基于GF-1遥感图像和ResNet50网络的PM_(2.5)浓度预测研究

Research on PM_(2.5) Concentration Prediction Based on GF-1 Remote Sensing Image and ResNet50 Network

基于GF-1卫星图像、结合Merra-2气象数据作为辅助预测变量,构建了长三角地区基于ResNet50网络的PM_(2.5)预测模型。其中气象参数可以为模型提供较为准确的PM_(2.5)浓度基准,而GF-1图像能帮助模型更合理准确地预测PM_(2.5)浓度的空间变化。利用十折交叉验证和测试集验证对模型进行检验,结果显示:模型的皮尔森相关系数R为0.948,预测PM_(2.5)的RMSE为6.6μg/m~3。反演得到分辨率为500 m的PM_(2.5)浓度分布图合理稳健。GF-1遥感图像和ResNet50网络适用于PM_(2.5)浓度预测,可以作为辅助监测手段,为长三角地区PM_(2.5)热点识别、后续流行病学研究提供数据支撑。

By referring to GF-1 satellite images and taking Merra-2 meteorological data into account as the variables of auxiliary prediction,PM_(2.5) prediction model based on ResNet50 network was constructed for the Yangtze River Delta.According to this model,meteorological data could be used to provide an accurate PM_(2.5) concentration benchmark,while GF-1 images could be used to predict the spatial change of PM_(2.5) concentration in a more reasonable and accurate way.The results showed that the Pearson correlation coefficient of the model was 0.948 and the RMSE was 6.6μg/m3.As obtained from model inversion,the distribution of 500m resolution PM_(2.5) concentration map was verified to be reasonable and robust,indicating that GF-1 remote sensing image and ResNet50 network were suitable for the prediction of PM_(2.5) concentration.The model could be used as auxiliary monitoring means to provide data support for PM_(2.5)hot spot identification and follow-up epidemiological research in the Yangtze River Delta region.