| Forest resources have a significant impact on ecological balance,climate change,biodiversity,human production and life,etc.Accurately grasping the distribution and scope of forest resources has an important guiding role for the country to further formulate relevant economic and environmental benefit policies.With the development of major projects of high-resolution earth observation systems in China,the use of highresolution remote sensing images for natural resource management has gradually become one of the main techniques.However,high-resolution remote sensing that contain rich spatial detail information and structural information,at the same time,have more complex spectral characteristics.Changes in spectral heterogeneity,differences between categories and within categories,which significantly causes difficulties in high-resolution remote sensing images processing and information extraction.When traditional machine learning algorithms based on low-level feature learning process high-resolution remote sensing,due to the reason of classifier,in the face of abundant detailed information,it is difficult to accurately express abstractly,which makes classification effect and classification accuracy significantly reduced.How to use the rich information contained in high-resolution images accurately and efficiently becomes the key to extracting information from remote sensing images.Under natural conditions,the distribution of forest types is relatively complex.There are different types of forest in the same area,and the spectral characteristics presented in highresolution images will be more differentiated.Therefore,aiming at resolving the practical problem of extracting forest types from high-resolution remote sensing images,this article proposed a classification method combining stratified classification strategy and feature learning method based on an object based convolution neural network for forest type determination.The work and results of this article are as follows:(1)Different from the total factor land cover classification,this article proposed a stratified classification strategy from the perspective of the natural attributes of the forest to extract forest information.At first,the vegetation area was extracted from the study area,then,according to the classification method,the forest identification of the vegetation area was realized.The experimental results show that the stratified classification strategy not only avoids the complexity of scale selection to some extent,but also improves the overall efficiency of forest extraction.(2)Based on the influence of salt and pepper noise brought by the pixel classification,this paper combined object-oriented method and CNN with multi-layer feature learning ability to identify forest information.The object-oriented image analysis method can not only avoid the salt and pepper noise in pixel classification,but also avoid the spectral isomerism in high-resolution remote sensing image.In addition,CNN can mine high-level abstract feature information high-resolution remote sensing images,the combination of them can further improve the accuracy of forest types classification.The overall extraction accuracy of the proposed method was 0.92.At the same time,the experiment is compared with different segmentation methods(superpixel segmentation method-SEEDS)and the traditional machine learning method random forest.The results show that the overall accuracy of the method in this paper is about 1% higher than the SEEDS-CNN method,compared with the random forest method,an increase of about 10%,and the extracted forest result is more complete,less “salt and pepper” phenomenon.In addition,the method proposed in this paper is more robust to the extraction of forest subclasses of high-resolution remote sensing images. |
PDF Full Text Request