摘要
数据集制作是利用深度学习方法进行地质灾害目标自动检测的重要基础。提出一种无人机影像地质灾害目标数据集(disaster event dataset,DED)的制作方法,该数据集包含坍塌房屋、滑坡和泥石流3种典型地质灾害目标,共有16 535个标注对象。使用Faster R-CNN模型进行实验评估以检验DED的有效性,并基于不同深度的预训练网络以及是否使用k-means聚类优化进行对比实验。结果表明,基于DED训练深度网络有较好的效果。
Dataset production is an important basis for the automatic detection of geological disaster targets based on deep learning methods. We propose a method to produce a disaster event dataset(DED)of low-altitude remote sensing images obtained by unmanned aerial vehicle(UAV). The DED consists of three typical geological disaster events,namely collapsed houses,landslides and debris flows,with a total of 16535 annotation objects. Faster R-CNN model for experimental evaluation is used to verify the effectiveness of the DED.And a comparative experiment is conducted based on the pretrained networks of different depths and that whether k-means clustering optimization is used. The results show that our dataset has a good training effect on the deep networks.
作者
詹总谦
黄兰兰
张晓萌
刘异
ZHAN Zongqian;HUANG Lanlan;ZHANG Xiaomeng;LIU Yi(School of Geodesy and Geomatics,Wuhan University,Wuhan 430079,China)
出处
《测绘地理信息》
CSCD
2022年第3期100-107,共8页
Journal of Geomatics
基金
国家重点研发计划(2016YFB0501403)
国家自然科学基金(61871295)。