首页 > , Vol. , Issue () : -
准确识别工业热源是我国大气污染防治的重要前提，然而因热源特征不明、类型判定不准等问题，工业热源遥感监测目前难以大范围推广应用。本文对此提出了一种耦合温度特征的工业热源人工神经网络遥感分类精准识别方法。该方法首先基于DBSCAN聚类算法和土地利用类型从Suomi-NPP VIIRS Nightfire数据中识别全部类型工业热源；其次利用频率统计方法分类建立工业热源温度特征模板；最后构建人工神经网络模型判别工业热源类型并由此分析了我国近年工业热源的时空动态演化特征。结果表明：（1）煤炭加工业、金属冶炼及压延加工业、水泥、石灰和石膏制造业、精炼石油产品制造业四类工业热源各级温度频率与分布形态、主次峰峰值均呈现明显类间差异，主峰温度分别为795 K、830 K、760 K、1725 K。（2）温度特征模板增强下的人工神经网络方法工业热源分类识别效果好，模型训练集与验证分类识别精度分别为99%和88.17%；（3）我国工业热源时空分布呈现出“地域集中”与“波动下降”双特征。工业热源空间上主要集中分布在北方地区，数量占比高达85.4%。煤炭加工业、金属冶炼及压延加工业、精炼石油产品制造业和水泥、石灰和石膏制造业的主要分布地分别为山西、河北、新疆以及安徽；以2015年和2018年为拐点，四类工业热源在2013-2020年间整体表现出先降后升再降的波动变化趋势。本研究提出的工业热源遥感识别方法可为基于卫星手段的大气工业污染源遥感监测提供技术支撑。
Objective: As one of main air pollution sources, the spatial-temporal distribution and category dependent determination of industrial heat sources are critical for policy making of air pollution control. However, due to the lack of identified characteristics, it is difficult to clearly differentiate the sub categories of the industrial heat sources in large geographical area using remote sensing technology. For that, we proposed a satellite-based artificial neural network (ANN) identification method for industrial heat sources by coupling with temperature characteristics in this study by taking the whole mainland China as a case. Method: The Suomi-NPP Nightfire products containing location and temperature information in mainland China from 2013 to 2020 were firstly collected and screened as industrial heat source clusters based on DBSCAN clustering algorithm and land use data. Then, four types of temperature characteristic templates depended on industrial heat source clusters were generated by combining the frequency statistical analysis with Gaussian function. Finally, a temperature characteristic template enhanced ANN model was developed to discriminate the sub categories of the recognized industrial heat sources and subsequently analyze their spatio-temporal changes. Result: Results illustrate that there are significant differences in temperature frequency, distribution pattern and major-minor peaks among four types of industry heat sources (i.e. coal processing (CP), metal smelting and rolling(MSR), cement lime and gypsum manufacturing(CLGM) and refined petroleum products manufacturing(RPPM)) with their major peak temperatures being 795 K、830 K、760 K and 1725 K, respectively. Moreover, with the enhancement of temperature characteristic template, the ANN model performs very well in identify the category depended industrial heat sources, with the training and verification accuracy of 99% and 88.17%, respectively. Besides, spatial-temporal distribution of industrial heat sources in Mainland China demonstrates the dual characteristics of "regional concentration" and "decreasing fluctuations". Industrial heat sources are mainly concentrated in the northern region, accounting for 85.4% of the whole country. The main locations of CP, MSR, RPPM, and CLGM are Shanxi, Hebei, Xinjiang, and Anhui, respectively. In the period of 2013 to 2020, the overall trend of fluctuations is “descent - ascension - descent”, taking 2015 and 2018 as the turning time. Conclusion: There are obviously difference in temperature frequency, distribution pattern and distribution statistics among four types of industrial heat sources. Based on these differences, the temperature characteristic templates constructed are reliable and credible to discriminate the sub categories of industrial heat sources. Temperature characteristic template enhanced ANN model would provide a newly promising way for satellite-based precise identification of industrial heat sources by combining the temperature feature of industrial source and the super self-learning ability of ANN method.