A framework for post‐processing bird tracks from automated tracking radar systems

Radar is an effective tool for continuous monitoring and quantification of aerial bird movement and used to study migration and local flight behaviour. However, systems with automated tracking algorithms do not provide the level of processing sufficient to guarantee reliable data. Therefore, post‐processing such radar data is required but often non‐trivial, especially in challenging environments such as open sea. We present a post‐processing framework that implements knowledge of the radar system and bird biology to filter the data and retrieve reliable, high‐quality tracking data. The framework is split into three modules, each with a specific aim: (I) sub‐setting based on prior knowledge of the radar system and bird flight, (II) improving bird track quality and (III) detecting and removing spatio‐temporal sections of data that have a clear bias for false observations. The effectiveness of the framework is demonstrated with a case study comparing track densities inside and outside an offshore wind farm, and by applying the workflow to a dataset of visually validated radar tracks. Application of Module I resulted in a dataset of 520.894 bird tracks (19.5% of total data) within a 10.4 km2 area. Additionally, 18.734 tracks were corrected for geometric errors in Module II, and Module III identified 236 of 719 observation hours and an area of 1.55 km2 as unreliable for spatio‐temporal analysis. No difference in track densities was found between the area inside and outside the wind farm when using the post‐processed data, whereas using the unprocessed bird tracks, lower track densities were observed outside the wind farm. Of the visually validated radar tracks, the framework removed 85% of false positive bird tracks, while retaining 80% of true positive bird tracks. The framework provides a logical workflow to increase the reliability and quality of a bird radar dataset while being adaptable to the radar system and its surroundings. This is a first step towards standardising the post‐processing methodology for automated bird radar systems, which can facilitate comparative analyses of bird movement in space and time and improve the quality of ecological impact assessments.


| INTRODUC TI ON
Radar is a remote sensing technique which has been used for decades to quantify different aspects of avian flight and monitor biomass flows and has great potential for ecological research and diverse applications (Bauer et al., 2019;Shamoun-Baranes et al., 2019).It is the only non-invasive technique for quantifying avian flight that can measure aerial movement across a broad range of avian masses and be deployed for continuous measurements in diverse habitats.Radar has been adopted as one of the main tools in aeroecology (Gurbuz et al., 2015;Shamoun-Baranes et al., 2019), and different radar systems have been used to study migration (Bradarić et al., 2020;Bruderer, 1997;Dokter et al., 2018), flight behaviour at sea (Assali et al., 2017;van Erp et al., 2021van Erp et al., , 2023) ) and the flight response to anthropogenic activities and structures (Aschwanden et al., 2018;Desholm & Kahlert, 2005;Van Doren et al., 2017).Additionally, radar is also used for environmental impact assessments, long-term monitoring and forecasting bird movements for applications such as aviation safety and wind energy (Shamoun-Baranes et al., 2019;van Gasteren et al., 2019).
Currently, several dedicated bird monitoring radar systems are commercially available and designed to be used 'off the shelf', such as the Robin Radar systems (Robin Radar Systems, 2023) and MERLIN Avian Radar System (DeTect, 2023) among others (Accipiter, 2023;AscendXYZ, 2023;Miltronix, 2023).These systems (from here 'bird radars') have been used to study local (van Erp et al., 2021(van Erp et al., , 2023) ) and migratory (Bradarić et al., 2020;Fijn et al., 2015) flight patterns in remote areas in relation to wind energy developments.They scan the horizontal and vertical plane and contain processing software on board for automated identification and tracking of bird targets from raw radar scans.Although this automated process is time-and storage-efficient in generating biological data, these data may still contain a considerable amount of false positive (reflections erroneously identified as birds) and false negative observations.This is problematic for answering ecological questions (Guillera-Arroita et al., 2017;Rempel et al., 2019), and therefore, post-processing is required before data can be interpreted adequately.Several aspects of bird radar systems, including radar image processing (Stepanian et al., 2014), target detection (May et al., 2017;Urmy & Warren, 2017) and automated tracking (Urmy & Warren, 2020), have been addressed in literature, however, post-processing has not received the same attention.To date, this is done ad-hoc, as radar system performance can vary between locations, weather conditions, radar software and hardware configurations.This difficulty of radar data processing can hinder data standardisation and harmonisation between radar systems (Liechti et al., 2019).
In this paper we provide a modular framework to create workflows for processing horizontal bird tracks gathered from bird radars.We focus on automated systems; however, several processing steps can also be applied to radar studies where birds are tracked manually or using external software (e.g.Capotosti et al., 2019).The workflows can be tailored to different circumstances and systems, resulting in more reliable and standardised datasets for studying bird flight quantitatively, and facilitating reproducibility and interoperability.The framework consists of three modules, each with a separate aim.To demonstrate the implementation and consequences of each module regarding data reduction and quality, we create a workflow for a subset of tracks observed by a radar positioned in an offshore wind farm and perform a comparative study of track densities inside and outside the wind farm using both the unprocessed and post-processed bird tracks.Additionally, we apply the workflow to a dataset of radar tracks visually validated with field observations to verify the accuracy of the framework.Offshore wind farms provide a strong testing ground as they are challenging environments for deploying radar due to high reflectivity that can be caused by waves and wind turbines.

| MATERIAL S AND ME THODS
The modules are ordered so that the first steps require only basic knowledge of the system, are computationally simple, and result in a major data reduction to facilitate more complex subsequent processing steps (Figure 1).To ensure applicability of the framework to most bird radar systems, only locations and corresponding timestamps of bird trajectories are required.Required weather data can be acquired from the ECMWF ERA5 reanalysis (Hersbach et al., 2020) which is freely available and has global coverage.
Module I removes data from areas with low observation reliability and identifies false positive observations.Module II improves the quality of the remaining tracks by removing false points within the track geometries.Module III removes remaining spatial and temporal sections with low reliability due to increased occurrence of false positive or false negative observations.Bird radar data processing is an iterative process because the consequences of processing steps often cannot be assessed beforehand.Hence, parameter choices need to be evaluated and visual verification of processing results is indispensable.Besides being the primary tool for finding the right bird movement in space and time and improve the quality of ecological impact assessments.

K E Y W O R D S
aeroecology, automated tracking, bird radar, movement analysis, remote sensing parameter values, it also functions as confirmation on the adequate performance of the modules.
Each processing step will be applied to a case study to create a workflow and show its effect on data quality and quantity.Data processing was carried out in R (R Core Team, 2020), using the birdR package (De Groeve & van Erp, 2023) that was specifically developed for this task.We used a subset of the data from a study of thermal soaring behaviour in seabirds at sea (van Erp et al., 2023), namely the data from the bird radar (Robin Radar 3D-Fix) positioned at the Luchterduinen Offshore Wind Farm (52.427827°N, 4.185345° E) collected from 1 to 30 June 2020 (2.668.345bird tracks).The 3D-Fix has two antennae (one horizontal, maximum range = 10 km, and one vertical, maximum range = 6 km) to capture horizontal and vertical information of birds crossing the respective beams.However, only data acquired exclusively by the horizontal antenna were used in the study (see van Erp et al., 2023) and, consequently, here as well.

| Module I: Sub-setting based on prior knowledge
The first module removes tracks from areas that have a low reliability for observing birds (step 1) and tracks that are likely to be false positive observations (step 2).The data reduction also facilitates more complex subsequent analyses.These steps only require basic knowledge of the bird radar and the least computational effort per track.

| Spatial filtering based on a defined inclusion area
Several factors affect what area within the radar window is reliable for observing bird flight.First, there is a minimum and maximum distance from the radar at which birds can be reliably detected.
Additionally, nearby large features might disrupt bird detection in the area around them or block the radar beam.Together these factors should be used to determine the radar area of inclusion (AoI, Figure 2) in which bird flight can reliably be measured.The first step is to estimate the AoI based on the properties of the bird radar and apply it as a spatial mask to the data.
At small distances the radar beam is powerful enough to reflect on many unwanted features in the landscape (Rinehart, 1991) and create clutter, partly because of the side lobes of the radar beam (Figure 2b).Due to increased clutter, the number of false positive observations increases and this structural increase at small distances leads to a positive bias in bird observations (positive observation bias, or POB).Therefore, it is necessary to set a minimum distance (MiD) from the radar for data inclusion.The MiD depends on the power of the radar antenna and, to a lesser extent, the radar placement height, where a more powerful beam and higher placed radar results in a larger MiD.
With increasing distance from the radar, a bird's theoretical detection probability (TDP) decreases (Bruderer, 1997;Schmid et al., 2019).Given uniform bird abundance and size distribution over the radar window, the observed number of birds will be lower at larger distances, leading to negative observation bias (NOB).To reduce the impact of reduced TDP and limit NOB, a maximum distance (MaD) for data inclusion should be set at a distance where the TDP for the smallest bird of interest is still high.We recommend setting the MaD to a distance at which the TDP is at least 80% (Figure 2c).
Note that using a MaD limits the maximum bias in the collected data, but does not correct for the decline in TDP with distance below the MaD.
In the area between the MiD and MaD, there can be locations where birds cannot be detected due to static terrain features (typically structures or elevated terrain).Such features may impact bird F I G U R E 1 Overview of the processing framework.The framework consists of three modules (I, II, and III), each working with a defined data input (blue boxes).'Radar Bird Tracks' in Module I refers to all tracks classified as birds by the radar system.The subsequent data inputs ('True Bird Tracks' and 'Clean Bird Tracks') are subsets.Each processing aim (dark green boxes) is realised by one or several processing steps (light green boxes).After a module is applied, the output is suitable for specific purposes (orange boxes).Solid arrows show the logical progress within each module.Dashed arrows indicate the sequence of module application based on data verification and newly gained knowledge.If post-processing is successful users can move forward to the next module (A), but if the results are unsatisfactory users can reiterate on processing steps in the current module (B) or previous modules (C).detection in opposite ways: by creating false positive observations or false negative observations.False positive observations occur through the interaction between the radar beam and a feature, leading to clutter.An example is the clutter caused by turbine rotors (dashed black box in Figure 2c).False negative observations occur when features block sections of the radar window or create increased background scatter.The structure on which the radar is placed can itself be a feature that obstructs the radar beam at certain angles (dashed purple box in Figure 2c).If the location of these features is known in advance, the areas surrounding them should be excluded from the AoI.These areas should consist of the geographic position of the features plus a buffer, or if the feature blocks the radar beam entirely, the angles at which the beam is blocked should be removed.The buffer should be set based on the estimated arc length of the radar beam at the location of the feature or the pulse length (whichever is largest), as these parameters define the distance from a feature at which it reflects the radar beam.The arc length at the location of the feature can be estimated if the beam width is known using the circumference formula for circles: where s is the arc length, θ is the beam width (°) and r (m) is the distance between the radar and a blocking feature.One would thus define smaller buffers around features nearby the radar than further away.However, applying identical buffers for all features would simplify things.This can be done by taking the MaD as a value for r in Equation 1, which comes at the cost of slightly larger buffers for some features than is strictly necessary.
The definition of the AoI may also account for spatial constraints resulting from the research aim.For example, van Erp et al. (2021) studied bird flight over the open sea, which required excluding the area enclosing a wind farm.Although such considerations seem obvious, this spatial sub-setting can tremendously decrease the data volume and should therefore be included as a first analysis step.
Once the AoI is set, it is applied as spatial mask and radar bird tracks occurring completely outside the AoI are removed.

| Case study
Based on the TDP for large bird targets (−13 dBm -2 , Figure 2b) the MiD was set at 1000 m and based on the TDP for small bird targets (−25 dBm -2 , Figure 2c) the MaD was set at 2500 m.The radar system was installed on the service platform of one of the wind farm (1) Area of inclusion (AoI, grey area) for the 3D-Fix (red dot) at Luchterduinen Offshore Wind Farm that was used in the case study.The AoI is determined by a minimum and maximum distance from the radar (1000-2500 m), two angles at which the radar beam is blocked or data are unavailable (between 287°-30° and 115°-135° relative to Geographic North respectively), and the area within 100 m radius of each wind turbine (black dots).(b) Theoretical detection probability (TDP) plot for large bird targets (−13 dBm -2 ); colour hue depicts TDP from 100% (purple) to 80% (blue).Close to the radar (<1000 m, blue line) the radar beam and side lobes hit the terrain with enough power that there is a high chance for false positive bird observations.(c) TDP plot for small bird targets (−25 dBm -2 ); colour hue depicts TDP from 100% (purple) to 80% (blue).At large distances (>2500 m, red line), the TDP for birds drops to the extent that small bird tracking becomes less reliable.(d) Terrain features within the radar window (nearby wind turbines, dashed black box) can cause increased false positive bird observations or increase the background scatter so that birds cannot be detected (leading to false negative observations).Large or nearby features (such as the turbine at which the radar system is installed, dashed purple box) can block the radar beam at certain angles (here between 287° and 30°).
turbines (dashed purple box in Figure 2d), which blocked the radar beam between 287° and 30° (relative to Geographic North).The dataset only included bird tracks observed by the horizontal antenna of the radar system; tracks containing both horizontal and vertical information were excluded.Consequently, the area where the two antennae overlapped (between 115° and 135°) would suffer from NOB and was therefore excluded as well.Lastly, the other turbines in the wind farm have a negative impact on bird detection in their proximity (dashed black box in Figure 2d).Using the pulse width of the 3D-Fix (30 m) and formula [1], the spatial resolution of the radar at MaD (2500 m) was calculated to be 78.5 m.Therefore, a conservative 100 m radius exclusion zone was applied around their locations.
Together, these limits created the AoI for the case study (Figure 2a).
Tracks occurring completely outside the AoI were identified through a spatial intersection and a sub-sample was visualised to verify the identification was successful (Figure S1).The tracks outside the AoI (2.092.252) were removed; 576.093 radar tracks (21.6% of total) were retained (Table 1).

| Removing non-bird movement
After the spatial filter, false positive bird tracks are identified by two methods.First, their average airspeed is evaluated.Empirical measurements show that the average cruising airspeed of a wide range of bird species ranges between 8 and 23 ms −1 (Alerstam et al., 2007).Alternatively, aerodynamic theory and scaling laws can be used to estimate a range of realistic airspeeds during flight (Pennycuick, 1975).Including a margin for deviation from average airspeeds and different flight modes and conditions, see for example Spear and Ainley (1997), we recommend setting the range of acceptable airspeeds to 5-30 ms −1 .If observing a certain bird species or species group, the minimum and maximum airspeed could be further limited to reflect a species' observed or theoretical airspeeds.
A second filter can be applied to identify false positive bird tracks from features unaccounted for with the AoI.Examples of such features are objects which have variable reflectivity and are misidentified as birds under certain conditions (high wind speeds causing trees to move) or features that only exist temporarily within the AoI (e.g. an anchored ship).As these features are fixed, the false positive tracks obtained from these clutter sources can be distinguished from bird targets by their net spatial displacement (the Euclidian distance between the start-and endpoint), which will be close to zero.
In contrast, the net displacement for birds tracked within the radar window tends to increase the longer they are tracked.By calculating the displacement divided by the duration of the track (displacement over time, DoT), clutter tracks can be identified as the data with the lowest DoT.As clutter tracks are also often visually distinctive from bird tracks (Figure 3a), the data can be split in percentiles of DoT   (Spear & Ainley, 1997) were removed.Success was verified with an overview of the average airspeed of removed and remaining tracks (Figure S2).Next, DoT was calculated for all remaining tracks, and the 0.2nd to 2nd percentiles of data were extracted in increments of .2% and visualised (Figure 3).A distinct pattern of clutter tracks caused by the turbines was visible in the 0.2nd to 0.8th percentile (Figure 3a).The clutter pattern could not be clearly distinguished after the first percentile (Figure 3b).Based on these visualisations, the threshold for track removal was set at the first percentile (DoT = 1.15 ms −1 ).   points in the identified tracks, or 1.7%).These tracks were corrected by removing the two points with the time interval below the 0.12 s threshold.After correcting the tracks, all averaged track properties for those clean bird tracks were re-calculated (track length, track duration, average ground speed, average airspeed).

| Module III: Removing data sections with observation bias
Subsets of the remaining data could still suffer from observation bias due to external conditions or poor radar performance.Module III is split into two processing steps: one dedicated to identifying temporal NOB in which the whole AoI is affected temporarily and one dedicated to identifying spatial observation bias in which an area within the AoI is affected constantly.Identifying these data sections requires intensive processing and is therefore done after reducing the data volume in Module I.

| Detecting temporal negative observation bias
Temporal NOB can occur due to weather conditions affecting measured reflection, such as rainfall or high waves at sea, as well as system malfunctioning or maintenance.Periods when the system is not working appropriately should be identified to distinguish between no birds being observed (true zeros) and no observations (no data).Depending on the operating system, system malfunctions or shutdown, can be detected in the event logs or the absence of event logs of the radar system.Weather conditions have an indirect effect on bird detection, and an analytic approach is needed to determine the time periods over which the data are suffering from unacceptable levels of NOB.Note that we only consider NOB here, as radar tracking software generally responds to a temporary increase in reflections by increasing the threshold for object detection to avoid the tracking of clutter.
However, the following step can also be applied to detect periods with POB if required.
During bad weather, for example, in case of a rough sea or precipitation, large sections of the radar window can show increased reflectivity.The radar software takes measures to reduce false positive observations, either by temporarily increasing the threshold for target detection or masking areas of the radar window completely.
This leads to NOB for the periods when these measures are active, and these periods can be identified with the following steps.First the number of true bird tracks measured by the radar is calculated over time.We advise an hourly time scale, as this resolution is sufficiently fine to capture changes in weather phenomena while being sufficiently coarse to contain suitable sample sizes.This time series is then annotated with the variable that is expected to introduce NOB in the data and visualised (Figure 4a).Either the weather variable causing NOB or the diagnostic information from the radar system, like the detection threshold or masking intensity, can be used.The effect of the independent variable is modelled with a general additive model (GAM, Wood, 2020), as the relation is often non-linear (Figure 4b).If negative bias is confirmed visually, a threshold to distinguish between reliable and unreliable periods can be set based on the trend line of the independent variable.For example, the first derivative of the GAM curve can be used to determine the value of the independent variable where the negative effect starts (first derivative is larger than 0) or the effect is the strongest (first derivative is at its maximum).Tracks occurring completely within periods where the independent variable exceeds the threshold are removed from the dataset.The 3D-Fix uses a spatial mask to prevent observations in areas with increased reflections and stores the intensity of this mask in the variable 'landmask'.This variable is recorded any time that the radar is operational.Hence, time periods without entries for the landmask variable denote moments that the radar was not operational, which in our case study was 1 h (12:00-13:00, 3 June 2020).
The data were coerced to a temporal dataset; for each hour (except the hour the radar was offline) the number of clean bird tracks was counted, the average hourly landmask intensity was calculated and labelled as masking intensity.The time series of bird counts and masking intensity showed an inverse relationship (see Figure 4a).Cells showing observation bias can be identified by determining an upper and/or lower threshold relative to the trend line (Figure 5c).
Tracks completely occurring within unreliable cells are then removed from the dataset.

| Case study
The lowest spatial resolution of the radar system within the AoI can be estimated with the arc length at MaD (78.5 m, see Module I) and the pulse length (30 m) which is consistent over distance from the radar.Given the size of the turbines and accounting for the spatial resolution of the radar, a cell size of 100 × 100 m was chosen.A grid was created spanning the complete AoI and the number of tracks within each cell was counted (Figure 5a).Cells were annotated with the distance between their centre and the radar location.The relationship between the number of birds per cell and distance from the radar was estimated by fitting a GAM to the data and visualised on top of a scatter plot of bird count against distance from radar to see if observation bias occurred (Figure 5b).
Several cells showed low bird counts (0-300 tracks per cell) relative to the trend, mostly within the wind farm and near the borders of the AoI (Figure 5a).Relatively high bird counts were observed near the north-eastern wind turbine (52.431°N, 4.205° E).Upon inspection, this did not seem to be due to increased false positive observations, as the trajectories in this area did not show a clear sign of increased clutter tracks (Figure S3).Therefore, only a threshold for NOB detection was set.The predicted value minus 10 times the standard error was chosen as the threshold for identifying cells with NOB (Figure 5b), which marked 148 of the 973 cells (Figure 5c).These cells were excluded from the final dataset and 9.814 clean bird tracks which occurred completely within the area of those cells were removed from the dataset (verified in Figure S4).

| Case study analysis: Track densities inside and outside an offshore wind farm
To provide an example of the qualitative and quantitative effects of the post-processing framework in an ecological context, we analysed the estimated track densities inside and outside the wind farm.Track densities, as approximation of bird densities, are relevant for collision risk estimates (Cook et al., 2018) and the relation between track densities within and outside a wind farm can indicate avoidance or attraction behaviour (May 2015).
To estimate track densities in these two areas we apply a similar method as described by van Erp et al. (2023).Tracks of the unprocessed dataset (n = 2.668.345) and the post-processed dataset (n = 520.894)were annotated by occurrence inside or outside the wind farm area.For the unprocessed dataset, the wind farm area was considered as the geometric box drawn around the outer turbines of the wind farm plus a 100 m buffer (17.3 km 2 ).The area within 10 km from the radar minus the wind farm area was set as the area outside the wind farm (295.7 km 2 ).For the postprocessed data, the overlap of the area of inclusion (Figure 2a) and the complete wind farm area was considered as the relevant wind farm area (4.3 km 2 ) and the difference between the area of inclusion and the complete wind farm area was considered as the area outside the wind farm (6.1 km 2 ).If a track occurred both inside and outside the wind farm area it was considered for both categories.Average track density per day (# km 2 /day) was calculated for both areas by dividing the number of tracks per area by the number of observation days and area size.Additionally, track density for the unprocessed data was visualised on a 100 m × 100 m raster similar to the spatial overview in Figure 5 for comparison with the processed data.

| Post-processing of visually validated radar tracks
The classification and tracking of the bird radar at Luchterduinen was validated by Waardenburg Ecology through visual observations of the radar tracks between 2019 and 2021 (Leemans et al., 2022).
Over this time 675 tracks measured by the horizontal antenna were validated and classified as true positive (tracks identified as birds both by the bird radar and by observers, 583) or false positive (tracks identified as bird by the bird radar, but as another object/feature by the observers, 92).To verify the accuracy of the framework, these tracks were post-processed with the workflow established in the case study.1).The workflow applied 13 different parameters to process the data (Table 2).Ten of these were specified a priori and three were specified interactively using visual inspection.

| Track densities inside and outside an offshore wind farm
When using the unprocessed bird radar tracks, average track densities inside and outside the wind farm were 1.397 and tracks was extremely high near the radar (distance <500 m) and dropped close to 0 further from the radar (Figure 6).Using the post-processed bird tracks, average track densities within and outside the wind farm were 2.116 and 2.350 birds per km 2 /day respectively.The number of recorded tracks was similar inside and outside the wind farm, with slightly higher numbers near the north-eastern-most turbine and close to the radar (Figure 5c).
The increase in track densities for the post-processed data is the result of a large reduction in the considered area relative to the unprocessed data (97.9%outside, 75.1% inside) which showed very low track numbers (Figure 6a), resulting in a proportionally large decrease in considered area relative to the decrease in track numbers.Additionally, the tracks of the post-processed data are on average longer in length (607 m) than the unprocessed data (412 m), and these longer tracks are more likely to cross both areas (inside and outside) in their lifetime and contribute to both density estimations.

| Visually validated radar tracks
Of the 675 validated radar tracks, the workflow identified 192 (28%) as unreliable.Approximately 20% of the true positive bird tracks (115 of 583 tracks) were removed (Table 1).The largest

| DISCUSS ION
We provide a post-processing framework to standardise methodology and improve the reliability and quality of biological data In our case study, which used data from a challenging environment for radar monitoring, a considerable number of observation hours and grid cells with low reliability (33% and 15% respectively) were removed.After post-processing the data, the outcome of a comparative analysis of track densities inside and outside the wind farm changed from a clear difference (1.397 birds per km 2 /day inside and 241 birds per km 2 /day outside) to higher densities that were similar between the two areas (2.116 birds per km 2 /day inside and 2.350 birds per km 2 /day outside).The difference in measurements can have major implications for environmental impact assessments, as they affect estimation of collision risk (Cook et al., 2018) and can indicate the measure of avoidance to the wind farm (May et al., 2017).
The framework is designed to facilitate verification and iteration within and between processing steps with the help of data visualisations.Data visualisations are vital to verify whether the intended result is reached, and the framework's modularity facilitates iteration if the results are unsatisfactory.Reiterating previous post-processing steps can improve the reliability of data and subsequent research results.However, it is also important to realise that often 'good enough' practices should be adopted, as perfecting the post-processing methodology for these data is close to impossible.This is in part the case because considerations for post-processing are often at odds with each other, for example, the wish to remove any less-than-ideal data versus the need to keep enough data for good spatio-temporal coverage.This is exemplified by the visually validated radar tracks: a proportion of true positive bird tracks (20%) is removed during post-processing, but the data reliability is increased, as a much larger proportion of false positives (85%) is removed.
The outcome of a processing step might also provide new insights into the radar system and surrounding environment that can be utilised.For example, knowledge gained on spatial observation bias in Module III can inform the user on how to improve the and visualised to find a DoT value where clutter tracks are visually absent.This value depends greatly on the dataset and in previous studies was established between the 0.1st and 10th percentile (van Erp et al., 2021, 2023).Bird tracks with a DoT below the selected TA B L E 1 Overview of modules and processing steps and the effect of each processing step on data volume (number of tracks and percentage of tracks removed/processed) for the case study dataset and the true positive tracks of the validated dataset.
2.1.3| Module II: Improving the quality of bird tracksModule II improves the quality of the true bird tracks identified in Module I to make them suitable for fine-scale behavioural analyses of track properties.Observation errors within the tracks can occur due to anomalies or bugs in the radar tracking software.This can lead to isolated cases of extremely small time intervals between consecutive points and influence the classification of flight behaviour dependent on this information, such as thermal soaring(van Erp et al., 2023).Even if there are no clear indications of errors being present, it is good practice to verify the data quality before performing more in-depth behavioural analysis.Errors can be identified by determining the time interval between consecutive points in a track.Intervals that are considerably lower than the sampling frequency of the antenna likely indicate false observations.For example, an observation frequency of 0.5 Hz results in an approximate observation interval of 2 s for a flying bird.By setting the threshold for a false observation to 10% of this interval (0.2 s), we allow for considerable deviation from the 2 s interval due to bird movements, but still identify (near) instantaneous time intervals.Erroneous time intervals are linked to two observations; as there is no straightforward way to verify which was false, both points should be removed from the track.2.1.3.1 | Case studyThe time interval between consecutive points of each true bird track was calculated.The threshold for false observations was set at 0.12 s, 10% of the observation interval for static objects (rotational speed of the 3D-Fix horizontal antenna = 1.2 s per rotation).18.734 of the 554.608 true bird tracks (3.4%) had one or multiple false observation points identified (20.653 points of the total 1.230.214F I G U R E 3 Bird tracks (black) included in the 0-0.2nd percentile (a) and 1st-1.2ndpercentile (b) of displacement over time (DoT).Turbines of Luchterduinen wind farm are indicated as red dots.Within the green rectangles in panel (a) clear patterns of clutter tracks can be distinguished, which are absent in panel (b).

F
I G U R E 4 (a) Temporal overview (hourly scale) of bird radar data (Luchterduinen Offshore Wind Farm, 3D-Fix) captured between 8 and 15 June 2020.Grey bars depict clean bird track count per hour (left Y-axis) and average masking intensity per hour is shown as a black line (right Y-axis).A clear indication of negative observation bias (NOB), caused by increased masking intensity, is seen on 8 June.(b) A scatter plot of hourly clean bird count against masking intensity shows declining bird counts with increasing masking intensity, hence NOB at higher values.This relation is estimated by a GAM (blue line).The first derivative of the trend line (red line) is used to set the threshold when the bias is considered too large (minimum of first derivative at 0.243, red dashed line).(c) The same temporal overview as panel a, now with the threshold for NOB shown (red dashed line, 0.243).Red bars show time periods where masking intensity is above the established threshold and the unreliable data is removed.
The relation was modelled by a GAM (formula = clean bird count ~ masking intensity) and visualised over the data together with the first derivative (Figure4b).A decrease in hourly bird count is seen around a landmask value of 0.2, a decline which continues until most of the data show near-zero birds (landmask = 0.3).We calculated the landmask value where the first derivative was at its maximum to find the value where the decrease was largest and used this as the threshold for data exclusion (landmask = 0.243).We considered the hours where this landmask threshold was exceeded to have unacceptable NOB (Figure4c), and these hours were therefore removed from the final dataset (236 of 719 observation hours).A total of 12.309 clean bird tracks occurred completely within these hours and were removed.2.2.2 | Detecting spatial observation biasSpatial observation bias can still be present in the data if it is caused by features not known a priori and is, therefore, unaccounted for in Module I.In this processing step additional areas where observation bias occurs are identified.This is done by assuming total bird abundance over the AoI is only affected by distance from the radar due to decreasing TDP (Figure2b,c) and looking for outliers of this assumption.Note that using data covering a considerable time period or from events in which bird flight is assumed to be homogenous (such as mass migration) will greatly improve this step as the data will better match this assumption.By creating a spatial raster covering the AoI, the number of tracks per raster cell can be counted to show bird detection within the AoI (Figure5a).The resolution of the raster should be fine enough to distinguish spatial patterns in detection rate, but coarse enough that the sample size per cell stays large enough.Additionally, the minimum cell size is limited by the spatial resolution of the radar system.The relation between bird counts and distance from the radar per cell can be estimated by a GAM (Figure5b).Cells showing unusually low bird counts (i.e.far below the mean value of the GAM trend line), may be the result of nearby features reducing bird detectability or because these cells occur near the border of the AoI.For cells showing unusually high bird counts, visualisation of the bird tracks occurring in these cells is advised to determine F I G U R E 5 (a) Spatial overview of number of birds per 100 × 100 m cell (blue to orange hue).The black outline depicts the AoI, black dots depict the individual turbines of Luchterduinen Offshore Wind Farm, the red dot indicates the location of the radar.(b) Scatter plot of bird count against distance from the radar for each cell.The trend is estimated with a GAM (middle dashed purple line).The purple area shows data lying within the predicted bird count ±10 times the standard error of the prediction.The lower limit of this area (solid purple line) was used as lower threshold for detecting cells with negative observation bias.(c) Same overview as a, but now the cells identified as having negative observation bias are removed.Removed cells were located around the wind farm turbines and along the edges of the AoI.
number of tracks is due to clutter missed in Module I, causing POB, or due to high local bird flight activity.
proportion were identified in Module I because they were located within 1000 m of the radar (78 tracks, Figure S5) or had an airspeed <5 ms −1 (23 tracks).Of the false positive tracks, 85% (78 of 92 tracks) were removed during post-processing.Step 1 of Module I removed 77 tracks, and one track was removed in step 2 of Module I.

TA B L E 2 (
Overview of all parameters and recommendations for choosing the parameter value.Parameters with an asterisk are set during data processing based on visualisations and analysis of the data.width and arc length of the radar beam at maximum distance; take the highest valueAngles of the radar window where the radar beam is blocked by features(2) Removing non-bird movement Minimum average airspeed 5 m s −1 for general purposes, can be further specified based on species of interest Maximum average airspeed 30 m s −1 for general purposes, can be further specified based on species of interest Minimum displacement over time (DoT) * Based on visualisations of the lowest percentiles of DoT tracks.Set at point where no clear occurrence of clutter is present width and arc length of the radar beam at maximum distance; take the highest value Exclusion threshold * Based on data visualisation and the relation between track counts per cell and distance from the radarextracted from bird radars.The associated R-package birdR (De Groeve & van Erp, 2023) makes the application of this framework accessible for users with a range of goals and data processing experience.We demonstrated the framework by implementing a workflow for data from a radar system installed at an offshore wind farm.This environment is extremely challenging due to the dynamic reflectivity of the sea surface and the nearby wind turbines.The application of the workflow resulted in a significant reduction of data in the case study and limited the area covered by the radar to the most reliable region.This led to a different outcome in the comparative analysis of track densities inside and outside the wind farm and shows that post-processing can have important implications for ecological inference.Furthermore, by applying the workflow to a dataset of ally validated radar tracks, we demonstrate that the majority of true positive bird tracks are maintained during post-processing, while a large proportion of false positives is removed.The three modules prepare the data for distinct types of analyses.Module I can be implemented based on a priori knowledge and improves the data to allow for reliable exploratory analysis and summary overviews.This module can be readily applied to other bird radar systems and could be used in near-real time applications such as air traffic safety warnings(Colón & Long, 2023;van Gasteren et al., 2019).Additionally, the data reduction in Module I reduces the computational demand for in-depth analysis and is therefore critical to making the data workable.In contrast, Module II improves the quality of bird tracks by correcting their geometries.This makes identifying fine-scale flight behaviour possible and provides new possibilities for utilising bird radar data, such as studying offshore thermal soaring (van Erp et al., 2023).Lastly, Module III removes spatial and temporal data sections of lower reliability due to environmental factors.It is, therefore, vital for users interested in studying time series or spatial patterns in bird flight in cluttered environments.

F
I G U R E 6 (a) Spatial overview of number of birds per 100 × 100 m cell (blue to orange hue) for the unprocessed radar data.The black dots depict the individual turbines of Luchterduinen Offshore Wind Farm, the red dot indicates the location of the radar.(b) The same spatial overview, zoomed in for comparison with the spatial overview of the post-processed data (Figure 5).The black outline shows the area of inclusion as applied for post-processing in Module I, step 1. definition of the AoI in Module I and aid subsequent analyses.Our case study showed that bird observations become difficult where the signal is blocked by several rows of wind turbines, limiting the distance to which bird flight can be measured reliable inside the wind farm.Alternatively, this information can improve the future positioning of radar systems, especially in cluttered environments.For example, for comparisons of flight behaviour inside and outside wind farms the radar should be positioned at the edge to provide an unrestricted view outside the wind farm.Despite the proprietary algorithms and slight differences in characteristics, bird radar systems all record target position and time to track birds.By focusing on the utilisation of this information, our framework is applicable to a wide variety of different radar datasets.Furthermore, the modularity of the framework makes it straightforward for slight adjustments depending on the research question and data structure.Nevertheless, post-processing parameterisation might differ between radar systems and environments (e.g.see van Erp et al., 2023 for two radar locations).To further facilitate comparison of results among studies or different radar systems, workflows should be shared and well-documented.Studies where different radars are validated and cross-calibrated to try and harmonise their observations(Liechti et al., 2019;Nilsson et al., 2018), should also be extended to also include the post-processing workflows.Bird radar is a powerful remote sensing technology to study bird flight.The unique combination of properties makes it an important tool to obtain high resolution tracking data of multiple birds flying in remote areas of interest, for example, to assess the impact of offshore wind farm developments on bird life.Given the increasing demand for continuous and large-scale biodiversity monitoring(Schmeller et al., 2017), we believe bird radars will see increased usage and continued development.However, the cumbersome data and clutter issues can turn potential users of radar systems away.This framework aims to address these issues.It allows users to standardise where possible (e.g. the definition of concepts, the ordering of the modules and processing steps within them) while allowing for flexibility where required (e.g. the specific parameter settings and model choices to filter the data).Due to its modular nature, new processing steps can be added, or redundant steps removed to account for future developments.We believe the framework will increase the quality of future bird radar studies, broaden the range of possible applications and encourage more scientists to use bird radars in their research.AUTH O R CO NTR I B UTI O N S This study was conceived and designed by Jens A. van Erp, Judy Shamoun-Baranes and Emiel E. van Loon.Jens A. van Erp designed the final framework, supported by Judy Shamoun-Baranes and Emiel E. van Loon.The case study was performed by Jens A. van Erp.The R-package birdR was designed by Johannes De Groeve and Jens A. van Erp.Jens A. van Erp led the written manuscript, and all authors contributed to the writing of the manuscript and gave final approval for publication.Judy Shamoun-Baranes acquired funding for this project.