In response to the issues of"salt-and-pepper phenomenon,"low degree of classification precision,and efficiency in wetland information extraction,this paper proposes an M&E-DeepLab network model,an improvement based on DeepLab V3+.Firstly,the model utilises MobileNet V2 as the backbone feature extraction network of DeepLabV3+,to reduce the model's parameter count,and enhance the network's training efficiency and accuracy.Secondly,it performs layer-by-layer feature transmission among the three parallel dilated convolution branches in the Atrous Spatial Pyramid Pooling(ASPP)module,to expand the receptive field and improve the efficiency of information utilisation.Lastly,a channel attention mechanism is introduced after the ASPP module to strengthen the channel features of deep feature maps,thereby enhancing the network's segmentation performance.The results show that the overall accuracy of the model is 90.0％,with a Kappa coefficient of 0.878.Compared to the original DeepLab V3+and other related models,this model demonstrates significant advantages in the extraction accuracy of various types of land such as Suaeda pterantha,Phragmites australis,and mixed wetlands.It has a promising application prospect in the extraction of information from the Liaohe wetland.
wetlands information extractLiaohe river wetlandDeepLab V3+semantic segmentationattention mechanismPanjin
NETL
NSTL
万方数据
