HALO performed 15 research flights on 15 different days in support of EUREC4A, as listed (with an evocative moniker) in Table . Flight IDs in the format HALO-MMDD (where M is month and D is day), rather than an enumeration of the research flights, are used to distinguish the different flights. This helps avoid confusion arising from non-coincident flights among the various research aircraft contributing to EUREC4A. Of these, 13 (HALO-0122 to HALO-0215) are designated as local flights, as they had both the takeoff and landing at Barbados' Grantley Adams International Airport. With the exception of HALO-0130 – a short flight that took advantage of overlap in crew duty to make some additional measurements of opportunity – each local flight lasted about 9 h, with roughly 7 h of circling on what call the “EUREC4A-Circle”. This circle was largely defined by the HALO flight pattern, which was fixed before the beginning of the campaign to support the deployment of dropsondes around a geographically fixed circle positioned windward of the BCO Barbados Cloud Observatory;, far enough upwind to not interfere with commercial air traffic but not so far as to be out of range of a C-band polarized research radar (POLDIRAD). The EUREC4A-Circle is easily identified as the darkest circle area in the heat map of flight tracks in Fig. , with a center at 13.3∘ N, 57.717∘ W and an approximately 220 km diameter.

An important and unusual aspect of the HALO (and EUREC4A) flight strategy was that it did not target specific meteorological conditions. Flight days were scheduled in coordination with the ATR aircraft so as to maximize the utilization of the aircraft subject to crew duty restrictions. Variations in takeoff (and landing) times were implemented to better sample the diurnal cycle and were staggered to accommodate crew duty considerations, rather than to target specific meteorological conditions. On most flights some time was also dedicated to flight elements other than the EUREC4A-Circle, for instance to allow an underpass of a meteorological satellite (e.g., the Terra satellite during Bernhard's Bicycle) or to sample the upwind conditions that were being monitored by other platforms. Only Mario's Snail (HALO-0130); the southeast excursion on Manfred's Escape (HALO-0126), which coordinated sampling of a cirrus deck with the R/V Meteor; and the choice of flight levels on Under Cover were influenced by meteorological conditions. The moniker associated with each flight (Table ) was chosen to strengthen the mental image associated with that flight and in most cases remind the reader of the principle investigator (PI) of each flight.

To aid in the analysis of flight data, all HALO flights are segmented via timestamps into a system of hierarchical identifiers. Non-exclusive segments are defined by two (YYYY-MM-DD hh:mm:ss) timestamps, the first one defining the start of the segment and the second denoting the first time step after the end of the segment. Timestamps have a temporal resolution of 1 s, and times are given in UTC. Every segment belongs to a “kind” – a categorical type for segments defined in Table . It helps to think of segments as an interval of flight time and the corresponding kinds as describing how the aircraft was being operated during this time interval (Fig. ).

Flight segmentation data are provided as YAML (YAML Ain't Markup Language) files that can be accessed at 10.5281/zenodo.4900003. This section provides a description of the YAML files and the reasoning behind their structure and method.

By adopting non-exclusive segments, a timestamp can belong to multiple segments that differ in kind. For example, timestamps belonging to the kind “lidar leg” will also belong to the kind “straight leg” if they match the definition of the latter. Segment start and end times were first roughly categorized based on timestamps from the flight reports and aircraft navigation features such as roll angle, altitude, pitch, etc. However, the final attribution of timestamps to segments was performed manually by the listed contact in the YAML files. At least one other person later tested the segmentation for errors or avoidable deviations from the segment definitions (Table ). This was done to maintain the objective segment classification whenever possible so that the user of these data can expect the segments to match the definition as closely as possible. As the segmentation was performed manually, segments are defined by the time intervals that are assigned to them, rather than by their kind.

The flight segments in the YAML files also contain a field called “dropsondes”, which provides a list of the dropsondes, whose times of launch are associated with the respective segment. The dropsondes are provided with classifications of good, bad, and ugly, based on their QC classification types from the EUREC4A dropsonde dataset JOANNE Joint dropsonde Observations of the Atmosphere in tropical North atlaNtic meso-scale Environments;. The list is in the form of unique dropsonde IDs that correspond to the variable “sonde_id” provided in JOANNE and are the “cf_role” variable therein. This field makes it convenient for selection of the dropsondes based on flight segments. In a few instances the launch time of a dropsonde will fall outside of the segment with which it is associated – for instance if the last sonde of a circle was inadvertently launched too late, after HALO had completed a circle.

Flight segments that deviated from the allowed kinds are flagged by an irregularity field. For instance, the inclusion of sondes launched before or after their associated flight segment constitutes an irregularity and is marked. The irregularity field takes the form of an explanatory string describing the irregularity. As the segmentation process revealed some oft-repeated irregularities, standardized irregularity tags (keywords) were defined (Table ) and are prepended to the explanatory string of the irregularity field when applicable.

In total 220 segments were defined over the 15 flights. These included 72 circles (69 regular, 1 with a smaller diameter, 1 outside of the EUREC4A-Circle, and 1 without dropsonde launches) within 26 periods of circling. Fifty-one straight legs were flown. The segmentation data are published by .

To give further insight into the large-scale conditions of each flight, satellite movies overlaid with the time-evolving flight tracks are created. Snapshots of these movies are shown in Fig.  for each flight of HALO. The snapshots were chosen to capture the cloud scene roughly 3 h after takeoff. Like the snapshots, the actual movies are based on the 1 min meso-scans of the Advanced Baseline Imager (ABI) on board the GOES-16 satellite , when these are available. During the daytime reflectance (channel 2, 0.64 µm) is used, and during the nighttime brightness temperature (channel 13, 10.35 µm) is used. On a few days the ABI did not provide meso-scans over the EUREC4A domain. In these cases, 10 min full-disk scans were substituted. To foster the generation of movies with different overlays by users, the source code is available and relies purely on publicly available data sources.

During all research flights photographs were taken to visually document the conditions being sampled (Fig. ). Most photos were taken by the principal investigator, through either the left or the right window in the middle of the cabin forward of the wing; a few were taken from the cockpit. A subset (50–100 per flight) of these photos has been selected and further curated as described below. These photos, with their extended metadata, are included as part of EUREC4A's HALO dataset.

The data curation involved manually correcting camera timestamps by calibrating the camera's internal clock with photographic evidence of flight-level time data from GPS watches or instrument panels synchronized with the aircraft sensor system time (BAHAMAS – BAsic HALO Measurement And Sensor system, Sect. ). GPS location and altitude tags are added to each photo using BAHAMAS location data at the capture time. For photographs taken on the apron, where aircraft position data are not available, the position of the usual parking position (13.08∘ N, 59.4828∘ W) was used. With a cruising air speed of 200 m s-1, the estimated 1 min accuracy of the capture time implies a GPS location accuracy of about 12 km.

Additional metadata were added using standard IPTC (International Press Telecommunications Council) metadata conventions. The IPTC tag “description” is used to describe the scene photographed. The IPTC tag “keywords” contains information about the orientation (viewing direction), the platform HALO, pictured cloud types, or other notable objects. In cases where the orientation could not be determined, a default is adopted, usually to the left or right of the PI seat. Because most of the photos were taken with a shared camera, some may have been taken by different members of the flight crew; when this was not documented, the PI of each flight is set as the creator. The supplementary photo documentation is written into each photo's IPTC tags as part of its extended metadata. The photographs can be viewed and downloaded from the database .

The BAsic HALO Measurement And Sensor system (BAHAMAS, Table ) provides aircraft location and attitude data for all HALO flights, in addition to atmospheric measurements. A subset of the BAHAMAS data, consisting of the aircraft altitude, heading, latitude, longitude, roll angle, pitch angle, and true air speed with a time resolution of 10 Hz, has been created . Figure  uses the data subset to present the tracks of all flights in the vicinity of Barbados as well as the ferry flights from and to Germany. The roll and pitch angles of all flights are shown in Fig. . The distribution of the roll angles (blue) shows two peaks. The one centered at 0∘ indicates straight legs, while the other centered at 2.2∘ arises from circling in a clockwise (positive roll angle) manner. The distribution of the pitch angle (orange) shows a peak near 3∘. This pitch changes systematically as fuel is burned through the flight. Although the constant roll angle on the measurements during circling is sometimes raised as a concern, this analysis shows that – for the large circles flown during EUREC4A – the non-zero pitch results in a larger deviation from the true nadir of the downward-staring instruments than does the constant roll.

EUREC4A's HALO was designed to observe different properties of clouds using the richness of their interaction with electromagnetic radiation. Different instruments (Table ), by virtue of their differing measurement principles and footprints, see clouds in different ways. Figure  provides a snapshot for a 5 min flight segment from flight HALO-0205 on a circle segment (HALO-0205_c2, Table ), which represents typical cloud conditions of EUREC4A. WALES and the HAMP radar provide vertical cross sections; specMACS and VELOX provide a two-dimensional horizontal view of the clouds along the flight path, and other instruments provide a scalar time series of measurements along the flight path.

To provide an overview of the cloud fields sampled by HALO, a trinary cloud mask is created for each cloud-sensitive instrument, as described in Appendix . Access to the cloud mask data is outlined in Table . The value of the cloud mask denotes measurements that each instrument identifies as cloud-free (0), probably cloudy (1), or most likely cloudy (2). The introduction of the probably cloudy flag reflects the ambiguity in cloud detection faced by many instruments. Especially for the passive instruments (HAMP radiometer, specMACS, KT19, VELOX), a range of thresholds were applied to separate cloudy and cloud-free observations. Cases where the lower and upper threshold give a different decision are marked as probably cloudy. A comparison of the cloud masks (Fig. ) shows how the cloud amount is sensitive to the manner of detecting clouds. The radar is sensitive to large drops, which form through the collision and coalescence of cloud droplets, a process that becomes active as clouds deepen and increase their condensate burden. The lidar, on the other hand, is also sensitive to optically thin clouds with a very small condensate burden. This explains the differences in the measured cloud cover by these two instruments for the 5 min segment shown in Fig. . The sensitivity of the passive instruments is influenced by the contrast of the cloud and surface reflection or emission. A time offset is also apparent in different cloud flags, which arises from slight differences in the instrument orientations (more forward pointing instruments detect clouds earlier than more backward pointing instruments), rather than lack of synchronicity.

The campaign average cloud-cover estimates as detected by the instruments are stated in Table . Most instruments define a minimum cloud cover based on the cloud flag most likely cloudy and a maximum cloud cover that additionally includes the uncertain cloud flag probably cloudy. WALES stands out as there is no probably cloudy flag in the cloud mask algorithm (Sect. ) and the minimum and maximum cloud cover are equal. The HAMP radar seems to have very few uncertain cases.

To provide context for the variations in cloud cover, Fig.  shows a time series of circle-mean (minimum) cloud-cover estimates for all research flights and from all instruments. HALO typically flew six circles per research flight (per day). In addition to the research flight mean, the whiskers span the range from the smallest to the largest circle-mean (minimum) cloud cover for each flight. In general, the individual instruments show a tendency towards higher cloud cover at the beginning as well as towards the end of the campaign. For most cases the cloud-cover estimates from passive instruments and the radar agree well. WALES systematically detects more clouds. It is more aligned with the circle-mean (maximum) cloud-cover estimates of the other instruments, as it does not include an uncertain cloud flag and is very sensitive to optically thin clouds. The flight HALO-0215 is an exception to the systematic difference between WALES and the other sensors, which is due to a deep stratocumulus layer with a strong reflection at the cloud top that blinded the lidar while the radar was still able to provide reasonable estimates.

To further investigate the differences among the sensors and their cloud-masking algorithms, we display the cumulative fraction of circle-mean cloud-cover estimates in Fig. . In particular, the bars show the range defined by the circle-mean minimum and circle-mean maximum cloud-cover estimates for the cloud-cover ranges stated on the x axis. The differences between minimum and maximum cloud cover originate from the uncertain cases with the cloud flag probably cloudy. The first thing to note is a disagreement between the instruments for cloud cover ranges up to about 0.5 due to their different detection principles. Geometrically and optically thin clouds can have a significant impact on circle-mean estimates in low-cloud-cover situations and lead to uncertain pixels depending on the detection principle . As WALES is able to detect optically thin clouds with few condensates, the cloud-cover estimates are generally higher, and the change in cumulative fraction is strongest between 0.2 and 0.6. The radar stands in contrast to WALES, with most circle measurements exhibiting a cloud cover below 0.2 as it cannot detect the small and optically thin clouds at the operating wavelength. The VELOX cloud mask includes a high fraction of uncertain pixels leading to a large difference (large bars) between the minimum and maximum cloud cover visible in Fig.  at cloud covers of up to 0.4. In the case of VELOX as well as for all other passive instruments, the cloud-cover estimates shift to higher numbers when the thresholds are reduced (from minimum to maximum cloud cover).

In general we find that only a few circles have a cloud cover higher than 0.6. At such high cloud cover the instruments agree remarkably well and, also, minimum and maximum cloud cover are almost equal, meaning that there are few or none probably cloudy measurements. Or, to put it another way, about 90 % of all circles have a cloud cover below 0.4 for most instruments except VELOX with 90 % of cloud-cover estimates below 0.6. Furthermore, about 50 % of the time cloud-cover estimates are below 0.2. The analysis of circle-mean cloud cover suggests a high abundance of low-cloud-cover situations. WALES as well as the passive instruments is capable of detecting the thinner cloud edges and small and optically thin clouds. The HAMP radar is not sensitive to these cloud parts which typically consist of small cloud droplets. The comparison illustrates the potential of the multi-instrument cloud-cover product to study cloud macro- and microphysical properties.