A workflow of open-source tools for drone-based photogrammetry of marine megafauna

Sensor calibration

Many commercial drones (e.g., DJI) automatically process and reformat wide-angle imagery to correct for distortion, which produces different field of view (FOV) and camera parameters from what is reported by the manufacturer. Such a discrepancy between reported and actual camera specifications can lead to errors in GSD calculation, causing increased measurement errors due to incorrect scaling (Eqs. (1) and (2)). Furthermore, different camera resolution settings between collected video or imagery by the same drone will yield different adjusted FOV and camera specifications than reported by manufacturer (Segre, 2025). While collecting images in RAW format should typically be immune to this internal processing, many photogrammetry software require RAW images to be converted into a compressed, lower resolution image, such as .jpeg or .png file. Thus, before collecting data in the field, researchers should confirm that the reported specifications match the actual specifications of the camera settings they intend to use. Some studies use software to correct for image distortion (Burnett et al., 2018) and calculate an “adjusted focal length” (e.g., Dawson et al., 2017). Segre (2025) provides a step-by-step protocol to manually check the effect of automatic image processing by calculating the true, corrected FOV for each camera-resolution combination. Alternatively, researchers can simply calculate an “adjusted focal length” (f_adj) in a controlled setting (drone not in flight) by measuring a known-sized object at a fixed distance (i.e., distance, or “altitude”, between object and camera sensor on the drone resting on a table) by rearranging Eqs. (1) and (2) to solve for f: ${\displaystyle f_{adj}=\frac{a\times S_w\times L_p}{L\times I_w}.}$

Note that following the same logic, an adjusted sensor size can be calculated instead of an adjusted focal length. Researchers should check to see the effect, if any, of automated processing and distortion for each resolution they intend to use and determine if they should use their calculated adjusted specifications instead of what is reported by the manufacturer.

LiDAR Altimeter

The accuracy in the recorded altitude is often a major contributor to photogrammetric error. Many researchers self-install a LiDAR altimeter system to their drone to increase the accuracy of recorded altitude (Dawson et al., 2017). LidarBoX is a 3D printed open-source hardware enclosure for a LiDAR altimeter system that can be swapped and installed between multiple off-the-shelf drones, e.g., DJI Inspire 2, Mavic 3, Phantom 4 (Bierlich et al., 2023a). Custom-built LiDAR attachments for a variety of drones are also available (i.e., O3ST (https://www.o3st.com/). Prior to take-off, researchers should install and initiate a LiDAR altimeter system to record altitude (Fig. 1). Note, the distance from the LiDAR to the camera lens should be added/subtracted to the recorded LiDAR altitude. Incorporating real-time kinematics (RTK) technology with drones is another option to help improve accuracy in altitude and position (Ekaso, Nex & Kerle, 2020; Román et al., 2024), although this system is restricted to terrestrial work or when flying from shore.

In some cases, adding a LiDAR altimeter (or RTK system) to a drone is infeasible, and thus using the barometer is the only option. Bierlich et al. (2021b) developed the Bayesian statistical model in part to provide methods for incorporating measurement uncertainty when only using a barometer. Development and application of this Bayesian model ensure measurements are accurate, robust, and comparable to measurements that were collected using a LiDAR altimeter. We converted this Bayesian framework into the Xcertainty R package, which we incorporate into our presented workflow (Fig. 1). See ‘Flying over a calibration object’, ‘Incorporating uncertainty’, and ‘Propagating uncertainty of morphometric measurements’ for details on collecting training data, setting up Xcertainty, and propagating measurement uncertainty into analysis (Figs. 1–3).

GPS times

Since the LiDAR unit is often a separate system from the drone camera (and most cameras do not use GPS satellite time), the timing of recorded imagery needs to be synced with the timing of recorded altitude. Failing to collect GPS time can lead to a mismatch of timestamps between the camera and LiDAR, causing improper scaling of images due to incorrect altitudes used in calculating GSD (Eqs. (1) and (2)) (Raoult et al., 2020). Unless the drone model integrates LiDAR and camera units on-board, the two data streams need to be integrated during post-processing, which can be achieved by collecting a photo or video of a displayed high precision GPS timestamp prior to each take-off (Figs. 1 and 2). The offset between the displayed GPS time and the recorded camera timestamp is used to sync the LiDAR altitude with videos or images. CollatriX v2 is open-source software with functions to automatically match altitude with the camera imagery corrected by the GPS timestamp (Fig. 1) (Bird & Bierlich, 2020).

Launch height

While LiDAR altimeters provide greater accuracy in altitude measurement, they are prone to recording nulled or erroneous values. For instance, Bierlich et al. (2021b) reported that 17% of all measurements contained nulled laser values, which may be due to refracted laser pulses through the water or scattering due to atmospheric particles, including mist. Thus, less accurate barometer altitude may be required in cases where the LiDAR unit fails. Since barometer altitude is zeroed upon launch, launch height—measured distance from the water surface to the camera lens—must be recorded before take-off and later added to the recorded barometer readings (Burnett et al., 2018) (Figs. 1 and 2). Adding launch height is particularly important when drones are hand-launched and recovered from boats.

Fly over calibration objects

Calibration objects are often small, known-sized objects that are easily deployed from research vessels, e.g., a wooden board attached to a line (Fig. 2). Training data should be collected with the calibration object floating flat at the surface of calm water at altitudes that encompass the range of altitudes expected during flight over the target species, with the camera fully nadir (i.e., pointing straight down). While the frequency of collecting training data depends on the drone system, the weather conditions, and project logistics, it would be ideal to collect training data at least once per day of fieldwork to help capture variability in altimeter accuracy due to changing weather and sea state (Bierlich et al., 2023a). If once per day is infeasible, we advise collecting training data at least once per field season. However, for drones with large variability in measurement error (e.g., no LiDAR), it may be best to collect training data more frequently (i.e., per flight) to help improve overall uncertainty predictions. Since different camera resolution settings can yield different levels of error for the same drone (see Segre, 2025), studies should use training data that appropriately match the observation data—i.e., with the same camera settings, pixel resolution, altimeters, and recording either videos or photographs. As such, the same drone system may have multiple sets of training data that correspond with the settings used during data collection. The training data and measurements of the animal are then input into the Xcertainty R package to generate posterior distributions for each measurement, which can be used to describe the measurement and its associated uncertainty (see section ‘Incorporating Uncertainty’) (Figs. 1 and 3).

Fly over animal

Raoult et al. (2020) provides guidance for flying over several different marine species. Here we provide a few additional details to consider. While flying, it can be challenging to spot and stay-over mobile animals, particularly marine species. Many drone systems use dual camera systems with the ability to switch to a second camera with a wide-angle view to help increase the efficiency of staying over the target animal at full frame. For drone systems with a single camera, it may be helpful to start a flight at a higher altitude (to increase the field of view) until the individual or group is located and in sight, and then gradually lower altitude until the animal is in full frame. Note that the camera must be fully nadir while collecting imagery. To reduce errors from environmental conditions and animal behavior, imagery should be collected when the target animal is in a straight body orientation at, or just below, the surface of the water, with minimal interference from wave refraction or glare. Collecting photographs in bursts or collecting video instead of photographs can create more opportunities to select a specific frame with minimal interference from the environment or animal position.

It is also important to note that different species may have variable responses to drone flights. While aquatic species in general tend to exhibit less of a behavioral response to drone disturbance compared to aerial and terrestrial species, it is important to identify and minimize potential disturbance caused by the drone to the targeted species (Afridi et al., 2025). For instance, smaller cetaceans seem to exhibit more of a behavioral reaction to drones at lower altitudes compared to larger cetaceans (see Afridi et al., 2025 for a review and guidance). Researchers should thus collect data within altitude ranges that cause the least disturbance to their target species and monitor any changes in behavior during flight.

Manual

Once images are ranked and filtered for quality, they are ready to be measured. Objects are measured in pixels and converted to standard units (e.g., meters) using Eqs. (1) and (2). MorphoMetriX is an open-source photogrammetry software developed as a user-friendly graphical user interface (GUI) for customized measurements (in pixels and meters), including lengths, segmented widths, angles, and areas (Torres & Bierlich, 2020) (Figs. 1 and 2). The latest version (v2) can be downloaded from https://github.com/MMI-CODEX/MorphoMetriX-V2. Since its release, MorphoMetriX has been used in a variety of drone-based photogrammetry studies on marine megafauna. While several other photogrammetry software tools are available and can also be used within our presented workflow, such as ImageJ (Schneider, Rasband & Eliceiri, 2012) and custom built software in R (Christiansen et al., 2016) and MATLAB (Dawson et al., 2017; Burnett et al., 2018), we highlight MorphoMetriX as it was designed to provide a fast and easy-to-use GUI that includes a powerful zoom, functions to help enhance measuring, and the flexibility for customization with no required knowledge of any coding language (Torres & Bierlich, 2020). See Supplementary Material for additional details on MorphoMetriX v2.

Automated

Automated applications developed to study cetaceans have typically focused on detection and identification of animals in imagery (Borowicz et al., 2019; Cheeseman et al., 2022), with only a few studies focused on extracting morphological measurements using photogrammetry from drones (Gray et al., 2019; Bierlich et al., 2024) or from photographs of specimens collected from bycatch (Zhang et al., 2023). Gray et al. (2019) used convolutional neural networks to automatically identify blue (B. musculus), humpback, and Antarctic minke (B. bonaerensis) whales and extract estimates of body length from drone-based photographs. More recently, Bierlich et al. (2024) developed XtraX, a user-friendly GUI to automatically extract body length and body condition measurements in pixels and meters from drone-based images with similar accuracy as manual measurements (Fig. 1). Together, DeteX and XtraX provide a pathway for automatically extracting morphological measurements from drone-based videos nearly 90% faster than manual analysis, saving analysts hours of processing time (Bierlich et al., 2024). See Supplementary Material for additional details on XtraX.

Working examples

Xcertainty follows these main steps to produces body length and width estimates: (1) prepare calibration and observation data, (2) build the Markov chain Monte Carlo (MCMC) sampler, and (3) run the sampler. Once body length and width estimates are obtained, several commonly used body condition metrics can be estimated using the body_condition function (Fig. 4). The temporal scale of output estimates can be specified by the user, for example, whether at the daily scale (useful for widths in estimating body condition) or at the annual/seasonal scale (useful for length in growth models). We provide three detailed vignettes following these steps to demonstrate the functionality of Xcertainty. The data used in each vignette is embedded within the R package (https://cran.r-project.org/web/packages/Xcertainty/index.html). The data used in the vignettes were collected under research approved by Oregon State University’s Institutional Animal Care and Use Committee (IACUC, permit no. IACUC-2023-0388) and National Oceanic and Atmospheric Administration (NOAA) and National Marine Fisheries Service (NMFS) (permit no. 16011 and 21678). See Supplementary Material for detailed descriptions of the package’s functions, their inputs and outputs, and code for each vignette. While Xcertainty contains several pre-packaged functions, users can also develop customized functions and samplers for their specific needs, see the Xcertainty GitHub for more details (https://github.com/MMI-CODEX/Xcertainty).

Vignette 1 and 2: Length, width, and body condition

Vignettes 1 and 2 use an example dataset of measurements of gray whales foraging along the Oregon Coast, USA (full details of the study system and data collection protocols can be found in Pirotta et al., 2024). Vignette 1 uses imagery from a DJI Inspire 2 (the red drone in Figs. 3 and 4) and Vignette 2 uses imagery from a DJI Phantom 4 Pro (the blue drone). Both examples demonstrate how to generate posterior distributions for body length and width (Fig. 3) and then estimate several body condition metrics using the body_condition function (Fig. 4). When building the MCMC sampler, users can specify arguments for building prior distributions related to image altitude, altimeter bias, altimeter variance, altimeter scaling, pixel variance, and object length (see Supplementary Material for more details). To enable outputs to be driven by the observed data, we strongly recommend starting with non-informative priors, which includes an overly wide range or distribution than what is typical or expected for each parameter. Vignette 1 demonstrates how to set up the MCMC sampler using non-informative priors, while Vignette 2 demonstrates how to identify when to use and how to set up a sampler using informative priors (Figs. 3 and 4).

The body_condition function uses the posterior distributions of body length and widths to calculate several common body condition metrics with associated uncertainty (Fig. 4): standardized widths are the body widths divided by the individual’s body length (Fernandez Ajo et al., 2023); projected area (m²) corresponds to the dorsal projected area calculated as a series of trapezoids (Christiansen et al., 2016); body volume (m³) is calculated assuming either a circular (Christiansen et al., 2018) or elliptical (Christiansen et al., 2019) cross-section of the animal; Body Area Index (BAI) is a standardized metric where the projected area of a specified region of the body is divided by the length squared of that region (Burnett et al., 2018). While uncertainty increases with dimensional scale (single width to projected area to volume) (Bierlich et al., 2021a), studies should use the metric that most appropriately addresses their research objectives. For example, body volume can be converted to mass if average tissue density is known to then infer energy dynamics (Christiansen et al., 2019; Christiansen et al., 2024), and BAI and single standardized widths can be used to compare relative body condition across individuals and populations (e.g., Barlow et al., 2023; Bierlich et al., 2022; Perkins-Taylor et al., 2024) and can help detect pregnancy (Fernandez Ajo et al., 2023; Cheney et al., 2022). Since the body_condition function easily calculates all these metrics, researchers should consider including the ancillary body condition metrics in their study’s Supplementary Material; these values may be of interest to other researcher’s’ objectives, thus facilitating greater comparison and knowledge sharing.

Vignette 3: Growth curves

Vignette 3 uses the function growth_curve_sampler in Xcertainty to reproduce results from Pirotta et al. (2024) (data available at https://osf.io/4xabg/), who analyzed the growth patterns of 130 individual gray whales over a 7-year period using five different drone systems (Fig. 5). Pirotta et al. (2024) noted that, for many individuals with replicate samples across years, the raw, observed body length measurements from older drones (e.g., DJI Phantoms) with only a barometer were larger and more variable compared to newer drone models (e.g., DJI Inspire) with LiDAR altimeters attached (Fig. 5). To account for large discrepancies in an individual’s measurements over time, Pirotta et al. (2024) developed a growth model to ensure that individual estimated sizes of whales over time were constrained by a growth function, ensuring that repeated estimates of whale lengths were biologically realistic. Additionally, they also accounted for improved drone technology such that growth trajectories were estimated while accounting for the differences in measurement precision and bias from newer versus older drones, improving the precision of estimated true sizes of whales across their measurement histories. The function growth_curve_sampler is based on the von Bertalanffy-Putter growth model fit in a Bayesian hierarchical framework described by Pirotta et al. (2024), and reproduced results are shown in Fig. 5. Users can also create customized functions using different growth models, see the Xcertainty GitHub for more details (https://github.com/MMI-CODEX/Xcertainty).