Feature Optimization Combined With UPerNet-ConvNeXt For Eucalyptus Extraction From Remote Sensing Images

Eucalyptus is an important forest tree species,with fast growth,good adaptability and high yield characteristics,can be used in paper,oil refining,construction,furniture and other multipurpose.However,the unscientific planting of eucalyptus has some negative effects on the ecological environment,such as the occupation of cultivated land and basic farmland planting,leading to the destruction of cultivated land and basic farmland.Therefore,rapid and accurate grasp of eucalyptus distribution information is particularly important for scientific planting and eucalyptus management.Although using remote sensing images to extract eucalyptus has become an important technical means,there is still room for further improvement in selecting and combining eucalyptus extraction features,as well as combining emerging deep learning models.Therefore,this paper takes the junction area of Xixiangtang District and Long ’an County in Nanning,Guangxi as the research area.Based on the random forest Gini index and OOB accuracy,the characteristics of Sentinel-2A and OHS images in the preprocessed research area are optimized.The UPerNet-ConvNeXt deep semantic segmentation model was combined with a new combination in this paper to extract eucalyptus.Meanwhile,the UPerNet deep semantic segmentation model of FCN,Unet,DeepLabV3 and UPerNet were compared with other feature combination schemes.The results show that:(1)Feature optimization variables with optimal extraction accuracy and effect of eucalyptus,Sentinel-2A images were b3,b2,b4,NDeI,NDVI,b12,RVI and b5.OHS images were B8,B7,B6,B10,B11,B9,B12,B16,B5,B13,B14,NDeI,RVI,NDVI and B15.(2)Eucalyptus index NDeI,which was created based on spectral reflectance information combined with eucalyptus characteristics,showed a good effect and could effectively improve the accuracy of eucalyptus extraction.This conclusion was valid in both remote sensing images.After addition of NDeI,F1 scores and IoU of UPerNet-ConvNeXt network model were increased by 0.24% and 0.40% in semantic segmentation Sentinel-2A images,while F1 scores and IoU of OHS images were increased by 0.30% and 0.51% respectively.(3)Eucalyptus extraction based on the preferred features can effectively shorten the model training time and improve the semantic segmentation accuracy of eucalyptus extraction,and this conclusion is valid in both remote sensing images.The F1 scores of UPerNet-ConvNeXt model extraction and semantic segmentation Sentinel-2A images increased by 0.77% to 91.96%,the IoU increased by 1.3% to 85.11%,and the training time decreased by 24 minutes,while the F1 scores of OHS images increased by 0.67% to 91.24%.IoU increased 1.15% to 83.88%,and training time decreased 2 hours and 39 minutes.(4)The UPerNet-ConvNeXt model in this paper could effectively improve the semantic segmentation effect of eucalyptus extraction,and the conclusion was valid in both remote sensing images.Compared with the original UPerNet model,such as FCN,Unet,DeepLabV3 and UPerNet,under the same feature combination scheme,the semantic segmentation effect of UPerNet-ConvNeXt model in this paper was more significant,with fewer missegmentation and missing segmentation,and more detailed segmentation edges.The F1 fraction and IoU of eucalyptus extraction accuracy in descending order were UPerNet-ConvNeXt > Unet >UPerNet > FCN > DeepLabV3.