The Research On Spatiotemporal Fusion Of Urban Land Surface Temperature Based On Random Forest Method

Land surface Temperature(LST)has important applications in monitoring urban climate research and land surface temperature with high spatial and temporal resolution.However,the thermal infrared sensors have a trade-off between temporal resolution and spatial resolution due to the technical limitations.For example,the remote sensing images with high spatial resolution tend to have low temporal resolution,but images with high temporal resolution tend to low spatial resolution.How to obtain the images with high temporal and spatial resolution have important significances,the downscaling and spatiotemporal fusion of remote sensing provide an effective way to solve the problem about the contradiction of spatial and temporal.Random forest(RF)and Spatiotemporal fusion method were combined to generate LST data with high spatial and high temporal resolution in this study.The main city areas of Xi'an and Xianyang was selected as the research object,and Sentinel 3 SLSTR and Landsat 8 data were used for spatiotemporal fusion.First,downscaling the Sentinel 3 SLSTR to generate 300 m land surface temperature data.Then,the STARFM and EFSDAF were tested to perform spatiotemporal processing using the downscaled 300 m Sentinel 3 and Landsat 8 LST data.Finally,the true land surface temperature data of Landsat 8 was used to verify the spatiotemporal results.The research content and results are as follows:(1)Research on LST downscaling based on RF.Sentinel 3 OLCI data?DEM data?Land use data and NDVI data are selected as regression factors,using random forest method to downscale the Sentinel 3 LST images to generate 300 m LST images.The accuracy of downscaled 300 m LST images are compared with Sentinel 3 1000 m LST images.The results show that random forest has high accuracy in downscaling the urban surface temperature.Comparing areas with nature surface(weak heterogeneity)and urban surface(strong heterogeneity),the former shows better results and higher accuracy.(2)Research on Spatiotemporal fusion of urban LST.This study proposed a new spatiotemporal fusion method(Enhanced Flexible Spatiotemporal Data Fusion Model,EFSDAF).The downscaled Sentinel 3 LST images and Landsat 8 LST images are used to perform spatiotemporal fusion using STARFM and EFSDAF method,and the accuracy of fusion results are verified using the true Landsat 8 LST images.The results show that the EFSDAF has the high accuracy for the fusion results of urban LST,while STARFM cannot be used for urban LST.In addition,the fusion results of EFSDAF method in urban heterogeneous areas have high accuracy than natural surface areas.(3)Influencing factors for LST spatiotemporal.The quality of the input images for the fusion results and the downscaling process are discussed and analyzed respectively in this study.The Landsat 8 LST images are aggregated to 300 m as the simulated Sentinel 3 LST images which are used for spatiotemporal with the 30 m Landsat 8 LST images.The results show that the quality of the input images directly affect the final fusion results,and the accuracy based on simulated images is higher than the true images.In addition,the accuracy based on RF and EFSDAF has better than the EFSDAF,especially in small local areas of study.