Abstract:

Objective: To explore the effective network construction method of three-dimensional convolutional neural network (3D-CNN) in tree species classification supported by hyperspectral data. Method: Taking the southern Sierra Nevada, California, USA as the study area, the canopy height model (CHM) obtained from LiDAR data was divided into single trees and used as a supplement to establish samples. A 3D-CNN network structure with simpler structure, higher classification accuracy and no need to preprocess hyperspectral data was improved for forest species identification. Result: Compared with the conventional supervised classification methods(support vector machine, random forest), the traditional two-dimensional convolutional neural network model and the latest MSR 3D-CNN model, the overall classification accuracy of the 3D-CNN model proposed in this study is 99.79%, and the mean intersection over union(MIoU) is 99.53%. Compared with SVM and RF method, the overall classification accuracy is improved by about 5%, and the new 3D-CNN model has the characteristics of more accurate extraction of tree species' boundaries and less pepper and salt phenomenon; Compared with 2D-CNN, the overall classification accuracy is improved by about 10%, and MIoU is improved by about 7%; Compared with MSR 3D-CNN, the overall accuracy is not much different, but in the process of training and testing, this model takes much less time than MSR 3D-CNN model. Conclusion: The 3D-CNN model proposed in this study can efficiently process the original hyperspectral images, classify and map tree species, and add forest vertical structure information to make classification labels, which can obtain higher accurate classification results.

Key words: hyperspectral, LiDAR, convolutional neural network, tree species classification, 3D-CNN