Abstract:

Larch plantation has become an important tree species in north China,and the afforestation area is increasing year by year.Therefore,the extraction of larch plantation is of great significance for rational utilization of forest resources in China.Based on Mengjiagang Forest Farm within the territory of Huanan County in Heilongjiang Province as the research district,combined with phenological characteristics of larch,the study selected a typical period of GF-1 PMS images as the data source,used forest resources subcompartment data and field survey data as the sample data,extracted image spectral characteristics,texture characteristics,vegetation index and topographic features,and used Random Forest (RF)algorithm to extract the spatial distribution of larch plantation from multi-date and multi-feature angles,so as to obtain the best classification time phase and feature combination of larch plantation.The experimental results showed that the optimal window size was 9×9 by using gray co-occurrence matrix (GLCM)to classify texture features under different windows.The importance of all features was evaluated based on the Gini coefficient,and the highest overall accuracy was selected as the preferred subset.When 84% of all features (spectral,texture,exponential and topographic features were 11,5,9 and 2 respectively)were used for classification,the overall accuracy reached the highest 82.67% (Kappa coefficient was 0.76).The contribution rate of the vegetation index was the highest among all the characteristics.Compared with spectral features,spectral features+ vegetation index,spectral features + texture features and spectral features + terrain factors classification,constructing multi-feature optimization RF classification model can effectively reduce the dimension and improve the classification accuracy of larch plantationforest.

Key words: larch plantation forest, texture features, random forest, feature preference