4.5 Performance of MODIS LST for the study area
4.5.1 Summer surface temperatures
The scene recorded by the thermal imaging system (Fig. 3.5) features a wide range of surface covers that are typical for the wider area around the study area. However, it cannot necessarily be assumed representative for the much larger target area of the MODIS sensor (Fig. 3.7). A comparison of satellite and terrestrial measurements is still meaningful, particularly if weekly averages are considered, where spatial differences are strongly reduced. The available MODIS L2 LST data points and the almost continuous record of the average scene temperature from the thermal imaging system are displayed in Figs. 4.26 and 4.27. It is evident, that the average surface temperature in the footprint of MODIS depends on the fractions of wet and dry areas. For the comparison, the spatial average of the scene recorded by the thermal imaging system is used, while we keep the spatial variability of the weekly average surface temperatures (Fig. 4.19) in mind.
0 10 20
4-Jul 11-Jul 18-Jul 25-Jul 1-Aug 8-Aug 15-Aug
-2 0 2 4 K MODIS L2 LST
Thermal imaging system °C
deviation of averages
Figure 4.26: Land surface temperature measurements derived from the MODIS L2 LST product (black diamonds) and average scene temperature (blue line) obtained from the terrestrial thermal imaging system for the study period in 2009. The bar diagram displays the deviation between the weekly average tem- peratures inferred from the satellite record and the “true” value for the scene derived from the thermal imaging system.
terrestrial measurements are below 2 K. In the first three weeks, frequent clear- sky conditions lead to an excellent data density of MODIS LST measurements. In the first week, the deviation of 1.3 K of the weekly averages is largely caused by the three satellite measurements that are significantly warmer than the ter- restrial measurements, which can most likely be explained by different fractions of snow in the scenes of the satellite and the thermal camera. In the following period from 11 to 18 July, five days feature medium to excellent data densities of MODIS LST, while two on average colder overcast days are not represented at all, resulting in a net positive bias of 1.3 K of the satellite record. With a relatively balanced data density, the third week from 18 to 24 July shows only an small offset of 0.5 K. The following three weeks are dominated by overcast conditions (see Fig. 4.19) causing prolonged periods without satellite measure- ments, so that the agreement within 2 K of the weekly averages of satellite data and terrestrial observations appears rather fortuitous.
Except for the first week, the study period in 2008 is characterized by more frequent cloudy conditions compared to 2009, which results in a smaller density of satellite measurements. More pronounced deviations between satellite and terrestrial measurements are observed, for which two reasons can be identified. Firstly, the overrepresentation of cloud-free periods with on average warmer surface temperatures can lead to a strong positive bias of the satellite averages, particularly if the averaging period is almost entirely dominated by clouds, while
-10 0 10 20 -4 -2 0 2 4 K
24-Jul 31-Jul 7-Aug 14-Aug 21-Aug 28-Aug 4-Sep 11-Sep
°C MODIS L2 LST MODIS erroneous measurement Thermal imaging system deviation of averages
Figure 4.27: Land surface temperature measurements derived from the MODIS L2 LST product (black diamonds) and average scene temperature (blue line) obtained from the terrestrial thermal imaging system for the study period in 2008. Data points diagnosed as erroneous measurements (see text) are marked by crosses. The bar diagram displays the deviation between the weekly average temperatures inferred from the satellite record and the “true” value for the scene derived from the thermal imaging system.
the available satellite measurements stem from short cloud-free periods. Sec- ondly, there exist measurement errors in the MODIS data set, some of which are obvious, such as a measurement of -20.5◦C obtained on 9 August. At this day,
terrestrial observations of the cloud fraction cf (www.eklima.no, 2010) indicate fully overcast skies, so that presumably cloud top temperatures are measured instead of land surface temperatures, which range around +5◦C according to the record of the thermal imaging system. A significant number of satellite measure- ments is considerably colder than the terrestrial measurements, which indicates at least a partly contamination by cloud top temperatures. These measurement errors can invoke a serious bias of the weekly averages, as they are the only data points in prolonged cloud-covered periods and thus receive a strong weighting in the employed incremental averaging procedure. The erroneous measurement on 9 August would for example result in a negative bias of more than 9 K of the weekly average, which is clearly intolerable for the purpose of permafrost mon- itoring. The possible contamination by cloud top temperatures is supported by the quality flags provided in the MODIS L2 product, which indicate that more than 99% of the data set may be affected by nearby clouds. The calculation of meaningful weekly averages is obviously no longer feasible, if only few data
points are employed, so that we only discard cases with extreme deviations be- tween the two data series. Therefore, we define a data point a measurement error if the satellite measurement is more than 10 K colder than the terres- trial measurement. These data points are excluded from our evaluation of the weekly averages, which in total concerns four measurements with temperatures between -7.0◦C and -20.5◦C. Particularly in the second half of the 2008 study
period, many more satellite measurements exist, that are considerably colder than the terrestrial measurements, which suggests an admixing of cloud top temperatures. However, it does not seem appropriate to a priori exclude them as measurement errors due to the large scaling gap between the terrestrial and satellite footprints. In some cases, e.g. for the week between 21 and 28 August, these probably cloud-influenced, cold-biased data points are found to balance the warm-bias due to the overrepresentation of clear-sky periods to some extent. In September, with almost entirely overcast conditions, cold-biased data points dominate, so that the satellite average is more than 4 K colder than the average of the terrestrial observations. As the available MODIS LST observations are mainly below 0◦C after 4 September, this would lead to a false estimation for
the onset of freezing, which in reality occurs after the end of the study on 15 September.