Facilitating long-term 3D sonic anemometer measurements in hemiboreal forest ecosystems

Eddy covariance flux measurements and micrometeorological methods have become standard in assessing the exchange of greenhouse gases (GHG) and reactive trace gases between forest ecosystems and the atmosphere (Baldocchi, 2014; Burba, 2013; Keronen et al., 2003; Krasnova et al., 2019; Noe et al., 2011). For the production of meaningful results, controlled series of long-term measurements of the forest ecosystem-atmosphere exchange are needed (Bravo et al., 2019). The flux tower sites’ footprint area depends on accurate measurement of 3-dimensional turbulent air movements over the course of a year and is a necessary parameter in the assessment of the capacity of the sites’ forest ecosystems to exchange matter and energy with the atmosphere (Kljun et al., 2015; Vesala et al., 2008). This allows for linking forest growth to flux measurements over a spatial explicit area (Kollo et al., 2021). The quality of forest carbon sequestration estimates depends therefore on the data availability over the whole seasonal cycle, the amount of which needs to be maximised.

The challenge to keep up measurements in northern climate regions during winter months with the occurrence of snow and ice has been discussed by Makkonen et al. (2001). The problem and conditions of ice formation on measurement equipment and infrastructure has also been reported in earlier studies (Makkonen, 2013; Makkonen & Laakso, 2005). Needed corrections for flux measurements when anemometer heating took place have been developed e.g., for measurements in Siberia and when assessing winter carbon dioxide (CO₂) fluxes (Kittler et al., 2017; Panov et al., 2021). The reported measurements hint at changes in turbulent flows but no one reports a stronger impact on horizontal wind parameters (Goodrich et al., 2016; Kittler et al., 2017; Skelly et al., 2002).

Measurement of turbulent air movements utilising 3D ultrasonic anemometers has become a standard task at flux tower sites. In winter conditions, icing of the ultrasonic anemometer sensor heads leads to loss of data and introduces measurement gaps of some short periods up to months given that winter conditions may last from October to April. Different heating strategies were reported in a comparison study. Applying a constant heating strategy improved the annual data coverage by 64% and led to a data loss of 3.2% by removing biased data due to the heating. Utilising an intermittent heating strategy reduced the loss of data even further (Goodrich et al., 2016). Furthermore, boreal forest ecosystems may experience night frosts throughout the year and short-term heating might prevent night-time ice formation.

Different brands of ultrasonic anemometers have a possibility to heat the sensor heads to melt ice or snow and keep them operable also during freezing periods. While heating is desired to be operable it also leads to biases in measurements because heat-driven buoyancy is part of the turbulent flow. Therefore, artificial added turbulence by heating should be minimised (Goodrich et al., 2016; Kittler et al., 2017).

In this technical paper, we aim to quantify the temperature ranges when icing occurs using statistical methods and online measured data from the SMEAR Estonia station. We hypothesised that 1) icing takes place at different temperature ranges, 2) we can determine the temperatures with the highest anemometer sensor head icing probability and 3) a regulation feedback loop algorithm can minimise data loss.

At the SMEAR Estonia station, we measured temperatures below zero degrees Celsius (Figure 2) from October to April, if measured above the forest canopy on the atmospheric mast. On ground level (2 m above the surface), temperatures can fall below zero for short periods also during May. Therefore, ice formation occurs frequently on ultrasonic anemometer sensor heads which are exposed to the ambient weather conditions during these months.

The probability of icing on the anemometer sensor heads was found to span from approximately +5 °C to −25 °C (Figure 3 and 4). Icing events with slightly positive temperatures happen due to a change from cold dry to warmer wet air masses when the anemometer sensor head is still below zero degrees. We calculated PDFs for each height (Figure 3) and each month when icing occurred (Figure 4) by pooling the data from all years. For all heights, the calculated mixture distributions were composed of two or three normal distributions (Wolfram Research, 2016). The PDFs maxima denote the temperatures or temperature ranges where icing of the anemometer sensor heads is most likely to happen. Given Figure 3, the highest probability of icing lies at temperatures between zero and −3 °C. The second temperature range with a medium probability for icing is between −5 °C and −15 °C, and finally the lowest probability but still with a distinctive peak in icing probability lies at the range between −15 and −25°C. The latter peak was most pronounced on the highest instrumentation levels (90 and 110 m) but also near the forest canopy at 30 m. Almost all instrumentation heights have mixed distributions with at least two or three normal distributions that can be expressed as N (μ, σ) samples. In Table 1, we summarize the parameters obtained from the fitting procedure of the PDFs to the data. For all instrumentation heights, except for 90 m, we found rather high Pearson’s χ² probabilities (> 0.5) that denote a good to very high (30 and 70 m > 0.9) confidence that the mixture models of the PDFs describe the data well, which confirms our hypothesis #1. Pearson’s χ² goodness-of-fit test compares a histogram of the data to a histogram of the distribution with the null-hypothesis H0: “The data are drawn from the fitted distribution” and rejected when p < 0.05 falls below the significance level (Wolfram Research, 2010). Only at 90 m, the probability was weak (< 0.05). A reason for that might be related to the situation that the data is not evenly well represented by continuous mixed normal distributions. However, for the sake of simplicity, we did not test different blends of mixed distributions where also other than normal distributions can be included. The lumped data, where all the data from different instrumentation heights were combined into one dataset also achieved very low Pearson’s χ² p-values (< 0.0005) which indicates that continuous normal distributions cannot map these combined datasets well. It might also reflect that the icing process is a rather heterogenous process and simply differs substantially between different instrumentation heights. Lumping all data together, therefore, fails to improve the prediction quality. However, even in the dataset combining all heights, the highest probability of icing was found at the range of zero to −5 °C and a second peak between −5 °C and −15 °C.

Figure 3

Figure 4

To get more insight, we further compiled the data with respect to the months when icing occurred and calculated PDFs for each month (Figure 4). Overall, in October and April, the months that usually mark the beginning and end of the ice forming period, showed only one peak of high probability, and the temperature range found was between +5 °C and −5 °C. November showed the largest deviation from the general pattern with four peaks, one between +2 °C and +1 °C, one at zero degrees, and one between −1 °C and −4 °C. November’s lowest temperature range was found between −5 °C and −15 °C with the lowest probability for icing. December followed with a single-peak pattern and the probability maximum was found at −3 °C, but employed a wide span from +5 °C to −15 °C. January employed a two-peak probability pattern with wide ranges of +5 °C to −13 °C for the peak with its maximum at −5 °C and the second peak lied between −13 °C and −25 °C with the maximum at −19 °C. February showed also a two-peak shape but its ranges were found narrower, +5 °C to −6 °C and −6 °C to −15 °C. March then moved back to a single-peak probability function pattern with its maximum at −2 °C and a span between +5 °C and −10 °C. Table 2 summarizes the PDF fitting procedure parameters for the monthly data and shows good to excellent matches for October, November, December and March (> 0.8). A medium good fit was found for February (> 0.5) but a small probability for March (< 0.3) and a very small probability for January (< 0.08). These findings support our hypothesis #2 that we can determine the most probable icing temperature ranges. Altogether, the highest probability for icing is slightly below zero degrees and mostly a problem in weather conditions which are expected to be more likely in Estonia with climate warming (IPCC, 2021; Jaagus & Mändla, 2014; Kupper et al., 2011; Lõhmus et al., 2019).

To ensure as small data loss as possible we analysed the time it takes for the ultrasonic anemometer to get its sensor heads ice free using the de-icing algorithm (Figure 1). Figure 5 shows a Box-Whisker chart where the length of sensor head heating for each month is visualised. We found that the median time of heating is around 30 seconds with a minimal time of 3 seconds and a maximal heating time of 72 seconds. We grade this result as very positive because assessing forest ecosystems’ fluxes from both eddy covariance and inverting the gas concentration profile measured at five instrumentation heights of the atmospheric tower in Järvselja over the whole year, there is need to decide how much data must be deleted and how many gaps filled in carbon cycle estimations because of the heating effect that introduces extra turbulence by buoyancy and thus leads to biases. The box width in Figure 5 is scaled in relation to the number of icing events in the given month and we can see that most ice formation events occur during December and January. Such short time periods allow us to create strategies of correction and compensation of added turbulence in the data processing beside the removal of heated data sets. Also, the determination of the flux footprint is, given such short-term icing events, based on very robust data even in winter. The short time intervals found confirm our hypothesis #3 and tell us that it cannot be considered to let the ultrasonic anemometer sensor heads remain frozen and risk long periods of data loss. The risk of biased GHG budgets due to heating periods of the ultrasonic anemometer sensor heads is very small or absent, if after the short heating period the data are discarded for some 5 or 10 minutes. It is further important to minimise the lack of data during winter conditions and especially when considering that the period with temperatures near zero degrees are important for the tree frost hardening and de-hardening processes at the onset and the end of winter and based on that determine the vegetation period.

Figure 5

We found that icing on micrometeorological measurement equipment takes place over a wide range of temperatures. The highest probability is near zero degrees, but it depends on the instrument height and the particular month. The lowest icing temperatures were recorded in January and February, but the largest variation in icing temperature ranges was observed in November. Given the possible time range of five months when frost and ice formation can occur frequently, it is necessary to estimate the times when heating of instrumentation is needed to ensure the lowest possible data loss. The median 30 seconds of heating to get the 3D sonic anemometer sensor head ice free proves the intermittent heating algorithm successful. The risk of missing out long time intervals during autumn, winter and spring would jeopardise the GHG budget calculations. To reduce the risk of biased carbon sequestration estimations, the forest ecosystems’ carbon and energy budgets need to be properly assessed. The quality of estimating the eddy tower’s footprint area has a high impact on the potential of applying the greenhouse gas budget and forest growth in carbon sequestration regulations. With the smart heating algorithm leading only to intermittent heating of the anemometer sensor heads, the risk of biased carbon sequestration estimations due to large data gaps is strongly reduced at SMEAR Estonia.