Extraction Of Larch Plantations Using Texture Features Within High Spatial Resolution Images

Larch is one of the most important planted timber wood in China,and its planting area ranks in the top ten of the planted arbor species.Larch,which are widely planted,has produced enormous economic benefits and satisfied the demands of timber resources and ecology.However,the spatial distribution of larch plantations is unclear because of its extensive area and long-period.It is critical for forest management,monitoring programs and ecosystem function assessment to accurately obtain the spatial distribution of larch plantations by remote sensing technology.In this paper,larch plantations in Mengjiagang Forest Farm of Jiamusi,Heilongjiang Province and Dagujia Forest Farm of Qingyuan Manchu Autonomous,Fushun,Liaoning Province were studied.Based on GF-1 and GF-2 remote sensing images,the extraction of larch plantations was carried out from different phase and different spatial scale.In the Mengjiagang Forest Farm,this paper discussed the influence of texture feature extraction method and parameter selection on classification accuracy.And then the effects of multi-temporal and multi-feature synthesis and feature selection on classification accuracy were studied to determine the appropriate remote sensing identification features of larch plantations.In the Dagujia Forest Farm,the effects of spectral and texture features of remote sensing images with different spatial resolution on the extraction of larch plantations were studied to determine the appropriate spatial resolution.The main conclusions are as follow:(1)Extraction of larch based on single-phase GF-1 image texture feature.The parameters of Gray Level Co-occurrence Matrix texture have a great influence on classification accuracy,among which the window size has the most significant effext on texture features,the direction is the second,and the pixel distance is the last.Through Maximum Likelihood classification experiments of texture features in different directions and different windows,the results show that the optimal window size is 15×15 and the optimal direction is 90°.By analyzing the variation of texture feature values with pixel distance,it is found that when the pixel distance of mean,variance,homogeneity,dissimilarity,entropy and second moment is 1,the pixel distance of contrast is 2,and the pixel distance of correlation is 3,larch plantations are most divisible from other objects.There is no significant different in improving classification accuracy among Uniform Local Binary Patterns texture with differet radius.Uniform-LBP texture feature with radius 1 and sampling point 8 can not only improve the classification accuracy,but also improve the calculation efficiency.The combination of GLCM texture features and Uniform-LBP texture feature can improve the classification effect of single-phase image.Compared with the spectral featrues classification results,the overall accuracy and the extraction accuracy of larch plantations have been improved by 3.35%and 2.67%respectively.(2)Extraction of larch plantations based on multi-temporal GF-1 images.The spectral features,texture feature and vegetation indices of multi-phase images,terrain features and texture feature difference rate were used to construct feature space for extraction of larch plantations,and then Random Forest was used for ranking the feature importance and classification.The results show that the data dimension can be reduced,the appropriate feature set can be determined and the classification accuracy can be improved by feature selection.In the remote sensing information extraction of larch plantations in Mengjiagang,the overall accuracy and extraction accuracy of the suitable festure subset reached 89.25%and 91.08%respectively.Different types of features have different importance,vegetation indices are relatively important,followed by spectral features and texture features.By comparing the importance of different texture features,it is found that the importance of GLCM texture features is higher than that of Uniform-LBP texture feature,and the contribution of GLCM texture to classification of October images is higher than that of May and July images,among which mean is the most important texture feature.(3)Extraction of larch plantations based on different resolution images.In the Dagujia Forest Farm,Maximum Likelihood,Support Vector Machine and Random Forest were used to extract larch plantations from three classification combination of different resolution images.The results show that 4m is the appropriate spatial resolution for the extraction of larch plantation,and the overall accuracy and extraction accuracy of larch plantations reached 91.85%and 94.2%respectively.Texture features can improve classification accuracy,but the appropriate spatial scale should be selected according to the target objects.The higher the spatial resolution is,the more significant the improvement effect of texture features on classification accuracy is.The classification accuracy of Support Vectore Machine and Random Forest is better than that of Maximum Likelihood Classification in different resolution images,which can meet the precision requirement of remote sensing information extraction of tree species.