A Machine Learning Based Downscaling Approach to Produce High Spatio-Temporal Resolution Land Surface Temperature of the Antarctic Dry Valleys from MODIS Data

• Published in: Remote Sensing
• Main Authors: Lezama Valdes L-M, Meyer H, Katurji, Marwan
• Format: Article in Journal/Newspaper
• Language: English
• Published: MDPI AG 2021
• Subjects: downscaling; land surface temperature; Antarctica; McMurdo dry valleys; MODIS; machine learning; Fields of Research::37 - Earth sciences::3701 - Atmospheric sciences::370108 - Meteorology; Fields of Research::37 - Earth sciences::3702 - Climate change science::370201 - Climate change processes; Fields of Research::37 - Earth sciences::3709 - Physical geography and environmental geoscience; Antarctic; McMurdo Dry Valleys; The Antarctic; Antarc*
• Online Access: https://hdl.handle.net/10092/103291; https://doi.org/10.3390/rs13224673

To monitor environmental and biological processes, Land Surface Temperature (LST) is a central variable, which is highly variable in space and time. This particularly applies to the Antarctic Dry Valleys, which host an ecosystem highly adapted to the extreme conditions in this cold desert. To predict possible climate induced changes on the Dry Valley ecosystem, high spatial and temporal resolution environmental variables are needed. Thus we enhanced the spatial resolution of the MODIS satellite LST product that is sensed sub-daily at a 1 km spatial resolution to a 30 m spatial resolution. We employed machine learning models that are trained using Landsat 8 thermal infrared data from 2013 to 2019 as a reference to predict LST at 30 m resolution. For the downscaling procedure, terrain derived variables and information on the soil type as well as the solar insolation were used as potential predictors in addition to MODIS LST. The trained model can be applied to all available MODIS scenes from 1999 onward to develop a 30 m resolution LST product of the Antarctic Dry Valleys. A spatio-temporal validation revealed an R2 of 0.78 and a RMSE of 3.32 ∘C. The downscaled LST will provide a valuable surface climate data set for various research applications, such as species distribution modeling, climate model evaluation, and the basis for the development of further relevant environmental information such as the surface moisture distribution.