多源数据融合智能识别煤矿山场景特征AI模型  

AI model for intelligent recognition of coal mine scene features through multi-source data fusion

作　　者：王立兵 任予鑫 马昆 王蕾[6] 刘峰[6] 翟文 董霁红[1] WANG Libing;REN Yuxin;MA Kun;WANG Lei;LIU Feng;ZHAI Wen;DONG Jihong(School of Environment and Spatial Informatics,China University of Mining and Technology,Xuzhou 221116,China;Engineering Research Center of Mine Ecological Restoration,Ministry of Education,Xuzhou 221116,China;School of Public policy&Management School of Emergency Management,China University of Mining and Technology,Xuzhou 221116,China;School of Mechatronic Engineering,China University of Mining and Technology,Xuzhou 221116,China;Ningxia Coal Industry Co.,Ltd.,China Energy Investment Grop,Yinchuan 750000,China;China Coal Society,Beijing 100013,China;Department of Strategic Planning,China Energy Investment Group,Beijing 100011,China)

机构地区：[1]中国矿业大学环境与测绘学院,江苏徐州221116 [2]矿山生态修复教育部工程研究中心,江苏徐州221116 [3]中国矿业大学公共管理学院,江苏徐州221116 [4]中国矿业大学机电工程学院,江苏徐州221116 [5]国家能源集团宁夏煤业有限责任公司,宁夏银川750000 [6]中国煤炭学会,北京100013 [7]国家能源集团战略规划部,北京100011

出　　处：《煤炭学报》2023年第12期4617-4631,共15页Journal of China Coal Society

基　　金：宁夏煤业有限责任公司金家渠煤矿企业资助项目;国家自然科学基金资助项目(52061135111);山东省煤田地质局重点科研专项(鲁煤地科字(2022)14号)资助项目。

摘　　要：矿山场景数据是智慧矿山建设和智能管理的基础数据,如何利用包括遥感影像在内的多源数据快速识别和提取出复杂的矿山场景是重要的研究方向。采用2020年Sentinel-2影像、GF-6影像、GF-2影像进行最优数据集筛选,使用2023年谷歌影像(Google image)数据扩充数据集,并与深度学习算法相结合,建立了2种露天煤矿场地识别模型。研究主要结论:(1)利用10 m Sentinel-2影像、8 m GF-6原始影像、2 m GF-6融合影像、3.2 m GF-2原始影像、0.8 m GF-2融合影像建立矿山识别模型,量化选择不同数据产生的模型精度。结果显示,遥感图像空间分辨率从10 m增加到0.8 m,通过相同的方法建立的矿山场景识别模型的精度逐渐提高。其中使用0.8 m空间分辨率的GF-2融合影像建立的矿山场景识别模型的精度最高,平均精准度P_(A)和(MIOU,Mean Intersection over Union)分别达到了0.702和0.824。(2)从多源遥感图像中采集了3162个多场景、多时段、多尺度矿山场景样本对所有样本进行统一融合处理,建立了矿山场地场景识别模型(MSSRM,Mine Site Scene Recognition Model)和矿山场地边界识别模型(MSBRM,Mine Site Boundary Recognition Model)。MSSRM的P_(A)达到了0.758,MSBRM平均交并比达到0.864。(3)对比了Faster R-CNN(FasterRegion-basedConvolutionalNeuralNetwork)、YOLO-v5(YouOnlyLookOnce-v5)、DETR(Detection Transformer)3种目标识别方法与Mask R-CNN、U-Net、DeepLabV3+三种图像分割方法建立的煤矿场地识别模型精度,其中,DETR方法建立的识别模型与Faster R-CNN和YOLO-v5相比P_(A)分别提高了7.6%和8.3%。DeepLabV3+建立的分割模型与Mask R-CNN和U-Net相比MIOU分别提高了14%和10.8%。(4)建立了从大范围的遥感影像中自动化、智能化、批量化识别矿山场地场景并绘制矿山场地边界的方法,以干旱、半干旱典型矿区(鄂尔多斯)露天煤矿场地识别应用为例,验证了智能识别矿山场景边界方法的性能,模型制Mine site data is a crucial foundation for the construction of smart mines and intelligent management.The rapid identification and extraction of complex mine sites from multi-source data,including remote sensing images,is an important research direction.This paper uses Sentinel-2 images from 2020,GF-6 images,and GF-2 images to select the optimal dataset.Google image data from 2023 is used to expand the dataset,which is combined with deep learning algorithms to establish two types of open-pit coal mine site recognition models.The main conclusions of the study are:(1) A mine recognition model was established using 10 m Sentinel-2 images,8 m GF-6 raw images,2 m GF-6 fusion images,3.2 m GF-2 raw images,and 0.8 m GF-2 fusion images.The accuracy of the model produced by different data was quantitatively selected.The results show that as the spatial resolution of remote sensing images increases from 10 meters to 0.8meters,the accuracy of the mine site recognition model established by the same method gradually improves.Among them,the mine site recognition model established using GF-2 fusion images with a spatial resolution of 0.8m has the highest accuracy,with an average precision(P_A) and mean intersection over union(MIOU) of 0.702 and 0.824 respectively.(2) A total of 3 162 multi-scene,multi-time period,and multi-scale mine site samples were collected from multi-source remote sensing images.All samples were uniformly fused to establish a Mine Site Scene Recognition Model(MSSRM) and a Mine Site Boundary Recognition Model(MSBRM).The P_(A) of MSSRM reached 0.758 and the average intersection over union of MSBRM reached 0.864.(3) The accuracy of coal mine site recognition models established by three object recognition methods:Faster R-CNN(faster region-based convolutional neural network),YOLO-v5(You Only Look Once-v5),DETR(Detection Transformer),and three image segmentation methods:Mask R-CNN,U-Net,DeepLabV3+ were compared.Among them,compared with Faster R-CNN and YOLO-v5,the P_(A) of the recognition model established by DE

关 键 词：矿山场景 多源数据 深度学习算法 智能识别 AI模型 

分 类 号：TD171[矿业工程—矿山地质测量] TP79[自动化与计算机技术—检测技术与自动化装置]