The RSS Lower Tropospheric Dataset has been upgraded to V4.0
The RSS lower tropospheric temperature dataset (TLT) has been upgraded to Version 4.0. The upgraded processing system uses the same techniques to optimize the adjustments for changing measurement times that we used for the middle tropospheric dataset (TMT V4.0). The upgrade is described in a paper accepted for publication in the Journal of Climate. (Published online on June 26, 2017, https://doi.org/10.1175/JCLI-D-16-0768.1). Unfortunately, the paper is behind the JCLI paywall until it is formally published in the journal, despite the fact that RSS paid the extra fees for open access.
What is the “TLT” dataset?
TLT stands for “Temperature Lower Troposphere”. We use measurements from microwave sounding instruments mounted on weather satellites to assemble a long-term record of atmospheric temperature. The “TLT” product is a weighted average temperature of a thick layer of the atmosphere extending from the surface of the Earth to an altitude of about 7000 meters. This is the part of the atmosphere that we live in, and thus it is an important indicator of global change.
Microwave sounders infer temperature by measuring the intensity of microwaves emitted by oxygen molecules, which “glow” in the microwave part of the spectrum. The intensity of this "glow" depends on the temperature. By observing at different microwave frequencies, different layers in the atmosphere can be monitored.
The long-term record is constructed by combining measurements from 16 different satellites (9 MSU satellites and 7 AMSU satellites). The measurements begin in late 1978 and continue to present day. The dataset construction is challenging because the satellites are not perfectly calibrated and many of the satellites’ orbits drift over time, causing them to make observations at different times of the day throughout the course of their lifetimes. And, obviously, the temperature typically changes with time of day. This “diurnal” drift needs to be removed from the data to prevent spurious changes in the observed temperature record.
What is the main issue addressed in the paper?
The paper describes an important update to the methods we use to construct our lower tropospheric temperature (TLT) data record. The methods used in this TLT update are very similar used for our latest version of the middle tropospheric temperature (TMT).
The most important differences between the previous version (Version 3.3) and the new version (Version 4.0) are:
1. The method used to make adjustments for drifting satellite measurement time was changed. In the new method, the model based diurnal cycle climatology used for these adjustments was optimized so that differences between satellites making measurements at different times of day are removed in a more accurate manner. The new method steps away from adjusting the satellite data using model output by incorporating infomation from the measurements themselves. (This will be discussed in greater detail below.)
2. Inter-satellite offsets are now calculated separately for land and ocean scenes. This prevents possible errors over land, where the adjustment for changing measurement times are large, from adversely affecting measurements over the ocean, where the diurnal cycle is close to zero.
3. Several periods of suspect data were removed (see below for more details).
4. Two new satellites, NOAA-19 and METOP-B, are now included in the processing. This serves to reduce sampling error as well as any remaining errors due to the diurnal adjustment during the last part of the record, where NOAA-18's measurement time is drifting rapidly.
How does Version 4.0 compare to the earlier version?
The new version shows more warming that the previous version at most locations. This difference is summarized in the plots show below:
Fig. 1. These two figures show a comparison of large-scale time series between the new version (V4.0) and the previous versions (V3.3). The top panel shows the near global (70S to 80N) time series. The bottom panel shows the tropics (30S to 30N). In both regions, the long-term warming is greater in the new version. The amount of the increase is larger for the regions outside the tropics, so the global increase is larger.
The most important differences between the two datasets occur during 2000-2007, where the new version shows more warming due to the effect of the optimized diurnal adjustment on NOAA-15, and after 2012, where the removal of data from NOAA-15 and AQUA, combined with the new data from NOAA-19 and METOP-B, results in increased warming. The spike in November 1980 is due to an increase in the amount of data available for that month. This occured when we changed the source of the raw data to the NOAA CLASS System from the earlier data source (which we received in tape(!) in the late 1990s). It looks dramatic, but has little effect on the long-term trend.
The difference in maps of the trend is easiest to see if we flip back and forth between the two images.
Fig. 2. These two maps show the geographical patterns in the trends for versions 3.3 and 4.0. The overall pattern of warming is similar for the two datasets, with the new version showing more warming than the old version in most regions outside the deep tropics.
When I go to the time series browser, I see different trends than are in the paper and in Fig. 1 above. Why are they different?
The figures and trends in the paper are for January 1979-December 2016. The trend values on the website (currently 0.184 K/decade through May 2017) are updated each month as new data come in. For each update, additional data are added to the end of the time series (obviously) and both the optimized diurnal cycle and the "merging parameters" (inter-satellite offsets and the target factors) are recomputed using the additional data. This means that new data can theoretically change the entire time series, though in practice these changes are small. Another small difference is that the trends in the paper and Fig. 1 above are for 70S to 80N, while the trends on the webpage are computed for 70S to 82.5N (this is for the sake of historical consistency) . There is not much area between 80N and 82.5N, but it is warming very rapidly. All of the monthly averages for 2017 are above the 1979-2016 trend line, so they serve to pull the trend value upward as they are added.
What motivated the update?
The new study was mostly motivated by a problem in the differences between co-orbiting satellites making measurements at different times of the day. We routinely monitor these differences as part of our quality control procedures. These differences were particularly large over land, leading us to suspect that there were problems with using the “raw” model-based diurnal cycle to correct for changing measurement times. Over the last few years, we have upgraded the diurnal cycle adjustment for both TLT (described here) and TMT (described in an earlier paper, http://journals.ametsoc.org/doi/10.1175/JCLI-D-15-0744.1, which has been available for the last 15 months).
What method did you use to fix this problem?
We started with the model-based diurnal cycles at each location, and then made small, latitude dependent adjustments to the model diurnal cycles. The adjustments were determined by minimizing the differences between measurements made by different satellites at different times of day. In other words, we assumed that the differences between satellite measurements were caused by errors in the model derived diurnal cycles and adjusted these diurnal cycles to make the differences go away. We then used these new, optimized diurnal cycles to adjust all the satellite data, even when only one satellite was operating. Different adjustments were made for land and ocean scenes. For the most part, these adjustments moved the warmest part of the day over land to a time slightly later in the afternoon.
Fig. 3. A comparison of the “raw” model-derived diurnal cycle and the adjusted diurnal cycle for near global averages, land and ocean separately. The adjusted diurnal cycle over land has an afternoon peak slightly later in the afternoon than the raw model results. Another important change is the reduction in slope in the pre-dawn hours from midnight to ~7:00 AM.. (Adapted from Figure 5 of Mears and Wentz, 2017)
One of the criticisms of our method of diurnal adjustment is that it is based on climate model output. Some people dislike the idea of using information from models to adjust measurements. We note that this new method moves a step away from using the model results directly. In the new method, the measurements themselves are used to deduce and fix errors in the modeled diurnal cycles.
The “raw” and adjusted diurnal cycles don’t look so different. Why were the final results changed as much as they were?
The biggest effect of the adjusted diurnal cycles is to reduce the amount of adjustment made to NOAA-15 as its equatorial crossing time drifted from about 7:30 in 1998 to about 4:00 in 2010. For both the PM and AM passes of the satellite, the adjusted diurnal cycle shows less change during these times. We should note that the largest change is in the morning. Because of this, in the old version, we were overcorrecting early morning cooling for NOAA-15. Of course, changing the diurnal correction also changes the results for all the other satellites, but the largest effect was for the time period when NOAA-15 was drifting rapidly (1999-2007), both because of the time of day that is was operating, and because of the lack of other co-observing AMSU satellites (prior to the launch of AQUA).
What other changes did you make to the data?
We also found that NOAA-15 and AQUA appeared to undergo calibration drifts late in their respective lifetimes that were unrelated to the diurnal adjustment. Part of the data near the end of these satellites' missions is no longer used. (See the paper for more details.) We also found anomalous behavior for NOAA-18 at the beginning of its mission. The periods excluded in V4.0 are:
NOAA-15 : Excluded after Dec 2011.
AQUA: Excluded after Dec 2009.
NOAA-18: Excluded before Jan 2009.
These “which data were used” changes are very similar to those that were performed for TMT, with the exception of the exclusion of NOAA-18 data before 2009. The NOAA-18 problem apparently only affected the near-limb views that are used for TLT, but not TMT.
In the figure below, we show plots of the inter-satellite differences for the AMSU satellites after various levels of adjustment have been applied. When the raw model based adjustments are used (second and third rows down), there are still substantial differences between the co-orbiting satellites, particularly between satellites that drift a lot in local observing time. After using the optimized diurnal adjustment, these differences are mostly removed. All of the choices we made were driven by the desire to reduce inter-sallite differences.
Fig. 4. Time series of inter-satellite differences for the AMSU instruments after different levels of adjustment and data removal were applied. (Figure 4 from Mears and Wentz, 2017)
How do the new results compart with the UAH versions of the TLT dataset?
Our new dataset shows more warming than either the current version of the University of Alabama, Huntsville (UAH) data (V6.0), or the previous version (V5.6). The plots below show a comparison between the three datasets.
Fig. 5 Global (A) and tropical (B) times series for RSS V4.0, and UAH V5.6 and V6.0. The green lines show the differences, offset to make them easier to evaluate (adapted from Fig. 9, Mears and Wentz 2017).
The newer version of the UAH data agree fairly well with RSS V4.0 from 1979 to about 1998 over the entire globe, and warms slightly relative to RSS V4.0 in the tropics during this period. After 1998, the RSS version shows more warming than UAH V6.0 in both regions.
Note that neither NOAA STAR nor the University of Washington products are TLT products, so we are unable to compare with those data sources.
What about comparison with radiosondes (weather balloons)?
In our paper, we compare our results to the four homogenized radiosonde datasets that are currently being updated and are available in gridded or individual station form (we have shown that it is important to sample the satellite data at radiosonde locations when doing these kinds of comparisons). In all the radiosonde datasets, the measurements have been "homogenized" in an attempt to remove the effects of changing radiosonde instrumentation, siting, and observing practices. The UAH researchers like to say that their data agree better with radiosondes. This depends on which radiosonde dataset is under consideration, and what one means by "agree better." We did find one thing that the radiosondes datasets all agree on. During the main period of disagreement between RSS V4.0 and UAH V6.0 (i.e., ~1998-2007), a comparison with homogenized radiosonde datasets shows generally better agreement with RSS V4.0 than UAH V6.0.
Fig. 6. Difference time series between near-global satellite and homogenized radiosonde TLT temperature anomalies. The top panel shows difference time series for UAH V6.0, and the bottom panel shows the differences for RSS V4.0. In all cases, the satellite data are sampled only at locations that have valid radiosonde data for the year, month, and dataset in question. The resulting time series are then smoothed to remove variability on time scales shorter than 6 months.
In the top panel, I drew a green line (by hand, not a fit) that represents the difference trend between UAH and the various radiosonde datasets for 1998-2007. In the bottom panel, I drew a pink line that represents the difference trend between RSS V4.0 and the radiosonde. The green line in the bottom panel is identical to the one in the top panel, just moved down for comparison. At RSS, we do not use radiosonde data to guide our choices when constructing long-term satellite datasets. This is done in order to try and keep the two types of data independent of each other. That being said, these results do suggest that our changes to the AMSU data are supported by the radiosondes (RSS V3.3 also shows a large cooling signal relative to the radiosondes over the 1998-2007 period). Note that all satellite data warm relative to radiosondes before about 2000, and then cool after about 2000. We don't know if this overall pattern is due to problems with the radiosonde data, with the satellite data or (most likely) both.