摘要
青藏高原地区的湖泊对当地环境、生态、人文的影响巨大,从遥感高分影像上准确、迅速提取大型湖泊是开展相关研究的重要基础。遥感影像的湖泊提取方法已从传统人工目视解译向自动化发展,但NDWI等常用水体提取技术在应用于不同地域对象时精度不一,特别在提取高原大型湖泊时精度仍有待提高。针对这一问题,设计了一种Res-Unet相结合的水体提取深度学习神经网络,选取青藏高原地区的Landsat高分影像作为训练数据,利用训练所得模型对青藏高原的色林错湖进行提取并分析其20年内的面积变化情况。结果表明:(1)基于Res-Unet的神经网络提取湖泊面积的mIoU为0.9279,Kappa系数为0.9252,远高于NDWI的提取精度,适用于色林错湖面积提取;(2)2000—2020年,该湖泊总面积增长了468.70 km^(2),与青藏高原北部的各拉丹东冰川消融带来的径流补给增加存在联系。
Lakes in the Qinghai-Tibet Plateau have a huge impact on the local environment,ecology,and humanities.Accurate and rapid extraction of large lakes from high-resolution remote sensing images is an important basis for related research.The lake extraction method of remote sensing images has developed from traditional manual visual interpretation to automation.However,common water extraction techniques such as NDWI have different accuracy when applied to different geographical objects,especially when extracting large lakes in the plateau.The accuracy still needs to be improved.In response to this problem,this paper designs a deep learning neural network for water body extraction combined with Res-Unet,selects the Landsat high-resolution images in the Qinghai-Tibet Plateau as training data,and uses the training model to extract the Selincuo Lake in the Qinghai-Tibet Plateau.And analyze its area change in 20 years.The experimental results show that(1)the neural network based on Res-Unet has an mIoU of 0.9279 and a Kappa coefficient of 0.9252,which is much higher than the extraction accuracy of NDWI,and is suitable for the extraction of Selincuo Lake area;(2)The total area of the lake increased by 468.70 km^(2)from 2000 to 2020,which is related to the increase in runoff recharge brought about by the melting of the Geladandong glacier in the northern Qinghai-Tibet Plateau.
作者
吴渊
陈剑
陈鹏
WU Yuan;CHEN Jian;CHEN Peng(Tongji University,Shanghai 200092,China;Southwest Jiaotong University,Chengdu 611756,China)
出处
《环境生态学》
2022年第11期6-12,共7页
Environmental Ecology
