Forest Classification Of Remote Sensing Image Based On Deep Learning ResNet Model

As one of the most important indicators of forest resources monitoring,the classification of tree species has important research value in forest resources investigation.The traditional classification method of tree species in remote sensing image adopts artificial feature extraction,which was time-consuming and laborious,and the recognition effect was poor.Considering the characteristics of remote sensing images and the advantages of convolutional neural network in image classification,the combination of the two can improve the classification efficiency and accuracy of tree species.In this paper,we propose a semantic segmentation model for tree species classification of U-Net remote sensing images based on the optimization of ResNet50 structure of residual network in convolutional neural network,and used the transfer learning method to retain the parameters of large data sets for network model training.Both methods could effectively solve the problem of insufficient learning depth of small samples.At the same time,the efficiency of the experimental process and the precision of the experimental optimization model were improved.Finally,the tree groups of six species in the study area were extracted by using the ResNet50&U-Net model constructed in this paper.The conclusions are as follows:(1)The Resnet50&U-Net model in this paper was trained in data B432,the results are as follows:the pixel accuracy of the overall classification is 93.46%.The Class Pixel Accuracies(CPA)of different classes are as follows:the pixel accuracy of pine and cypress is 80.03%,the pixel accuracyof flowers,seedlings and shrubs is 78.87%,the pixel accuracy of economic trees is 60.39%,the pixel accuracy of broad-leaved trees is 81.23%,the pixel accuracies of Chinese fir is 84.78%,the pixel accuracies of bamboo forests is 81.85%,and the pixel accuracies of other classes is 89.55%.In this data,the Mean Pixel Accuracy(MPA)is 82.09%,and the Mean Intersection over Union(MIoU)is 72.79%,The overall pixel accuracy of B843 is 93.71%,and the Class Pixel Accuracy(CPA)is 75.24%for broadleaf,84.16%for pine and cypress,84.33%for Cunninghamia lanceolata,63.00%for economic tree,83.37%for bamboo forest,79.21%for flowers,seedlings and shrubs,and for other classes are 88.97%.The Mean Pixel Accuracy(MPA)is 82.28%,and the MIoU is 73.26%.There is little difference in the accuracy between the data of different band combinations,which indicates the stability of the model in this paper.Compared with other methods,the accuracy is relatively great,which shows that the ResNet50&U-Net model in this paper has certain advanced nature and feasibility.(2)Using the ResNet50&U-Net model to calculate the area of Xinshao County:in the data B432 prediction results,the area of broad-leaved group is 1.6400×10~2km~2,the area of pine and cypress group is 3.7566×10~2km~2,the area of Chinese fir group is 1.8749×10~2km~2,the area of economic tree group is 0.2136×10~2km~2,the area of bamboo group is 1.9883×10~2km~2,the area of flower seedling and shrub group is 0.4834×10~2km~2,and the area of other groups is 7.6538×10~2km~2,In the data B843 prediction results,the area of broad-leaved group is 1.4036×10~2km~2,that of conifer group is 4.0547×10~2km~2,that of Cunninghamia lanceolata group is 1.8668×10~2km~2,that of economic tree group is 0.2196×10~2km~2,that of bamboo group is 2.0603×10~2km~2,that of flower seedling and shrub group is 0.493 7×10~2km~2,and that of other groups is 7.5118×10~2km~2.The results are close to the actual area of forest resources in Xinshao County,which verifies the feasibility of ResNet50&U-Net model.(3)Resnet50&U-Net model predicted the area of each tree groups in 2017 in Xinshao County,and then compared with the area of each tree groups in 2020.The total area of forest in 2017 is 10.1504×10~2km~2,while the total area of forest in 2020 is 9.9568×10~2km~2.The forest area was reduced to 0.1936×10~2km~2 in 2020.The area of broad-leaved group increased by 0.1205×10~2km~2,that of pine and cypress group remained unchanged,that of Cunninghamia lanceolata group decreased by 0.0309×10~2km~2,that of economic tree group decreased by 0.0863×10~2km~2,that of bamboo forest group decreased by 0.1564×10~2km~2,and that of flower seedling and shrub group decreased by 0.0405×10~2km~2.It is proved to be feasible in the application of forestry resource information updating.