利用集合深度学习方法融合多源数据开发全国能见度网格数据

利用集合深度学习方法融合多源数据开发全国能见度网格数据

Fuse Multiple Data Sources with an Ensemble Deep Learning Approach to Estimate Nationwide Gridded Visibility

查看全文

作者:

  • 吕宝磊 华云升达(北京)气象科技有限责任公司 北京 102299
  • 胡泳涛 佐治亚理工学院土木与环境工程学院 亚特兰大 30332
  • 李林 北京市气象局 北京 100089
  • 梁海河 中国气象局气象探测中心, 北京 100081
  • 刘钧 华云升达(北京)气象科技有限责任公司 北京 102299
  • 王晓江 中国华云气象科技集团公司 北京 100081

中文摘要:

能见度对于生产生活安排具有重要的指导意义,目前我国的能见度监测网络站点覆盖相对稀疏,观测数据具有空间离散性和局地性。利用集合深度学习模型融合多源数据估计能见度,该集合模型包括了深度神经网络模型、随机森林模型、梯度强化模型和广义线性模型四种机器学习器,融合了气象观测、大气成分观测、模式模拟和土地利用类型等多源数据,并利用Barnes客观分析进一步消除深度学习融合误差,开发出空间连续的能见度网格数据集。该数据的空间分辨率为12 km,经过独立样本评估,融合数据的R20.61。相比较于空间插值和线性模型等方法,提出的方法具有更好的准确度,且具有很好的空间解析力,可进一步用于开发更高分辨率数据。该方法具有较高的计算效率和较好的数据兼容性,可以部署于业务化平台中可靠运行。

中文关键词:

深度学习,能见度,多源数据融合,Barnes客观分析

KeyWords:

deep learning, visibility, data fusion, Barnes objective analysis

Abstract:

Ground visibility had significant influences on transportation and human outdoor activities. Stationary visibility observations are usually spatially sparse, inadequate to meet the demand for complete spatial coverage data. To tackle the problem, we put forward a novel data fusion framework by incorporating an ensemble deep learning method and a Barnes objective analysis approach. The ensemble deep learning method works to generate high-quality first-guess field by building complex linkages between data sets. The Barnes objective analysis method was used to remove remaining residual errors and biases. Furthermore, it was designed to be hierarchical to better address the environmental factors such as particulate concentrations and relative humidity that greatly affect visibility in a sub-layer. The data fusion framework is able to include multiple data sources of different types, i.e. gridded WRF meteorological predictions, CMAQ air quality predictions, stationary meteorological and atmospheric composition observations and other supporting land use/cover data sets. The model was implemented to generate gridded hourly visibility data sets in China. The fused gridded data was evaluated against observations from independent monitors, with an R2 = 0.61. Comparatively the R2 values of interpolation approach and linear regression fusion were respectively 0.55 and 0.57. Besides, our gridded fusion data have much more and detailed spatial information than that of the smoothed interpolation data. Our data fusion approach is relatively easy to implement in operational system and also has good extensibility to handily include more data sets.

版权所有:《气象科技进展》编辑部

联系电话:(010)58993291 68409927

Email:qxkjjz@163.com

气政邮:《气象科技进展》编辑部/中国气象局图书馆(气象科技史研究中心)/干部学院/中国气象局
地址:北京市中关村南大街46号中国气象局气象干部陪训学院 邮编:100081