Improving the disaggregation of MODIS land surface temperatures in an urban environment: a statistical downscaling approach using high-resolution emissivity

Spatially and temporally dense land surface temperature (LST) data are necessary to capture the high variability of the urban thermal environment. Sensors on board satellites with high revisit time cannot provide adequately detailed spatial information; thus, the downscaling of LST is recognized as being an important and inevitable intermediate process. In this paper, improvement in the downscaled LST accuracy is investigated, employing the statistical downscaling methodology in an urban setting. A new approach is proposed, where thermal radiances are disaggregated using multiple regression analysis and are then combined with emissivity values derived from a high-resolution image classification. Predictors include reflectance values, built-up and vegetation indices, and topographic data. Surface classification is performed utilizing machine learning techniques and fusing Sentinel-2 imagery with ancillary data. Thermal data from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor are downscaled from their original resolution to 100 m in the city of Athens, Greece. Validation of sharpened temperatures is performed using the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) surface temperature product and in-situ measurements. It is demonstrated that the proposed downscaling framework using ridge regression has the potential to produce reliable, high temporal LST estimates with an average error of fewer than 2 K, while consistently having a better accuracy than the reference, single-predictor downscaling of the MODIS LST product.