Land surface temperature algorithm calibration through meteorological stations

This methodologic paper arises from the necessity to gather Land Surface Temperature (LST) data over a relatively large period and territory: 2000-2018, Portugal. The computational power required to complete this task was found to be a major barrier. However, platforms such as Google Earth Engine (GEE) offer a vast data archive freely accessible through a web interactive development environment or an application programming interface, namely, Python's API. Additionally, the computation using GEE is hosted in Google's servers, drastically reducing the processing times. However, computing LST through Landsat-7 satellite imagery resulted on a difference of-8ºC±6ºC compared to the values from meteorological ground stations. As such, this paper aims to further calibrate computed LST through meteorological stations and make the methodology and corresponding code available, thus encouraging cooperation on the development and integration of local calibration methods. A sensitivity analysis of the representativeness of each station was performed using three methods of temperature extraction: station coordinate's pixel, buffers around the station, and surrounding soil occupation (identifying the area with the same soil occupation as the station's location). Pearson's correlation coefficient was on average significant at 0.81 in the raw data and increased to 0.89 after clearing data from outliers. The best representativeness method for meteorologic stations was the one based on soil occupation, which resulted on a Pearson's r of 0.91. As a result, we advise researchers to complement their remote sensing work with ground data whenever possible through the usage of a method like the one here described.