Abstract:High-spatiotemporal-resolution land surface temperature (LST) images are essential in various fields of study. However, due to technical constraints, sensing systems have difficulty in providing LSTs with both high spatial and high temporal resolution. In this study, we propose a multi-scale spatiotemporal temperature-image fusion network (MSTTIFN) to generate high-spatial-resolution LST products. The MSTTIFN builds nonlinear mappings between the input Moderate Resolution Imaging Spectroradiometer (MODIS) LSTs and the out- put Landsat LSTs at the target date with two pairs of references and therefore enhances the resolution of time-series LSTs. We conduct experiments on the actual Landsat and MODIS data in two study areas (Beijing and Shandong) and compare our proposed MSTTIFN with four competing methods: the Spatial and Temporal Adaptive Reflectance Fusion Model, the Flexible Spatiotemporal Data Fusion Model, a two-stream convolutional neural network (StfNet), and a deep learning-based spatiotemporal temperature-fusion network. Results reveal that the MSTTIFN achieves the best and most stable performance.

Spatiotemporal Temperature Fusion Based on a Deep Convolutional Network