Research On Tree Species Classification Using Airborne Hyperspectral Images Based On Deep Learnin

Tree species identification is one of the important elements of forest resource monitoring.Timely and accurate acquisition of tree species information is the basis for sustainable forest management and resource assessment.Traditional tree species identification mainly relies on manual fieldwork,which is time-consuming and laborious.The rapid development of remote sensing technology makes up for the deficiency of manual survey methods.Compared with other remote sensing data sources,airborne hyperspectral images have rich spectral and spatial information,which can detect the subtle differences between different tree species and realize high-precision tree species identification.Traditional hyperspectral image classification algorithms（e.g.,Support Vector Machine,Random Forest）require feature selection in advance,which can cause information loss and reduce classification accuracy.Deep learning provides an end-to-end learning paradigm that can autonomously learn complex nonlinear features from large amounts of data,avoiding the complexity and limitations of feature extraction by traditional algorithms.In this paper,airborne hyperspectral images and deep learning methods were used to classify tree species in three study areas.The main research contents and conclusions are as follows:（1）In order to explore the difference of tree species classification results based on pixel level and individual tree level,this study first improved the feature extraction part of Mask R-CNN network and add attention mechanism to Bottleneck block.It achieved the segmentation of individual tree in high-resolution images,and provided individual tree vector map for subsequent tree species classification based on individual tree level.The results showed that when the input data of the network was RGB information and the Sim AM attention mechanism was added to the network,the highest segmentation accuracy was obtained,with AP value reaching41.5%and AP₅₀value reaching 74.5%.（2）A joint spatial-spectral network based on double branch and SimAM attention mechanism was proposed for hyperspectral tree species classification.The traditional machine learning methods and deep learning methods were used to classify tree species in the three study areas of TEF,Tiegang Reservoir and Xiongan New Area respectively,and the classification performance of each classifier was evaluated.The results showed that the method proposed in this study achieved the highest classification accuracy in the three datasets.（3）The classification of tree species is discussed and analyzed from different perspectives.The results show that the classification performance of tree species based on individual tree level is better than that based on pixel level.The combined use of spatial-spectral information and the addition of attention mechanism are helpful to improve the classification performance of the network.The network proposed in this paper achieves the highest accuracy under different proportions of training samples and does not change greatly with different proportions of training samples,providing a solution for tree species classification of small sample data in forest areas.