基于深度学习的积雪覆盖区山地冰川识别研究
Research on Identification of Snow-Covered Mountain Glacier based on Deep Learning
王晶晶 1柯长青 1陈军2
作者信息
- 1. 南京大学 地理与海洋科学学院,江苏 南京 210023
- 2. 安徽建筑大学 环境与能源工程学院,安徽 合肥 230601
- 折叠
摘要
全球变暖导致冰川急剧退缩,及时的冰川监测和制图至关重要,而积雪覆盖一直是冰川识别的重要影响因素.以喀喇昆仑区域为例,选择春季Landsat-8 OLI、Sentinel-1和DEM数据,结合其光谱反射率、SAR散射以及地形等特征,基于不同主干网络的U-Net和DeepLabv3+深度学习方法,使用不同样本尺寸,不同特征组合进行冰川识别对比研究.结果表明:①对于 256×256、512×512和1 024×1 024像素样本尺寸,训练样本尺寸越大,空间上下文信息越丰富,识别精度越高,冰川末端范围更为精确.②基于MobileNet、VGGNet、ResNet以及EfficientNet主干提取网络的U-Net语义分割网络中,VGG19主干网络识别精度最好,且优于DeepLabv3+网络结果,其F1值(F1-Score)为0.899 6,均交并比(Mean Intersection over Union,mIoU)为0.875 4,总体精度可达0.948 4,在山体阴影、冰雪融水、薄雾覆盖和冰冻湖泊区域识别效果均较好.③随着训练特征数量的减少,精度随之降低,地形特征对于提高冰川识别精确度作用显著,SAR特征则可提升召回率.研究证明了深度学习方法识别积雪覆盖的山地冰川的可行性,为山地冰川快速大面积识别的模型选择和参数设置提供了可靠的参考依据.
Abstract
Global warming results that glaciers retreat rapidly.Monitoring and mapping glacier boundary are ex-tremely significant for research on global climate change and predicting related disasters.However,snow cover-ing is the main barrier all the time.Selecting Karakoram subregion as study area,the Landsat 8 OLI,and Senit-nel-1 images and DEM data in spring(March 24th,2019)were utilized.The spectral reflectance of green,red,near-infrared and short-wave infrared bands in Landsat 8 OLI images were selected as the optical image features.The backscattering coefficient of VH polarization channel,the coherence coefficient of VV polarization channel,local incident angle,polarization entropy H and scattering Angle α after polarization decomposition were gained from SAR data and used as SAR features.Topographic features included DEM and slope.These characters were employed as input of models.First,based on U-Net model,experiments compared the accura-cies using different-size samples.The 256×256-pixel-size samples were imported to U-Net network model based on different backbone networks(MobileNetv2,VGGNet,ResNet and EfficientNet)and DeepLabv3+ model.Finally,the best one among the above networks was employed to import samples with different feature combinations.Results show:①Using the bigger training sample with the richer spatial context information can obtain the higher segmentation accuracy and the glacier terminal boundary is more accurate.②Among the differ-ent backbone networks,VGG19 backbone network exhibits the highest accuracy,which is higher than that of DeepLabv3+.Its F1-value is 0.899 6,and the mean intersection over union(mIoU)is 0.875 4,and the overall accuracy is 0.948 4.The recognition effect of shadow,snow melt-water,mist covering and frozen lake area is comparatively good.③With the decrease in the number of training features,the accuracy also drops.Topo-graphic features can improve the precision rate,while SAR features can increase the recall rate by 4%or so.This study proves the feasibility of the deep learning methods on the identification of mountain glaciers covered by a large amount of snow and provides reliable basis on model selection and parameters setting for rapid and large-scale mountain glaciers mapping.
关键词
山地冰川/深度学习/主干网络/积雪覆盖/喀喇昆仑Key words
Mountain Glacier/Backbone Network/Deep Learning/Snow cover/Karakoram引用本文复制引用
基金项目
国家自然科学基金重点项目(41830105)
青年科学基金(41901129)
出版年
2023
NETL
NSTL
维普
万方数据