Estimation of Ardabil land surface temperature using Landsat images and accuracy assessment of land surface temperature estimation methods with ground truth data

Over the past two decades, the intense need for land surface temperature information for environmental studies and management and planning activities has made estimating the land surface temperature one of the most important scientific topics. On the other hand, different methods have been proposed to estimate the land surface temperature, each of which has resulted in different results for different regions. In this study, the algorithms that have had acceptable results in different studies have been selected and evaluated. In the field of thermal studies, what is considered as a major defect in monitoring the land surface temperature is the lack of sufficient meteorological stations to know the temperature values in places without stations and information limitations in preparing temperature data, especially for large areas. The study area is also facing this shortage, and this limitation further highlights the importance of the topic selected for this study to estimate the surface temperature using remote sensing technology. Verification and validation of results obtained from estimating the land surface temperature are other basic and discussed topics in thermal studies. The purpose of this study is an estimation of temperature in Ardabil city and evaluate the accuracy of the four single-channel algorithms, the improved mono-window, the Planck's inversion function method and the radiative transfer equation (RTE) method, to compare the accuracy of the two Landsat 5 and Landsat 8 satellites in estimating the land surface temperature.

Three types of data have been used in this study; Landsat 5 and 8 satellite images, data of two meteorological stations and ground data harvested with a digital thermometer. The images used are from the two satellites Landsat 5 and Landsat 8 with a time interval of 19 years. The meteorological data used were obtained from two synoptic stations in the study area. In addition to land surface temperature, relative humidity, minimum temperature and maximum temperature data of 24 hours were also obtained on two dates. Also, two points of the study area were selected and land surface temperature in the position of these two stations simultaneously with the satellite Recorded from two digital thermometers. MODTRAN web version calculator software version 6 has been used to model the radiation and the amount of atmospheric transmission. Emissivity with two methods of LSE methods based on NDVI and LSE NDVI Thresholds Method and land surface temperature with four algorithms: single-channel algorithms, An Improved mono-window, inversion of Planck’s function and radiative transfer equation using band 6 Landsat 5 and band 10 Landsat 8 bands. It was coded in MATLAB software for 2000 and 2019. Finally, the accuracy of the algorithms was evaluated using synoptic station surface temperature data and field sampling.

The collected data and results are analyzed and while presenting the output maps, the accuracy of the methods with terrestrial and meteorological data as well as the accuracy of Landsat 5 and Landsat 8 satellites in estimating the land surface temperature has been compared and evaluated. The results showed that for the three single-channel algorithms, the inversion of Planck’s function and RTE, the first method of emission and for the An Improved Mono-Window algorithm, the second method of emission had a higher accuracy. Land surface temperature data obtained from meteorological stations in 2000 differ by 12 minutes in terms of time and by 2019 differ by 4 minutes in terms of satellite transit time. The first meteorological station is located somewhat within the city limits and according to the results, it seems that the most important factor is the greater difference between the data of the first station and the estimated LST compared to the second station is the same factor because the heterogeneity of pixels and large changes in levels in urban areas interfere with a pixel value. And subsequently increases the likelihood of errors in estimating surface temperature within the urban anthropogenic range. For the ground station, two points with a homogeneous environment and outside the urban area with agricultural use (alfalfa) and barren use of the harvested product were selected and their surface temperature was measured at the same time as the satellite. The output results of land surface temperature estimation were compared and evaluated with two synoptic stations and two ground stations. In both histories, the single-channel algorithm showed the least difference with the temperature recording stations.

In this research, using Landsat 5 and Landsat 8 satellite images, four algorithms for estimating the land surface temperature of the earth, including single-channel algorithms, An Improved mono-window, inversion of Planck’s function and radiative transfer equation and land surface temperature maps of Ardabil city for two 2000 and 2019 were coded and extracted in MATLAB software environment. The band 6 Landsat 5 satellite was used for 2000 and the band 10 Landsat 8 satellite was used for 2019 due to less noise than the 11th band and the proximity of 9.66 (which is the highest radiation in this range). Comparison of land surface temperature maps obtained by the algorithms with synoptic and ground stations showed that in both 2000 and 2019, the single-channel algorithm was more accurate than the other methods. Comparison of the results of the single-channel method with the stations shows a difference of  +2.5 and 2- with stations 1 and 2 for the year 2000 and a temperature difference of  +3.3, +0.9, 1- and -0.9. Shows stations 1, 2, 3 and 4 for 2019, respectively. It seems that the direct use of atmospheric transmittance coefficients in the single-channel method process has been effective in the high accuracy of this method. In terms of accuracy, after the single-channel algorithm, the An Improved Mono-Window method, the RTE algorithm, and finally the Planck function inverse correlation algorithm were placed, respectively. The results of comparing the output of all four algorithms with the data of stations 1, 2, 3 and 4, show that the ground stations harvested with a digital thermometer are more accurate than the data of meteorological stations. One of the reasons for this is the location of meteorological stations (especially, Station_1) in the urban area due to the heterogeneity of the urban environment and the possibility of pixel interference and temperature interference of land uses, while ground stations from the out-of-town area. And was selected from an environment with homogeneous pixels (barren and agricultural). Also, the results of all four algorithms extracted from the Landsat 8 image show more accuracy compared to the results of the four algorithms obtained from the Landsat 5 image, and due to the improved spatial resolution of the TIRS sensor compared to the TM, the TIRS sensor output is more accurate, It was predictable.