Does NOAA have "Ghost Stations" for US Temperatures?

The Epoch Times (ET) is a newspaper operated by the Falun Gong cult, formed in 1992. The cult is (understandably) opposed to the Chinese Communist Party, but recently it has made some inroads into the US, where it supports far-right political agendas; its ET newspaper has promoted a number of conspiracy theories involving QAnon, anti-vaccine propaganda, and climate science denial. In a recent "premium report," the Epoch Times claims that the U.S. Temperature dataset operated by NOAA contains data from "non existent temperature stations" and asserts that there are "hundreds of ‘ghost’ climate stations" that are no longer operational, and data is just filled in from surrounding stations, with the implication that they do so to support Biden's climate policies.

The Fake Problem

The article relies mostly claims made by influential climate-tweeter named John Shewchuk and climate-blogger Anthony Watts. Strangely, almost all their claims have to do with the now-deprecated USHCN dataset, and as we'll see, they had to use this defunct dataset in order to make their lies seem plausible. The USHCN v2.5 dataset was built on a 2.5° longitude by 3.5° latitude gridded analysis of monthly temperatures from the stations in the USHCN network. This was the standard NOAA dataset for CONUS temperatures until about 2014. 

NOAA increased the number of stations until it reached 1218 stations in 1957, but then in 1990, the number of stations began to decrease as equipment failed and people retired. There are now far fewer stations in the network as about 30% of these stations have closed. When station closures happen (or if there is a period of time where a station doesn't report, the station temperature is estimated from surrounding stations. This may seem weird, but it makes sense. Suppose a 2.5° x 3.5° grid-cell has 4 stations reporting such that Tavg for the grid-cell would be the average temperature of 4 stations. If one station doesn't report one month (or if the station closes), you could simply average the remaining 3 stations. If, however, you take that Tavg and estimate the temperature of the non-reporting station to be that average temperature, you haven't change Tavg for the grid-cell at all. So for purposes of continuity in the dataset and making station-related calculations easier, USHCN interpolated values for the stations that closed, but they also marked these estimated values with an "E" for "estimated" in station data. These interpolations are not counted as sample data, and they cannot significantly affect temperature trends. Since NOAA calculates trend using area-weighted means on a 2.5° x 3.5° grid, the trends aren't affected by these infilled station values.

What Shewchuk is saying is that NOAA is counting these interpolated values as station-reported data, which they are fabricating to manipulate data and create spurious warming trends in US temperatures According to him, these stations “are physically gone—but still report data—like magic.” He refers to these stations as "ghost" stations (Tony Heller calls them "zombie" stations). Shewchuk (and Tony Heller) is essentially saying that  NOAA is closing USHCN stations and then fabricating data for those stations to fabricate global warming and support Biden's policies. But it's not difficult to show that in fact Shewchuk and Heller are flat out lying about what NOAA is doing.

USHCN is Defunct and is No Longer Used by NOAA

As of 2014, NOAA deprecated USCHN that Shewchuk and Heller use to make their accusations. There are public-facing data tables for the USHCN that are still populated by stations in the old network, but they only exist for historical purposes, and this is where Shewchuk and Heller are getting their data. In the place of USHCN, NOAA now uses two CONUS datasets.

1. USCRN (114 Stations). In 2005, NOAA introduced the USCRN network, which is made up of 114 ideally-sited, all-rural stations for monitoring changes in CONUS temperatures. From 2005 to 2014, NOAA monitored both USHCN and USCRN and documented a remarkable agreement between these two datasets over that decade. The two datasets were in close agreement with each other, even though USHCN had several times more stations than USCRN. Because of the nature of the USCRN network, the dataset requires no bias correction; it's essentially "raw" temperatures.
2. nClimDiv (10,000 Stations). In 2014, NOAA deprecated USHCN and replaced it with a newer nClimDiv dataset, which is built on a finer grid (about 5 km by 5 km) and is derived from the global GHCN-daily network. This network contains "several thousand more stations" (nClimDiv has over 10,000 stations) that are available in GHCN-daily. In addition, nClimDiv uses stations in Canada and and Mexico near US borders for the interpolation of US anomalies in areas near the US border. Any plots of USHCN station data following 2014 will not be representative of the US.

You would think that if Shewchuk and Heller were actually interested in evaluating NOAA's data handling and practices, they would use nClimDiv. It has over 8x more stations than USHCN ever had, but Heller and Shewchuk essentially ignore this. Instead, they are a decade out of date. It's true that stations have dropped from the now-defunct USHCN, but Heller and Shewchuk are lying about what's actually going on. NOAA has actually increased the number of reporting stations from 1,218 to over 10,000 stations, and nClimDiv doesn't use the interpolation method that Shewchuk and Heller are complaining about. So even if you were to count interpolated station values marked with an "E" as "fabricated" data (it's not), this not happening in either of the two current datasets. There are literally no ghost or zombie stations in the USCRN or nClimDiv networks. And their accusations about USHCN are absurdly wrong. We can essentially prove that by comparing results from NOAA's two current datasets.

USCRN and nClimDiv Agree

Since 2005, we have been able to compare USCRN with nClimDiv. As of now (updated 2025), we have 20 years of data to compare. USCRN has rural-only, ideally-sighted 114 stations requiring no bias correction, and nClimDiv is a long-term, homogenized dataset with 10,000 stations. And yet for the years they overlap, the two are nearly indistinguishable. In fact, USCRN shows marginally more warming than nClimDiv. Since CONUS trends are calculated from each using area-weighted means in both datasets, the number of stations doesn't really matter. Any sized network from 114 to 10,000 stations produces similar results, with or without homogenization and with or without infilling estimated station values.

When USHCN was operable, it too agreed closely with USCRN during the decade from 2005 to 2014. Whether this network contained 700 stations for 1200 stations, it still contained more stations than USCRN and fewer than nClimDiv, and it produced results that agreed with USCRN and nClimDiv, neither of which use the interpolation method Shewchuk is complaining about. In other words, it's clear that there are no practices by NOAA involving USHCN that are causing any spurious warming trends for CONUS temperatures. In fact, USHCN infilling has a negligible impact on USHCN trends..

USHCN Infilling Is Negligible

Zeke Hausfather about 10 years ago (around the time USHCN was deprecated) evaluated the impact of infilling in the old USHCN network. He has shared this in several places, but I'll share it from a post he contributed to Judith Curry's blog so you can see that this is not something that even informed contrarians really take issue with. Here Hausfather calculates how CONUS temperatures are affected with and without the infilled data. The difference between the two is plotted below with Tmin and Tmax temperatures.

After 1915, the Effect of Infilling is Negligible

ln the above graph, after about 1915, the trend in infilling adjustments is essentially flat, meaning that this has virtually no effect on CONUS temperature trends. As Hausfather explains, infilling essentially just adds the climatology of the station with missing data to the area-weighted anomaly of nearby stations, making the effect essentially the same as area-weighted averaging. It's essentially saying area weight for surrounding stations increases to "fill in" the area during times when the station doesn't report data.  Whether you use a network with 114 stations (USCRN) or with 10,000 stations (nClimDiv) you get very similar results for US temperatures, so so we shouldn't expect that losing a few stations out of a network with 1200 stations (USHCN) and then using infilling would change US temperature trends in anyway. Infilling quite simply has no significant effect on CONUS temperature trends, even within the now-deprecated USHCN network.

An Illustration

I thought it might be helpful to illustrate the concept of infilling and why it's not a tactic developed by those with nefarious intent, as Shewchuk and Heller want you to believe. I think I can illustrate Hausfather's point that infilling is statistically identical to calculating area-weighted averages. Let's set up a hypothetical grid with one station at the center of each cell in the network; this grid has 117 cells with 117 stations (similar to the USCRN network with 114 stations). We'll call this USCRN-like grid a "large cell grid" with the temperature of any cell in that grid the X temperature.

Hypothetical USCRN-like Grid with 117 X Stations

If you record temperatures at each X in the grid, then you can easily calculate the average temperature of the entire grid by averaging the values for each station. If your grid has cells of different sizes, you can multiply each cell temperature by the area of the cell and calculate an area-weighted average for the entire grid. But let's say we want a grid with finer resolution. Let's break up each of the 117 cells above in to 9 cells. If we do that, we can then "infill" these smaller cells with the temperature recorded in the center cell. These "infilled" values are not sample data; they do not increase the value of N, and they would be noted with an "E" by NOAA.  But if we do use infilling in this way, we'll get the exact same answer for the average temperature of the grid. For instance, see below. The average temperature of this grid is 20.50 C. Whether you count the center station reading only or all the infilled values, you're going to get the exact same average temperature for the grid.

Simple "Infilling" of X-station Data to a Higher Resolution USCRN-like Grid

We haven't changed the average temperature at all by infilling the X temperature into these smaller cells. However, we can see weaknesses in this approach. For instance, if the X temperature for one cell is 15 C and the X temperature for an adjacent cell is 24 C, it's not likely that the transition from 15 C to 24 C is abrupt; there's more likely a more gradual change, and in a finer grid there's likely to be an incremental change between 15 C and 24 C. How can we handle that? Well in our hypothetical network we also have other weather stations that can help us fill out the temperature of the whole grid with finer detail. But if we put these stations onto this finer grid, not every cell is filled. So let's construct a "small cell grid" in which 9 small cells corresponds to the 1 large cell above. This will be analogous to a USHCN-like grid. It contains X temperatures (same as above) and O temperatures from stations not in the 117 in my previous example. Essentially, we've added 468 USHCN-like "O stations" for a total of 585 stations, leaving 468 small cells with no stations.

Hypothetical USHCN-like Grid with 585 stations and 468 empty cells

If we use all the USCRN-like X stations and the USHCN-like O stations in a small cell grid, we can get better resolution, but since we've adopted a finer grid, we now have "empty" cells. The process of "infilling" simply fills in those empty cells with data extracted from the surrounding cells. In this hypothetical USHCN-like small cell grid, I used the same X station data as above and added hypothetical O station values. Then I infilled the empty grid cells by simply averaging the temperatures recorded at stations surrounding the empty small cell. Again, these infilled values are not counted as reporting stations and they do not increase the sample size of the network. They simply allow for an area-weighted mean to be calculated for the surface area contained in the network. And while this is clearly an oversimplification of infilling within the USHCN network, I think it illustrates the rationale behind it quite well. In my hypothetical example, the average temperature of this small cell grid is 20.51 C, essentially unchanged from the large cell grid above. This infilling does not generate new data. 

Infilling empty cells in USHCN-like Grid from Surrounding Stations

Now what happens if a small cell loses its station for a period of time? Well, the best way to handle that is to infer the temperature of that small cell from the surrounding cells. In effect, all that's happening is that the area weight of the surrounding cells expand to fill in that cell when data from that cell is missing. And notice in reality here, given that there are more, smaller grid cells in the USHCN-like small cell grid, there's actually a reduction of infilling compared to the USCRN-like large cell grid, if you were to break up large cell grid with the resolution of the small cell grid. If we impose a small cell grid on the large cell grid above, then infill based on the 117 stations in the USCRN-like large cell grid, then 936 cells are infilled, or about 89% of the small cells. If we use the USHCN-like small cell grid with all 585 stations, then only 468 cells are infilled or about 44% of the small cells. This is clearly an improvement over the large cell grid with only 117 stations. Because you're inferring temperatures from more stations instead of from fewer stations, overall you're getting a better understanding of the geographic distribution of temperature data without a significant change in the overall trends in temperatures. However, infilled values are not data points. The sample size N is the number of reporting stations; infilling only serves to provide a better area-weighted average.

Conclusion

It would seem there are multiple layers of dishonesty on the part of Shewchuk and Heller. There are no "ghost" or "zombie" stations in NOAA's datasets, USCRN or nClimDiv, and what they are calling "ghost" stations in the USHCN is intentionally deceptive, since on of the closed stations fabricate or report station data. Interpolations are marked for what they are. So there are multiple-levels of lie on the part of Shewchuk and Heller:

1. It's a flat out lie to refer to interpolated station values marked with an "E" as if they are data reported (or fabricated) by stations. 
2. USHCN is a defunct dataset that is no longer maintained, and neither nClimDiv nor USCRN use the infilling they wrongly refer to as "ghost"/"zombie" stations.
3. Comparing nClimDiv (or USHCN before 2015) to USCRN shows that they agree with each other. Since NOAA uses area-weighted means, CONUS trends are largely unaffected by station dropouts. Networks with 114 stations agree with networks with 10,000 stations. In fact, USCRN shows marginally more warming.
4. The impact of infilling on USHCN when it was active can and has been quantified, and the impact is negligible. Infilling does not generate spurious warming trends.
5. Infilling is statistically identical to using an area-weighted mean. 

Beyond this, Shewchuk and Heller both produce graphs of CONUS temperatures using simple averages of station data. This is the worst way to do calculate CONUS mean temperatures, but it's also statistically identical to infilling the entire area of CONUS that lack thermometers with their station means. In other words, Shewchuk and Heller are both "infilling," but instead of infilling areas with the closest available temperature data, they infill these areas with the simple mean of all reporting stations. 

The lies here are not limited to Shewchuk and Heller. ET knows about the update to nClimDiv, but it quotes Anthony Watts essentially lying about the newer dataset. Watts is quoted as saying, "The USHCN data set and the [new] nClimDiv climate division data set [which uses the same stations and has the same problems]." I suppose it's superficially true that nClimDiv uses the same stations current in USHCN, but Watts neglects to mention that nClimDiv also adds thousands more stations not used by USHCN. Neither Watts nor Shewchuk nor Heller nor ET have demonstrated that there are any "problems" with any algorithms and corrections that nClimDiv actually uses that would cause nClimDiv to report any spurious trends. There is simply no evidence of any "ghost" or "zombie" stations that are being manipulated by NOAA to affect Biden's climate policies, or any policy whatsoever.