The Forest Classification Combining Multidimensional Features Based On High-resolution Remote Sensing Images

The high resolution image has abundant texture and shape features, it can effectively improve the classification accuracy. Since the high resolution image includes less spectral information generally, to some extent, the advantage of forest classification could be weakened. Thus, to improve the classification accuracy of forest species, in this paper, the WorldView-2 high spatial resolution image with eight spectral bands was used as the research data, and the Jiangle Forest Farm in Fujian Province was selected as the study area.Firstly, multi-level segmentation was conducted based on the WorldView-2 image. The optimal segmentation scales for different target objects were determined based on the result of visual discriminant segmentation and subcompartment vector map matching. On this foundation, combined with the actual characteristics of different features to.establish three optimal scaling layers of the groundcover information extraction, respectively to extract road types; woodland types, construction land, bare land, pond; river. Secondly, the curves of spectral, vegetation index and texture features were analyzed under different of groundcover objects, and the shape features were extracted based on the mathematical morphological filtering. Then, the feature values with great difference among the classes were selected to establish the feature space and rules of classification. In addition, multidimensional features were established by using bi-temporal images, to analyze the groundcover feature values in different temporal phase. And the differences of vegetation feature were enhanced by the comparison of different seasonal phenology, to ensure the feature values of groundcover with strong divisibility in different temporal phase for the rules of classification. Finally, the bi-temporal and single-temporal multi-level classification single-temporal single-level decision tree classification based on the object-oriented method were used, with single-temporal pixel-based random forest classification, to realize the classification of forest groundcover, especially the high precision classification of vegetation species.The results showed that using texture and multi-temporal features on woodland types, shape features which based on the morphological filtering on construction land, road and water body types, and combining spectral, vegetation index features on all land types, can effectively improve the classification accuracy. The total accuracy of bi-temporal object-oriented multi-level classification and Kappa coefficient were 92.41% and 0.9141 respectively, which were higher than the accuracy of the other three results and satisfied the forest mapping and other applications. The object-oriented multi-level classification method could effectively overcome the phenomenon of "same objects with different spectrums" and "same spectrums with different objects" in pixel-based classification, and reduce the probability of wrong classification and contour information of various types of objects, additionally, strengthen the ability of resisting noise and avoid the phenomenon of "salt and pepper".