基于SCLSTM模型的MODIS地表温度产品重建方法  

MODIS LAND SURFACE TEMPERATURE DATA RECONSTRUCTION BASED ON THE SCLSTM MODEL

作　　者：宋冬梅[1,2] 张曼玉 单新建 王斌[1] SONG Dong-mei;ZHANG Man-yu;SHAN Xin-jian;WANG Bin(College of Oceanography and spatial information,China University of Petroleum,Qingdao 266580,China;The Laboratory for Marine Mineral Resources,Qingdao National Laboratory for Marine Science and Technology,Qingdao 266071,China;Institute of Geology,China Earthquake Administration,Beijing 100029,China)

机构地区：[1]中国石油大学(华东)海洋与空间信息学院,青岛266580 [2]海洋矿物资源实验室,青岛海洋科学技术国家实验室,青岛266071 [3]中国地震局地质研究所,北京100029

出　　处：《地震地质》2023年第6期1349-1369,共21页Seismology and Geology

基　　金：国家重点研发计划项目(2019YFC1509202);国家自然科学基金(41772350,61371189,41701513)共同资助。

摘　　要：MODIS(Moderate-resolution Imaging Spectroradiometer,中分辨率成像光谱仪)LST(Land Surface Temperature,地表温度)产品在大气物质和能量交换、气候变化研究及地震前兆热异常探测等方面具有重要价值。然而,由于云的遮挡导致MODIS LST数据产品中存在大量空值,限制了其广泛应用。为此,文中提出了一种基于混合模型的地表温度重建方法——SCLSTM(即SSA-CLSTM)。与传统方法相比,该方法无需建立复杂的回归关系模型。此外,由于CNN(Convolutional Neural Network,卷积神经网络)能够充分提取一维时间序列数据的局部特征,而LSTM(Long Short-Term Memory,长短期记忆)能够充分学习数据的长时间序列特征,因此将CNN和LSTM结合能够更加充分地学习数据的潜在特征。首先,使用SSA(Singular Spectrum Analysis,奇异谱分析)模型提取出地表温度时间序列中的趋势值用于填补缺值像元,实现地表温度的初步重建。然后,再利用SCLSTM(即1DCNN-3层堆叠LSTM)模型学习数据的局部时序特征和长期依赖关系,并实现对缺失像元的地表温度进行迭代预测,完成数据的精细重建。新疆和田地区和四川汶川地区的实验结果表明,文中方法与现有其他2种基于混合模型的重建方法相比,重建后的LST数据误差最小,与原始数据的一致性最高。其中,文中方法的RMSE可降至0.712K,AD为0.695K,重建后的LST数据与原始数据的相关系数可达0.95以上。此外,气象站的实测地表温度数据也进一步验证了该方法的可靠性。文中所提方法为基于深度学习的LST重建工作提供了一种新的技术手段和思路,同时也为基于LST的地表过程和地震热异常研究提供了坚实的数据基础。MODIS land surface temperature(LST)products are of great value in the exchange of atmospheric matter and energy,climate change research,and detection of thermal anomalies as earthquake precursors.However,due to the influence of the cloud,there are a large number of missing values in the MODIS LST data products,limiting its wide application.Therefore,in this study we propose a method of surface temperature reconstruction based on mixed model:SCLSTM(SSA-CLSTM).Compared with traditional methods,this method does not need to establish a complex regression relationship model.In addition,since CNN can fully extract local features of one-dimensional time series data,and LSTM can fully learn long-term time series features of data,the combination of CNN and LSTM is capable of fully learning potential features of data.Firstly in this study,the trend value of LST time series is extracted by SSA model to fill the missing pixel,and the initial reconstruction of LST is realized.Then,CLSTM(that is,1DCNN,three-layer stacked LSTM)model is used to learn the local temporal characteristics and long-term dependence of the data,and the iterative prediction of the surface temperature of the missing pixel is realized to complete the fine reconstruction of the data.Based on the experimental results in Hotan region of Xinjiang and Wenchuan region of Sichuan,it can be proved that compared with the other two existing reconstruction methods based on mixed models,the reconstructed return data error is minimum,and the consistency with the original data is the highest.The RMSE of this method can be reduced to 0.712K,the minimum is 0.695K,and the correlation between the reconstructed return data and the original data can reach more than 0.95.In addition,the reliability of the method is further verified by the measured surface temperature data of the meteorological station.In summary,the proposed method provides a new technical means and ideas for deep learn-based reconstruction work,and also provides a solid data foundation for the research of su

关 键 词：地表温度 SSA CNN LSTM MODIS 

分 类 号：P315.2[天文地球—地震学]