, and those of Betula platyphylla performed better at the object level. 4) SVM(support vector machine) classification algorithm performed consistently well with high accuracy under different units, different resolutions and different feature sets. Besides, the supervised statistical classification algorithms MLC and Bayes also had good performances. Conclusion: The use of GF-2 data might have promising performances in land cover classification and dominant tree species identification in the core area of Chongli Winter Olympic Games. The classification effect may be influenced by multiple factors such as spatial resolution, processing unit, feature set and classification algorithm.

Key words: Winter Olympic core area, GF-2 image, land cover, tree species classification, spatial resolution