Method And Application Research On Spatial And Temporal Fusion With Multi-source Remote Sensing Of Land Surface Temperature Data

Land surface temperature is an important parameter for the surface condition, which is of great importance to the study of hydrology, ecology, environment and biogeochemistry. The emergence of remote sensing technology has made it possible to obtain large-scale and continuous land surface temperature products. However, due to the limits of the load capacity of satellites and the bottleneck of sensors manufacture, it is difficult for current mainstream satellite sensors to obtain land surface temperature data with both high spatial and high temporal resolution, which restricts the promotion and application of land surface temperature data retrieved by remote sensing. Spatial and temporal fusion of the quantitative information from multi-souce satellites is an effective way to solve this problem. Base on the spatial and temporal fusion technology, this thesis studies the reflectance spatial and temporal fusion model and the land surface temperature downscaling method, and then proposes a new framework for land surface temperature spatial and temporal fusion, and finally, introduces the spatial and temporal fusion technology to the adaptive forest fire monitoring system. This thesis mainly includes the following four aspects:(1) Reflectance spatial and temporal fusion model based on convolutional neural network and sparse representation algorithm. The traditional reflectance spatial and temporal fusion model is implemented based on the interpolation of the neighborhood similar pixels, which may lead to obvious errors when significant change occurs on the phenology or type of the objects in the image for fusion. Aiming at this drawback, this thesis regards the reflectance spatial and temporal fusion procedure as the super resolution reconstruction of changed reflectance image, and reconstructs the contour and detailed feature of the changed reflectance image using the convolutional neural network and sparse representation algorithm, then implements the reflectance fusion of predicted time by time weighted integration of the reconstructed images from different observation time. Some areas with high heterogeneity are selected from Wuhan and Beijing for experiment. Compared with the traditional methods, the proposed method has better prediction on the change of reflectance made by phenology and land feature change, moreover, due to the step wise image feature learning by introducing the convolutional nerual network and sparse representation algorithm, the details of reflectance image fused by our method are clearer and more accurate. Aiming at different applications in high heterogeneous area, the proposed fusion model can provide high precision reflectance images with both high spatial and temporal resolution.(2) Land surface temperature downscaling method based on BP artificial neural network. The land surface temperature products always has the problem of insufficient information under high resolution, and requires information with higher spatial resolution to complete the scale change. Based on this theory, this thesis builds the scale transformation model which can describe the distributions between NDVI and LST on the low spatial resolution scale, according to the explicit and implicit relation, respectively, so as to improve the generalization ability of the model. Simulation is performed with the NDVI-LST distribution of the dry objects as the explicit relation, and the complex and implicit relation is learned using the BP neural network between the residuals of land surface temperature and a variety of surface parameters generated by visual and near infrared data, then a variety of surface parameters with higher spatial resolution is applied to the trained scale transformation model to generate the high spatial resolution LST image. The proposed method could effectively relief the mismatching phenomenon on the low spatial resolution land surface temperature images and the refinement earth system research.(3) LST spatial and temporal fusion model based on the downscaling technology. Traditional LST space-time fusion methods tend to be affected by the space-time variability of pixel emissivity, which may cause significant fusion errors when applied to the high heterogeneity areas. Aiming at this problem, this thesis proposes a new idea which regards the LST spatial temporal fusion model as the scale transformation of LST image on prediction time. The prediction time land parameters fused by reflectance spatial and temporal model are introduced to the LST downscaling method, and the scale transformation of LST image on prediction time is realized. Taking the complex surface areas and the effectiveness of the models in different seasons into consideration, experiments are performed to our model and to the traditional model using multi-source sensor data, and the results show that the proposed method is more suitable for the high heterogeneity areas and exhibits good fusion precision on data of different seasons. Reliable technical support could be provided by proposed fusion model for expanding quantitative applications of land surface temperature data in the field of global temperature change.(4) Adaptive forest fire monitoring model based on spatial temporal fusion. Traditional forest fire monitoring models tend to be affected by the geographic and seasonal factors, which leads to missing or false alarm of fires. This thesis builds a typical sky-clear surface brightness temperature model based on the spatial and seasonal distribution characteristics, and in order to improve the parameter accuracy of the model, use the high resolution brightness temperature images generated by space-time fusion technology to increase the training sample data. According to the typical sky-clear surface brightness temperature model and the surface reflective radiation theory, a functional dynamic threshold is established, and a forest fire monitoring model based on space-time dynamic threshold is built. Based on the experiment data acquired from the environmental disaster mitigation satellite (HJ-1B), a typical sky-clear surface brightness temperature model of Heilongjiang province is built, and the proposed fire monitoring model is applied to a forest fire case in the province, which proves the reliability of the fire monitoring model.