摘要
利用卫星遥感技术融合深度学习算法,可以快速、动态、高效识别露天煤矿开采区,以我国和其他煤炭资源大国的典型露天煤矿开采区为研究对象,基于高分二号多光谱遥感影像,制作数据集及标签,构建基于卷积神经网络(convolutional neural networks,CNN)的深度学习目标检测算法.通过加入特征金字塔网络,充分挖掘开采区及背景区的低分辨率语义信息和高分辨率纹理信息,实现快速卷积神经网络的深度学习目标检测算法模型的改进及参数优化.结果表明改进后的模型平均检测精度提高到98.48%,总体识别精度达到96.7%,有效提高了复杂背景下的多尺度、多类型露天开采目标的识别精度,为全球煤炭资源大国能源合作、生态环境保护及我国矿产资源的合理利用和修复提供了科学、精准手段.
Using satellite remote sensing technology coupled with deep learning algorithms could characterize open-pit coal mining area dynamically and efficiently.This study focused on typical open-pit coal mining areas from China and other major coal-resource countries as research objects.Based on GF-2 multi-spectral remote sensing images,data sets and labels are produced to construct faster-regions with convolutional neural networks(CNN) features target recognition model.Low-resolution semantic layer and high-resolution texture information of mining area and background area were fully mined by incorporating the feature Pyramid network,which could improve faster-regions with CNN features model by optimizing parameters.The results showed that the average detection accuracy was improved to 98.48%,and the overall recognition accuracy reached 96.7%.The improved faster-regions with CNN features model efficiently increases the identification accuracy of multi-scale and multi-type open-pit mining targets in complex backgrounds.These findings can provide a scientific and accurate technical way for global energy cooperation,environmental protection,and mineral resources utilization in China.
作者
包妮沙
韩子松
于嘉欣
韦丽红
BAO Ni-sha;HAN Zi-song;YU Jia-xin;WEI Li-hong(School of Resources&Civil Engineering,Northeastern University,Shenyang 110819,China;Changguang Satellite Technology Co.,Ltd,Changchun 130000,China;Hulun Buir College,Hulun Buir 021008,China)
出处
《东北大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2023年第12期1759-1768,共10页
Journal of Northeastern University(Natural Science)
基金
国家自然科学基金资助项目(52074063,U1903216)
内蒙古自治区高等学校科学技术研究自然科学项目(NJZY22278)。
关键词
高分影像
露天开采区
目标检测
特征金字塔
domestic high-resolution imagery
open-pit coal mining areas
object detection
feature pyramids
