Unlocking time series analysis with data-fusion land surface temperature

Consistent and continuous Land Surface Temperature at high temporal resolution is essential for many applications, such as anomaly detection (e.g. agricultural droughts), urban heat island monitoring or irrigation and crop water stress, among others. LST can be retrieved at high spatial resolution from spaceborne thermal infrared (TIR) instruments, like MODIS/VIIRS, Landsat, ECOSTRESS, and ASTER. But these data come with large temporal gaps due to cloud cover and orbit/sensor characteristics and consequently complicate time series analyses. 

To overcome these limitations, we developed a daily 100m LST product based on the synergy between passive microwave brightness temperatures from the Advanced Microwave Scanning Radiometer 2 (AMSR2), and optical data from Sentinel 2 within a novel disaggregation method [1]. This results in a dataset that allows for monitoring environmental systems consistently and continuously in near-real time. The passive microwave observations offer a distinct advantage in LST estimation due to the ability to penetrate cloud cover and measure thermal emissions at the surface. On the other hand, Sentinel-2, with its high spatial resolution multispectral bands, provides rich information on land cover and land surface properties. By combining these complementary datasets, we aim to leverage the strengths of both sensors to improve the accuracy and spatial resolution of LST retrieval. The method uses the abundance of overlaps between passive microwave footprints in combination with higher spatial information from S2 NIR and SWIR for downscaling at 100m resolution since 2017 at 1:30am and at 1:30pm. 

To assess the accuracy of the 100m LST, we compared the time series of microwave-based LST at 100+ locations against in situ measurements, MODIS and Landsat LST data. Currently, the temporal accuracy compared to these in-situ stations is ±3.1K with a correlation of 0.91 (for MODIS this was 2.6 and 0.94). We also performed a spatial comparison of our 100m LST data over agricultural regions against Landsat LST. While the few clear-sky Landsat LST observations is a limitation for the comparison, the preliminary results show a spatial accuracy between ±1.5K and ±4K.

Our results demonstrate that our LST data-fusion approach is a viable methodology to generate a temporally and spatially high resolution LST archive. We are aiming to bridge current and future missions for high-resolution LST by harnessing the complementary capabilities of multi-sensor data fusion. The proposed framework holds great potential for improving our understanding and monitoring Earth's complex environmental systems, such as local surface energy dynamics, climate processes or supporting various environmental applications requiring accurate and high-resolution LST information.

[1] de Jeu, R. A. M., de Nijs, A. H. A., & van Klink, M. H. W. (2017). Method and system for improving the resolution of sensor data. https://patents.google.com/patent/WO2017216186A1/en