| Tree species information is an important element of forestry resource survey and monitoring,and is an important indicator for describing the ecological value of forests.With the rapid development of hyperspectral remote sensing technology,it provides the possibility to realize the fine identification of tree species.In this paper,we use satellite hyperspectral and UAV hyperspectral images as data respectively,and improve the application model Deeplab V3+ of the null convolution algorithm to explore its effect on tree species identification in hyperspectral remote sensing images,so as to provide some ideas and basis for the application of hyperspectral data in tree species identification.The tree species in the “Millennium Forest” semi-artificial forest ecological zone of Xiong’an New Area were studied by using the principal component analysis method to remove the redundant bands in the images,adding the attention mechanism,feature fusion,replacing the trunk feature extraction network and improving the hole convolution in the Deeplab V3+model.The accuracy of the improved algorithm was evaluated using overall accuracy,m Io U value,Kappa coefficient and confusion matrix,and compared with support vector machine,random forest,FCN,U-net and Deeplab V3+ algorithms.When the UAV hyperspectral images were the data,the tree species in the Mahoiwan Village area of Xiongan New Area were used as the research object,replacing different attention mechanisms,feature fusion and adopting a denser feature pyramid structure based on the improvement of the former algorithm model.Principal component analysis and the best index method were used to filter the images by bands,respectively,to select the optimal band combination and fully exploit the potential of the algorithm model.The overall accuracy,Kappa coefficient and confusion matrix evaluation metrics were applied to compare the tree classification results with the random forest algorithm incorporating the feature information of eight texture features.The experimental results show that when using satellite hyperspectral imagery as data,the optimised algorithm model tree classification accuracy(58.64%)is higher than that of support vector machines(53.69%)and random forest models(48.32%),and the segmentation accuracy m Io U value of the model(0.3407)is higher than that of FCN(0.2758)and U-net algorithms(0.2774),and the training time is at least 2.3 hours.The highest accuracy of the optimised algorithm model for tree species classification(94.28%)was higher than that of the random forest algorithm(86.94%)when the UAV hyperspectral imagery was the data,and the classification accuracy was higher than that of the random forest for heavily mixed species,e.g.99.54% for acacia users and 88.90% and 90.13% for pear and compound maple respectively.In summary,the optimisation algorithm model has advantages in classification under small sample conditions.The application of null convolution with different null rates can effectively extract features with fragmented distribution and rich detail information,and the incorporation of attention mechanism and feature fusion methods can significantly improve the classification accuracy.The algorithm model has a certain degree of migration and has good performance in tree species classification recognition for hyperspectral images of different sources. |
