Vegetation Extraction Method Using UAV Remote Sensing Image

Land greening is an important content of ecological civilization construction,and is an important cornerstone to achieve sustainable development.As an important natural resource for land greening,vegetation is considered to be a sensitive indicator for the ecological environment changes.At the same time,according to its fixed role for carbon,vegetation also play a huge role in the global changes.The vegetation type division is the basis of vegetation extraction research,the early division of vegetation type is mainly carried out by artificial method,the high cost of manpower,material and financial has become the difficulty of vegetation type division.Now,with the development of remote sensing technology and UAV technology,high-resolution remote sensing images are one of the main data sources of vegetation extraction because of their rich texture features and geometric features.How to extract the terrain information quickly and accurately on the high-resolution image becomes the hot spot of computer vision,photogrammetry and GIS research at home and abroad in recent years.At present,high-resolution remote sensing image data acquisition is extremely convenient,and the extraction of terrain information,especially the classification of vegetation and the method of information extraction is lagging behind,which is largely dependent on the traditional classification method.When the traditional pixel-based classification method is applied to high-resolution remote sensing image classification,it can not make full use of its high-resolution and rich geometric texture information,but it will plague the phenomenon of "salt and pepper phenomenon" and reduce its classification accuracy.Therefore,the study of vegetation extraction method for high-resolution remote sensing images is of great significance to the division of vegetation types and the construction of ecological environment.Deep learning is a popular classification algorithm,which can represent the rich and internal information of data,which will have great research potential for vegetation extraction research and remote sensing image processing or image classification.Now,the vegetation extraction method based on remote sensing image is mainly visual interpretation,intelligent extraction,human-computer interaction extraction and so on.The methods of computer automatic extraction mainly include artificial neural network,supervised classification and unsupervised classification,fuzzy mathematics,object-oriented,expert system classification,decision tree and machine classification.These extraction methods have achieved some desired results.However,most of these vegetation extraction methods are classified and extracted based on the statistical characteristics of the classification of the pixels.It is difficult to solve the problems such as heterogeneous dissimilarity,same-spectrum foreign matter and mixed-pixel.The classification accuracy is not high and the separated polygons are cluttered.In this paper,we study the vegetation extraction with UAV images as data sources,eCognition Developer8.7 remote sensing software and Matlab as the experimental platform.Firstly,the optimal segmentation of remote sensing image is carried out,then the spectral-texture information of the vegetation is extracted from the UAV image,Using the FCM algorithm and the rough set theory to reduce the feature information which is extracted.The feature set of the vegetation extraction rule is established according to the reduced feature information,and using the optimal segmentation scale segmentation of UAV images for regular classification and extraction.At the same time,the deep learning method is applied to the vegetation extraction,which solves the problem of the lack of generalization of the rule set established in the rule classification extraction and the lack of the existing research literature on the deep study.The main findings are as follows:(1)Aiming at the problem of optimal scale segmentation in multi-scale segmentation of images,this paper proposed a method of optimal scale selection based on the maximum area,weighted mean variance and weighted mean variance rate.The experimental results show that the method can effectively avoid the subjectivity and inefficiency of the segmentation scale of artificial determination,and improve the classification accuracy and efficiency.(2)Aiming at the feature redundancy of remote sensing image classification and the need to establish rule set for a priori knowledge when extracting vegetation rules,this paper proposed the feature reduction using FCM clustering algorithm and rough set theory to reduce the feature redundancy and realize the eigenvector optimization,and establish the vegetation extraction rule set based on Prior Knowledge.The experimental results show that the rule classification is effective,which can effectively reduce the misclassification rate and reduce the appearance of "salt and pepper phenomenon",and express the true vegetation characteristics information on the image well.(3)The rule set of rule classification extraction is lack of versatility,and the deep learning method is applied to vegetation extraction.For using this window size to select the experimental sample will appear adjacent objects in the same window phenomenon,and resulting in some useless crushing plots and "salt and pepper phenomenon" after classification,this paper proposed a method that combine the optimal scale with the DBN method are used to study the vegetation extraction.The experimental results show that the segmentation method and the optimal separation scale selection method are practical in this paper,which can ensure the integrity of the feature extraction and reduce the over-crushing of the extracted objects;The optimal segmentation scale which realize the reasonable integration of the features and can effectively reduce the occurrence of "salt and pepper phenomenon",all kinds of features have a clear boundary;On the whole,the method is close to the overall classification accuracy of the rule classification,which is higher than the Kappa coefficient of the rule classification method.The experimental results can achieve the excellent quality standard,and the deep learning algorithm is more versatile.