Research On Multi-source Forest Data Classification Based On Double-branch Symmetric Network

With the advancement of science and technology,diversified data modalities provide more research directions for forestland remote sensing applications.Commonly used remote sensing images,such as hyperspectral images,contain the spatial dimension and spectral dimension information of category targets,which can effectively classify land.The lidar image has three-dimensional coordinate information,which is conducive to.the identification of the horizontal and vertical structure of the forest.The collaborative classification of multi-source forest land data synthesizes diverse forest information to obtain more accurate classification performance.Existing research often requires high technical skills of the staff due to feature selection.And because of the characteristics of mixed pixels in remote sensing images,it is difficult to extract the optimal feature expression.Focusing on the information extraction between multiple sources of data and the collaboration of the advantages of each source in the process of fusion,borrowing the powerful feature learning capabilities of deep learning methods to carry out fine classification of forest land,the main contributions are as follows:1.Aiming at the data set of pure forest category such as the Belgian forest land data,the feature extraction work of complex tree species categories is studied,and a two-branch symmetrical forest land fusion classification model is proposed.First of all,according to the data characteristics of the Belgian dataset,a preprocessing method for data intensity control is designed.This method can reduce the impact of excessive pixel differences in different bands in the multi-source dataset on the classification performance.Through the preliminary screening and processing of the data,the data intensity of the multi-source data is controlled to be consistent.A short-layer network structure is further designed,which can simply and effectively extract the characteristic information of each source data.The double-branch symmetrical information fusion network is a reasonable combination of short-layer networks to focus on reducing the mutual interference between spectrum and spatial features.It creatively constructs a special spectrum sensing tunnel for each data source,which can effectively combine the respective feature advantages of multi-source data.The similar but slightly different network structure ensures that the multi-source data performs the feature extraction of each source data.Functional consistency.2.Aiming at the classification of forest trees in a multi-scenario environment,research on multi-scenario remote sensing data sets that contain multiple categories in addition to forest tree categories.Based on the forestland fusion classification network,a cross-branch fusion classification network model is proposed,and a hierarchical cross-branch multi-source connection framework with a symmetrical structure is established.The overall symmetrical structure,combined with the specific connection between different layers of multi-source data,can equally represent or measure different feature sources,so as to distribute the information obtained from each data source proportionally,and further help the network balance the amount of information and adjust the direction of feature learning.