Remotely sensed Land Surface Temperature (LST) from a single source rarely has high temporal and spatial resolutions due to the sensors’optical characteristics. Spatiotemporal fusion uses data from multiple sources to retrieve LST with high temporal frequency and spatial detail, and the spatiotemporal contradiction is disentangled in the fusion process. According to an in-depth study of spatiotemporal fusion, LST exhibits unique features distinct from other land surface variables. However, the inherent mechanism and potential application of LST spatio-temporal fusion have yet to be compiled and extensively explored. Based on the intersection between LST and spatiotemporal fusion, this work collects, analyzes, and summarizes the state-of-the-art developments in LST spatio-temporal fusion. The research background, principles, methods, and applications of this field are systematically elaborated. In particular, the relations and differences with the ubiquitous spatiotemporal fusion technology are emphasized. In essence, spatiotemporal fusion methods extract exquisite temporal variation of pixels from the low spatial resolution images and obtain spatial correspondence from images at various scales to predict high spatial resolution images. The spatiotemporal fusion shows great promise over homogeneous and stable land surface, but has an unsatisfactory performance over heterogeneous landscapes with unstable thermal conditions. In comparison with Land Surface Reflectance (LSR), the spatiotemporal fusion for LST can be less sensitive to the land cover classification uncertainties because of its lower spatial resolution and lower diversity among different land types, but it is difficult to achieve using the general laws for accurate prediction due to the drastic temporal variation of LST. After spatiotemporal fusion was successfully implemented in LSR, several studies adapted it to LST with some improvements based on the thermal characteristics. In the existing five categories of spatiotemporal fusion models based on weight and learning, Bayesian and hybrid models have been applied to LST. Among these models, the weight models are more mature, robust, and effective, but they cannot easily capture the temporal change of LST. Furthermore, the improvement is relatively limited based on STARFM, ESTARFM, or other classical weight models. Learning models can realize a nonlinear prediction based on the structural similarity of training data when supported by reliable network architecture and abundant training. In particular, the deep learning models have more superior ability to depict and extract the LST with weak spectral characteristics, but suitable neural networks and model parameters must be selected and optimized. Although fusion studies based on the Bayesian framework (including maximum a posterior and Bayesian maximum entropy) are relatively rare, they have shown great potential for achieving unbiased and nonlinear predictions and low-quality requirements for the initial data as LST. The hybrid models can integrate the preponderances of the above-mentioned models and acquire more flexible, efficient, and accurate prediction results compared with a single fusion model, which could be the mainstream of the future spatiotemporal fusion model. Although the spatiotemporal fusion models are consistently developed, most of them only focus on generating fused products, with a lack of quantitative and qualitative analysis with respect to the practical applications of the fused LST products, such as agriculture and ecology. In this work, the applications in this field are divided into six aspects: land temperature, sea surface temperature, agroforestry, urban heat island, public health, and others, which cover the majority of remote sensing service fields. However, the breadth and depth of the application of the LST fusion products are less than those of LSR fusion products. The mutual development between theoretical research and application demand is urgently needed. The primary impediment to the application and dissemination of spatiotemporal fusion is the data itself, as evidenced by the diversity of multi-source data, the spatial continuity of image, and the sensitivity of temperature in time series. The angular effect, unstable inversion accuracy, and dramatical diurnal variation significantly constrain their potential applications. Considering these characteristics of LST and existing defects of the spatiotemporal fusion model, this work proposed the future work prospects, such as improving LST inversion accuracy, complementing the strengths of multi-source data, employing a deep learning model, enhancing algorithm flexibility, and constructing a spatiotemporal fusion integrated procedure. The implementation of these strategies will propel the development of theoretical research and operational application of LST with the spatiotemporal fusion technology. © 2022 National Remote Sensing Bulletin. All rights reserved.
