The thin border between cloud and aerosol: sensitivity of several ground

11 Cloud and aerosol are two manifestations of what it is essentially the same physical 12 phenomenon: a suspension of particles in the air. The differences between the two come from 13 the different composition (e.g., much higher amount of condensed water in particles 14 constituting a cloud) and/or particle size, and also from the different number of such particles 15 (10-10,000 particles per cubic centimeter depending on conditions). However, there exist 16 situations in which the distinction is far from obvious, and even when broken or scattered 17 clouds are present in the sky, the borders between cloud/not cloud are not always well 18 defined, a transition area that has been coined as the ―twilight zone‖. The current paper 19 presents a discussion on the definition of cloud and aerosol, the need for distinguishing or for 20 considering the continuum between the two, and suggests a quantification of the importance 21 and frequency of such ambiguous situations, founded on several ground-based observing 22 techniques. Specifically, sensitivity analyses are applied on sky camera images and 23 broadband and spectral radiometric measurements taken at Girona (Spain) and Boulder (Co, 24 USA). Results indicate that, at these sites, in more than 5% of the daytime hours the sky may 25 be considered cloudless (but containing aerosols) or cloudy (with some kind of optically thin 26 clouds) depending on the observing system and the thresholds applied. Similarly, at least 27 10% of the time the extension of scattered or broken clouds into clear areas is problematic to 28 establish, and depends on where the limit is put between cloud and aerosol. These findings 29 are relevant to both technical approaches for cloud screening and sky cover categorization 30 algorithms and radiative transfer studies, given the different effect of clouds and aerosols 31 (and the different treatment in models) on the Earth’s radiation balance. 32

Consequently, fundamental questions remain: What is the limit of visibility from which a 80 suspension of droplets must be considered cloud? Should this limit be set for an -average‖ 81 human eye, or can it be objectively established for some instrument as in Dupont et al. 82 (2008)? Or is it even reasonable to consider such a limit given that the aerosol/cloud particle   In general the distinction between a cloudy and a cloudless sky, and the separation between 113 cloud and aerosol, is appropriate for attribution studies and modeling radiative effects of 114 different climate forcing mechanisms, but imposing this classification may be unnecessary 115 (or inconvenient) in relation to new and advanced methods of observation and measurement. 116 If so, the distinction could also be unnecessary in radiative transfer models, or in future For example, Charlson et al. (2007) highlighted the importance that has been given to the 121 separation between the -cloud‖ and -clear‖ regimes in various fields of study including the 122 radiative forcing by clouds and the quantification of direct effects and indirect radiative 123 forcing by aerosols. The paper questioned the separation between the two regimes, and 124 suggested the desirability of treating the phenomenon as a continuum. Similarly, Koren et al. 125 (2007) described a transition zone (-twilight‖ zone) around the cloud in which the optical 126 properties are close to those of the cloud itself. The authors estimated that an appreciable 127 fraction (between 30 and 60%) of the part of the globe at any given time considered free of 128 clouds could correspond to that area of transition, a fact that could have important climate 129 implications. The question of the climatic importance of clouds that are considered -small‖ in cloud field as an area that includes detectable clouds and twilight zone, and found that the 133 cloud field fraction could be as large as 97% in an area where the detectable cloud fraction is 134 53%. In the cited works, several methodologies were used: spectral radiometry from the 135 surface in the visible and near infrared, satellite measurements, and modeling. Also long-136 wave spectral radiometry is being used for the purpose of studying the properties of thin 137 clouds and the transition region (Hirsch et al., 2014(Hirsch et al., , 2012.  The goal of the current paper is to quantify the importance and frequency of situations where 155 ambiguity between clouds and aerosol occur; in other words, situations where the suspension 156 of particles depend on subjective definition to be classified as either cloud or aerosol. These radiation measurements, and spectral measurements. Two sites are considered: Girona 163 (Spain), and Boulder (Co, USA). used to take images of the sky during daylight hours, at 1 minute time steps. The camera is a 186 conventional digital CCD camera, provided with a fish-eye (i.e. >180º field-of-view) lens and 187 mounted on a sun tracker, in such a way that a black sphere projects its shadow on the lens, 188 blocking the direct sun from entering the camera. In the current research, one year (2014) of 189 data and observations from each of these instruments will be analyzed.  The two locations are middle-latitude, Northern Hemisphere sites. However, they hold some 212 geographic and climatic differences that make pertinent the use of data from both sites in the 213 current research. First, Girona is at low altitude and close to the sea, while Boulder is at high 214 altitude and thousands of km away from the closest coast. Therefore, climate in Boulder is 215 much more continental, in the sense that warmer summers and colder winters are likely; more 216 important here is that the atmosphere above Boulder is in general drier and cleaner, so

221
Raw measurements and observations from the above instruments need to be processed in 222 order to obtain quantitative or qualitative information about the sky condition, clouds and/or 223 aerosol. In all processing and algorithms used (and explained below) decisions must be taken 224 to distinguish between clear sky (either clean or with a certain aerosol load) and clouds, or 225 between clouds and aerosol. These decisions usually take the form of thresholds, which are 226 somewhat subjectively selected after some tuning procedure. Sometimes, the human 227 intervention is obvious, for example when deciding which sky images are considered as 228 cloudless references. In the next paragraphs, we will explain the standard methods applied to 229 raw data, and will describe the sensitivity analyses that we have performed on them to reach 230 our goal.

231
Broadband solar radiation measurements at high temporal resolution (< 5 minutes) can be 232 used to infer the sky conditions. In this regard, after some initial attempts (Calbó et al., 2001; the measurement is the basis of all methods: the underlying assumption is that clouds make 288 solar radiation (either broadband or spectral) more variable in time than aerosols. Here we 289 will use the methodology as presented by Michalsky et al. (2010), which consists of two 290 filters applied consecutively on a moving time window of a given width (10 minutes in the 291 original paper). The first, coarser filter takes the difference between each adjacent 292 measurement, and also calculates the maximum minus the minimum OD in the window. If all 293 differences are less than a given threshold, and if the range of measured OD within the time 294 window is less than another threshold, then the points pass the first filter. The second, more 295 stringent filter scales the allowed variability according to the magnitude of the OD, which is 296 estimated by applying a low-pass filter on the series. Thus, the absolute value of the largest 297 difference between adjacent data must be less than a given fraction of the estimated OD at the 298 midpoint of the sample window, and the range must be less than another fraction of the same 299 estimate. The values of the four thresholds were 0.02 and 0.03 (absolute differences of OD at 300 550 nm) and 10 and 20% respectively in the original paper. In the present study, we will 301 change these four values, and also the time window where differences and ranges are 302 calculated, to assess their effect regarding -transition‖ cases. The final result of the MFRSR 303 cloud screening is every sample tagged as -good‖ or -bad,‖ meaning that can be 304 representative of aerosols or not. In the current paper, we will assume that samples labeled as 305 -bad‖ correspond to the presence of some kind of clouds.

306
As mentioned above, images of the whole sky are becoming more ubiquitous both in 307 atmospheric research and in solar energy management applications. Automatically captured 308 sky images allow a continuous (many such cameras take images every minute or even more  indicating that when such high values of diffuse radiation are in principle set to correspond to 335 clear sky, other tests for clear-sky detection filter out these cases anyway. It should be noted 336 that even with the lower threshold, the diffuse irradiance allowed as -clear sky‖ is well above 337 the Rayleigh limit, i.e., a certain amount of scattering particles larger than molecular is 338 always allowed. A summary of results is presented in Table 1. There are almost 11,000 339 minutes identified as clear when the higher threshold is used but labeled as not clear when the 340 lower threshold is applied. This means that almost 5% of the daylight hours (specifically, of  The difference in mean fsc when data are processed with one or the other threshold is 0.023 349 (0.022) in Girona (Boulder). This difference might not seem very large, but, as we will show 350 below, it is produced by larger differences for some particular conditions. Thus, differences  Table 1), for scattered to 354 broken cloud conditions, the average difference is 0.044 (0.046), which is more than 10% of 355 the average fsc of about 0.4 at both sites. Logically, since RadFlux uses the difference 356 between measured and estimated clear-sky diffuse as the basis for fsc estimation, estimated 357 fsc tends to be lower when Max_Diff is greater. In absolute value, differences tend to be 358 greater for lower fsc (see Figure 1 corresponding to Girona data). Table 2   The data points highlighted in red in Fig. 6 are all those that have passed the cloud screening 413 filter (i.e., labeled as -good‖). Initially (Fig. 6a)  and lower thresholds to produce a more -strict‖ filter. When the former is applied, almost 426 58% of points are considered aerosol (Fig. 5b), but when the latter is used, less than 19% of 427 the points pass the filter (Fig. 5c). With very few exceptions, all points with OD > 1.0 and 428 most points with negative AE (and OD > 0.1) do not pass the cloud screening even with the 429 relaxed filter.

430
The numbers in Table 3 allow an estimation of the frequency of transition cases between 431 cloud and -pure‖ aerosol. We start with about 420,000 instantaneous measurements for  The discussion above concerns results from Girona. When the same analysis is applied to 452 measurements from Boulder, the numbers obviously change, but not the main result of a large 453 percentage of cases in the transition zone. We started with 610,000 instantaneous measurements (note that 20-sec resolution was used in Boulder for the whole year) from 455 which about 158,000 (25%) were not processed by the MFRSR, due to thick clouds occulting 456 the Sun. Then we applied the three cloud screenings (default, relaxed, strict) to the rest of the 457 samples (452,000, see Table 3). About 242,000 additional points were labeled as -bad‖ (i.e.

539
More important than these aggregate numbers is to look at the particular cases with large (or 540 small) cloud fraction changes when the threshold is changed. In this sense, Table 4 shows the 541 differences in thin cloud fraction between the original processing and new processing using  where the change in the threshold produces a large change in the thin cloud fraction, because 558 a large part of that image is made up of what seems very thin clouds. In the example of Fig.   559 9d, the large differences seem to be related to a relatively high atmospheric aerosol load.

560
For the case of the increased threshold, the thin cloud fraction estimate in a little more than 561 80% of images decreases by less than 0.10 (Table 4). This includes a) some situations of 562 cloudless skies, b) situations with scattered to broken cloudiness but with a low amount of 563 thin clouds (in these two cases, the increase of the threshold of course makes it impossible to 564 get lower cloud fractions), and c) situations of overcast skies with thick clouds, that present 565 much higher values of the red to blue ratio (or that are set as cloudy because of very low light 566 intensity, i.e., very thick clouds). Again, this result confirms that the method is quite robust, 567 and also that almost 20% of time a change in the threshold produces a change in the thin 568 cloud fraction of more than 0.1. In Figures 9a-f the result of the cloud identification with the 569 higher threshold is also displayed, and we can see the moderate effect on cases of Fig. 9c and   570 9d, corresponding to clouds (or aerosols) with not well defined limits and that are mainly 571 visible when they are in front of the Sun due to their forward scattering characteristics. The 572 greatest effect of changing the threshold is found in situations such as those presented in Fig.   573 9e and 9f, where the hazy atmosphere (involving cumulus clouds formation) is too 574 problematic to be classified as -clear‖ or -cloudy‖ with thin or even opaque clouds by the 575 method applied to TSI raw images. It should be noted that in the cases where the effect of 576 changing the threshold is small (Fig. 9a and 9b), the optical depth as measured by the 577 0.2, which is the value that Dupont et al. (2008) found as the limit related to the more 580 common differentiation between cloud and not cloud. 581 5. Discussion, summary and conclusion 582 We have presented observations from three ground based, passive systems, that are intended 583 to detect clouds and aerosols in the atmosphere. Indeed the three systems share one 584 characteristic, which is that they are sensitive to the solar radiation flux once it has been 585 modified (affected) by the presence of suspended particles in the air (of course, solar 586 radiation flux is also affected by atmospheric gases). Thus, sky cameras -map‖ radiation 587 coming from the whole sky dome and record this radiation in three color channels (red, In these latter studies, a spatial approach 625 was considered, i.e., they accounted for the extension of this zone in a snapshot of the sky.

626
Our study, however, combines this approach (for sky camera images and partially for 627 broadband hemispheric solar radiation measurements) with a temporal approach, that is 628 accounting how often the atmosphere presents a state that cannot be distinctly categorized as 629 cloud or as aerosol (for broadband hemispheric radiation measurements and also for MFRSR, 630 Sun pointing, measurements). Therefore, our numbers correspond mainly to temporal 631 frequencies, are limited to two particular sites, and are quite conservative, but if we discard 632 the overcast conditions, the relative frequency of the transition cases increases to more than 633 15% of the remaining cases (this number is estimated by dividing the above overall value of 634 10% by the frequency of non-overcast cases, which is about 70% at the two involved sites).

635
Our results support the argument that clouds and aerosol are two extreme manifestations of 636 the same physical phenomenon, which is a suspension of particles in the atmosphere. This