Comparison of Land Surface Temperature Algorithm Using Landsat-8 Data for South Korea

Abstract

Land Surface Temperature (LST) is the radiological surface temperature which observed by satellite. It is very important factor to estimate condition of the Earth such as Global warming and Heat island. For these reasons, many countries operate their own satellite to observe the Earth condition. South Korea has many landcovers such as forest, crop land, urban. Therefore, if we want to retrieve accurate LST, we would use high-resolution satellite data. In this study, we made LSTs with 4 LST retrieval algorithms which are used widely with Landsat-8 data which has 30 m spatial resolution. We retrieved LST using equations of Price, Becker et al. Prata, Coll et al. and they showed very similar spatial distribution. We validated 4 LSTs with Moderate resolution Imaging Spectroradiometer (MODIS) LST data to find the most suitable algorithm. As a result, every LST shows 2.160 ~ 3.387 K of RMSE. And LST by Prata algorithm show the lowest RMSE than others. With this validation result, we choose LST by Prata algorithm as the most suitable LST to South Korea.

1.Introduction

Land surface Temperature (LST) can be defined as the radiological surface temperature and can be changed by surface condition, such as landcover, land moisture, and air condition (Becker and Li, 1990). LST can be used input data or validation data for numerical model and climate model which can predict heat island phenomenon, evaporation (Moran and Jackson, 1991; Kustas and Norman, 1996; Benali et al., 2012). If it is necessary for prediction with high accuracy about various climate phenomena, direct measurement is the best method. But it is impossible to set up measurement equipment is every place in whole area. Therefore, climate model or remote sensing data from satellite can be used to solve the spatial problem.

There are many land types such as building, park, road, cropland, forest in South Korea. There is very difficult to observe exactly with remote sensing data which has low spatial resolution. For example, Moderate resolution Imaging Spectroradiometer (MODIS) which is used in National Aeronautics and Space Administration (NASA) has about 1 km spatial resolution. NASA offer LST data from MODIS, but it cannot reflect various phenomena in land, like heat island, exactly. In contrast, if we can use high resolution satellite data, we can solve many phenomena in land easier. Therefore, we used Landsat-8 data which has 30 m spatial resolution to retrieve LST.

2. Data & Study area

Landsat-8 is satellite launched by United States Geological Survey (USGS) in 2013. It observes entire earth in every 16 days with from 30 to 100 m, high spatial resolution. With 30 m resolution, we can retrieve LST data which reflect various land types. Landsat-8 has 2 sensors, Operational Land Imager (OLI) and Thermal InfraRed Sensor (TIRS). OLI has 9 bands from 0.43 to 1.36 μm, from visible to SWIR wavelength which have 30 m spatial resolution. And TIR has 2 bands from 10.6 to 12.51 μm, thermal infrared wavelength which have 30 m spatial resolution but USGS serves TIR data convert to 30 m resolution (Table 1.).

Table 1. Characteristic of Landsat-8

Study area in this study is Landsat-8 images, 36.42 ~ 38.51°N, 125.25~127.98°E, and it observes this area every 16days. To retrieve LST exactly, air condition such as clouds, fog have to be removed. Therefore, we choose 7 scenes with very clear condition which has no clouds, no fog (2017.01.14, 02.15, 03.19, 06.23, 10.13, 11.14, 11.30).

3. Methods

1) Normalized Difference Vegetation Index

Normalized Difference Vegetation Index (NDVI) is one of important index which indicate vegetation growth and distribution. It reacts sensitively to climate change and observing climate change such as global warming and desertification. NDVI is calculated by normalizing the difference between the red band (band 3) reflectivity and the Near InfraRed (NIR, band 4) band reflectivity. In this study, NDVI is used to calculate emissivity. The formula for calculating NDVI is as follows.

\(\mathrm{NDVI}=\frac{N I R-R e d}{N I R+R e d}\)

2) Land surface emissivity

Land surface emissivity (LSE) is ratio of radiative energy emitted to actual radiative energy based on black body. LSE is changed by land roughness and satellite viewing angle (Sobrino et al., 2001; Sobrino et al., 2005). It is used for land development and erosion studies. LSE can be calculated by NDVI, and the equation is as follows (Han et al., 2004).

ε₁₁= 0.9897 + 0.029 ln (NDVI)

Δε= 0.01019 + 0.01344 ln (NDVI)

3) Land surface temperature

Infrared is influenced by water vapor and other gas. That means that land surface is easy to be influenced by water vapor and gases and have high variability. Therefore, the wavelength 10.5 ~ 12.5 μm, called the atmospheric window, can be used to eliminate water vapor effects. This technique is called split window method and can retrieve accurate LST with The Advanced very high-resolution radiometer (AVHRR) data. In this study, we used 4 LST retrieval algorithms to make LST using Landsat-8 data. And these algorithms are based on split-window method. These formulae for calculating LST are as follows.

Price equation (1984)

\(\begin{array}{c} L S T=T_{11}+3.33 *\left(T_{11}-T_{12}\right) * \frac{\left(5.5-\varepsilon_{11}\right)}{4.5}+ \\ 0.75 * T_{11} * \Delta \varepsilon \end{array}\)

Becker et al. equation (1990)

\(\begin{aligned} L S T=& 1.274+\left(1+0.15616 * \frac{1-\varepsilon}{\varepsilon}-0.482 * \frac{\Delta \varepsilon}{\varepsilon_{2}}\right) \\ & * \frac{T_{11}+T_{12}}{2} \\ &+\left(6.26+3.98 * \frac{1-\varepsilon}{\varepsilon}+38.33 \frac{\Delta \varepsilon}{\varepsilon_{2}}\right) \\ & * \frac{T_{11}-T_{12}}{2} \\ \varepsilon=\frac{\Delta \varepsilon}{2} & \end{aligned}\)

Prata equation (1991)

\(\begin{aligned} L S T=& 3.45 * \frac{T_{11}-273.15}{\varepsilon_{11}}-2.45 * \\ & \frac{T_{12}-273.15}{\varepsilon_{12}}+40 * \frac{1-\varepsilon_{11}}{\varepsilon_{11}}+273.15 \end{aligned}\)

Coll et al. equation (1994)

LST= T₁₁+ [1 + 0.58 *(T₁₁–T₁₂)] * (T₁₁–T₁₂) + 0.51 + 40 * (1 –ε) –75Δε

4. Result and Analysis

Fig. 1. Shows result of 4 LST from 4 algorithms. From (a) to (d) show LST on 2017.06.23 and from (e) to (h) shows LST on 2017.11.14. Fig. 1(a) ~ (d) has temperature range about 280~315 K and Fig. 1. (e)~ (h) has temperature range 270~300 K which are in reasonable range. There are no error value and no noise from cloud or other air condition. From this result, we consider that 4 algorithms retrieve each LST well.

Fig. 1. Result of LST by split window method algorithms (a) Price (2017.06.23), (b) Becker et al(2017.06.23), (c) Prata (2017.06.23), (d) Coll et al. (2017.06.23), (e) Price et al. (2017.11.14), (f) Becker et al. (2017. 11.14), (g) Prata (2017.11.14), (h) Coll et al. (2017.11.14).

After checking spatial distribution of LST, we compared each LST. In this comparison, reference data is the LST by Becker et al. equation because the Becker’s equation is used to retrieve LST with MODIS data and that LST is officially served by NASA. Fig. 2. Shows results of comparison with each algorithm. (a) show result of Becker et al. and Price and it shows R is 0.999 and RMSD 1.488 K. (b) show result of Becker et al. and Prata and it shows R is 0.999 and RMSD 1.941 K. (c) show result of Becker et al. and Coll et al. and it shows R is 0.998 and RMSD 2.119 K. 3 of results show very high correlation which are about 1. But Coll et al. show higher RMSD than other algorithms and we can find some different temperature pixels in the graph. As a result, although there are very small difference, Coll et al.’s LST is the lowest accuracy in 4 LST retrieval algorithms.

Fig. 2. Comparison of LSTs (a) Becker et al. vs. Price, (b) Becker et al. vs. Prata, (c) Becker et al. vs. Coll et al.

We tested change of LST in passage of time in 2 points (Fig. 3.). The first point is Gangnam-gu, the area has the form of downtown, which shows yearly maximum NDVI is about 0.2. We choose this area because it can represent characteristics of pixels with low NDVI, like urban area. The timeseries of LST shows about 270~318 K temperature range and reflects change of weather well. The second point is Bukhan mountain, the forest area, which shows yearly maximum NDVI is about 0.6. It can represent characteristics of forest area with high NDVI. The timeseries of LST shows about 267~312 K temperature range and reflects change of weather well and it shows lower change of temperature in timeseries. Generally, because humidity in vegetation effect to temperature, if there are many vegetations, temperature change range is lower than the place with less vegetation. Therefore, there are some temperature difference but LST trend of each algorithms were shown very similar and they reflect change of season well.

Fig. 3. Timeseries of LST. left: Gangnam-gu, right: Bukhan mountain.

4. Validation

1) Spatial representation

Generally, it is accurate to validate data with data which are collected by actual measurement device. But validation of satellite data is difficult with actual measurement data because there is no device which can covered satellite observation area at one time. Therefore, validation of satellite data is often replaced with other satellite data. To validate LSTs in this study, we used MODIS LST data. MODIS LST data used in this study is MOD11_L2 which has 1 km spatial resolution and 1-day temporal resolution. As I explained, spatial resolution of LSTs in this study are 30 m. So, it is necessary to check Spatial representation of LST in this study.

Fig. 4. Window-size test of LSTs.

As the first step of validation, we check spatial resolution of each LST by window-size tests (Fig. 5). From 3×3 to 27×27 window-size which were calculate average in each window-size were tested, Root Mean Square Error (RMSE) of 4 algorithms were showed Price has about 3.7 K, Becker et al. has about 2.8 K, Prata has 2.2 K and Coll et al. has 2.5 K. Bias were showed Price has about 0.5 K, Becker et al. has about 1.8 K, Prata has -0.5 K and Coll et al. has -0.7 K. RMSE and bias appeared similar values regardless window-size. That means MOD11_L2 and Landsat-8 LST can be matched and we can use MOD11_L2 data to validate Landsat-8 LSTs.

Fig. 5. Validation in NDVI cases.

We validated 4 LSTs in 5 NDVI cases, in 0 ~ 0.6 range (Fig. 5). Every LST has similar RMSE result but there are some differences in each case. In the lowest NDVI case (0~0.1), RMSE are the lowest values and in the third case (0.2~0.3), RMSE are the highest values. In case of bias, Coll et al. shows the lowest values, and Becker et al. shows the highest values.

After NDVI case test, we directly validated LSTs with MOD11_L2 (Fig. 6). Landsat-8 LSTs and MOD11_L2 data were match-up using their latitude and longitude data. Fig. 6(a) shows result of Price and R value is 0.9657, Fig. 6(b) shows result of Becker et al. and R value is 0.9566, Fig. 6(c) shows result of Prata and R value is 0.9643, Fig. 6(d) shows result of Coll et al. and R value is 0.9645. With this result, we think that all LST values are very reasonable.

Fig 6. Validation results of LST retrieval algorithms.

In comparison of RMSEs, Fig. 6(a) RMSE is 2.953 K, Fig. 6(b) RMSE is 3.387 K, Fig. 6(c) RMSE is 2.160 K and Fig. 6(d) RMSE is 2.396 K. And in each graph, they are very similar shape but result of Coll et al. shows some noises where pixels of Landsat-8 LST are over 320 K. With these results we think that Prata equation is the most suitable algorithm in this study.

5. Conclusions

In this study, we tried to find the most suitable LST algorithm to South Korea. We used Landsat-8 data which has high spatial resolution because South Korea has very complex and has variety land type. We used 4 algorithms for high-resolution satellite data, Price, Becker et al. Prata, Coll et al. which were developed by AVHRR data and is used with Landsat-8 data. We compared each algorithm to check they retrieve LST reasonable and they were showed over 0.98 of R values. We checked them by time series in 2 actual measure points, all LSTs showed they reflected change of seasons. After checking LST values, we tried to validate LST with MOD11_L2. Because spatial resolution of Landsat-8 and MODIS, we tested several window-size cases. The result shows there are few differences of RMSE and bias, and we can validate directly Landsat-8 LST and MOD11_L2. The result of validation shows LST of Prata equation has the lowest RMSE in whole scene (2.16 K) and each NDVI cases (1.6~8 K). As a result, we decided LST of Prata equation is the most suitable LST in this study.