Researchers develop method to obtain fine spatial and temporal resolution land surface temperature data

Scientists need fine spatial and temporal resolution land surface temperature (LST) data for many types of research and applications. Spatio-temporal fusion, a technique that combines data from multiple sources to create high-resolution images with both spatial (space) and temporal (time) details, is an important solution for researchers needing fine spatio-temporal resolution LST data. A team of researchers propose a new spatio-temporal fusion method based on Restormer (RES-STF).

Their work is published in the Journal of Remote Sensing on August 21, 2024.

LST data, the measurement of how hot the Earth’s surface is from a satellite’s point of view, is an important tool in monitoring global change. High spatio-temporal resolution LST data has many practical applications and plays a crucial role in guiding agricultural practices, irrigation, and drainage, as well as in studying urban heat environment changes. However, because of limitations in the remote sensing sensors’ hardware design, trade-off always exists between temporal and spatial resolution in satellite-derived LST data.

Scientists acquire the LST data from thermal infrared remote sensing images collected by various satellite sensors, such as Landsat TM/ETM+/TIRS, Moderate Resolution Imaging Spectroradiometer (MODIS), Advanced Very High-Resolution Radiometer (AVHRR), Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), and Visible Infrared Imaging Radiometer Suite (VIIRS).

Scientists can create 100-m, daily LST data by fusing 1-km, daily MODIS LST with 100-m, 16-day Landsat LST data. MODIS is a spectroradiometer instrument about the Terra and Aqua satellites. Landsat are Earth-observing satellites. But in recent years the quality of MODIS LST products has been decreasing noticeably, which greatly impacts fusion accuracy. So the research team proposes replacing the MODIS LST data with LST data from the Visible Infrared Imaging Radiometer Suite (VIIRS) in spatio-temporal fusion. The VIIRS instrument flies aboard Joint Polar Satellite System.

To handle the data discrepancy that results because of the large difference in overpass time between VIIRS LST and Landsat LST, the team proposes the spatio-temporal fusion method based on Restoration Transformer, or Restormer. Restormer is a model that handles large-scale data, and it combines the advantages of convolutional neural networks and the Transformer model. The convolutional neural networks offer multi-scale feature extraction and Transformer provides global dependency to effectively capture both local and global information in the images. The research team’s proposed Restormer spatio-temporal fusion method is called RES-STF and it copes with the challenges brought by different overpass time when fusing VIIRS and LANDSAT LST.

“The problem we aimed to address in this paper is the challenge of selection of suitable alternative daily time-series LST data for spatio-temporal fusion with Landsat LST, which provides fine spatial resolution, but coarse temporal resolution, to create 100-m, daily LST,” said Qunming Wang, a professor at Tongji University. “The strategy of using VIIRS LST presents greater quality and leads to more accurate fusion results,” said Wang. 

“The most important message is that VIIRS LST data are an effective alternative to MODIS LST in the spatio-temporal fusion task for creating 100-m, daily LST. Also, the proposed RES-STF method can cope with the data discrepancy brought by different overpass time when fusing VIIRS and Landsat LST,” said Ruijie Huang, a master student at Tongji University. The team’s research showed that RES-STF shows strong robustness through experiments in three different regions. It is also more accurate than one typical non-deep learning-based method, and three typical deep learning-based methods. Compared to the MODIS LST data, the VIIRS LST data contains richer spatial texture information which leads to more accurate fusion results.

Looking ahead, the team’s next step is to enhance the deep learning-based spatio-temporal fusion method RES-STF, especially when training data are difficult to collect, and combine other feasible data to deal with LST with great spatio-temporal heterogeneity. “Our ultimate goal is to obtain fine spatial and temporal resolution LST with great quality, providing critical data for studying urban heat environment changes and guiding agricultural practices, irrigation and drainage,” said Wang.

The research team includes Qunming Wang and Ruijie Huang from the College of Surveying and Geo-Informatics, Tongji University, Shanghai, China.

This research was supported by the National Natural Science Foundation of China.