Object-Oriented Forest Stands Classification Based On Deep Learning And Remote Sensing Data Collaboration

Forest stands are the basic units of forest resource survey,and their distribution is one of the key factors in forest survey.Forest stands play a vital role in forest resource planning,forest resource management,ecosystem construction and economic sustainable development.Remote sensing technology can break through the limitations of traditional artificial ground mapping,and is more and more widely used in the dynamic management of forest resources.Precise identification of forest stands is one of the main tasks of forest investigation in remote sensing study.Limited to the spatial resolution of remote sensing data,the current forest stands identification commonly use data with a resolution of less than 10 meters,such as Landsat multispectral images,Sentinel-2 multispectral images,etc.In addition to the above-mentioned medium-resolution images,multi-spectral data within a resolution of 10 meters are also becoming more and more popular,which contain richer information about forest stands（categories,distribution,growth）.Moreover,remote sensing data such as LiDAR and Synthetic Aperture Radar（SAR）also can provide more information about forest stands（type,distribution,structure）.Remote sensing data of different types and different band settings can provide complementary spectral and spatial details.For example,high spatial resolution multispectral images can provide fine forest structure information.The red edge band,which has been gradually popularized in recent years,can provide more fine vegetation category information.LiDAR can provide information such as the three-dimensional shape and structure of forest stands.This study uses RapidEye,Sentinel-2,high-resolution aerial image and LiDAR point cloud data,employs forest stands delineation,convolution neural network whice based on object-oriented classification and other techniques to improve the accuracy of stands classification.The purpose of the study is explore the complementarity of medium-resolution and high-resolution multi-source remote sensing data in forest stands identification.The main research and results of the study are as follows:（1）Making use of the Multivariable Analysis（MV）fusion method,the RapidEye image was sequentially merged with the two 20-meter resolution red-edge bands of Sentinel-2 to improve the spectral information of the RapidEye image;and further merged with the 2.5 meter aerial image to increase the texture information of RapidEye image;and finally the height information provided by the LiDAR point cloud was added to them.The results show that the accuracy of classification after data collaboration has been greatly improved the classification result.（2）The One-Dimensional Spectral Difference（ODSD）multi-scale segmentation method is used to delineate forest stands,clarify forest stand boundaries,and lay the foundation for subsequent fine forest stands classification.（3）The object-oriented method and convolutional neural network（CNN）technology are used to classify the medium-resolution and high-resolution multi-spectral images.Among several CNN structures,ResNet₁8 provides the highest classification accuracy,numerically exceed 4.53%,5.49%,5.97%,4.53%and 5.72%respectively in 5m RapidEye,5m RapidEye plus RE2,5m RapidEye plus RE2 plus RE4,and 2.5m RapidEye than support vector machines which acquired the highest accuracy among several traditional classification methods.The above results show that the collaboration of medium-resolution and high-resolution multi-source remote sensing data can effectively improve the recognition accuracy of forest stands,and forest stands delineation is of great significance.At the same time,the combination of object-oriented method and convolutional neural network technology has obvious feasibility and superiority in forest stands classification.