We developed FACA to accelerate the application of the well-established co-alignment method, enabling streamlined deployment at any location with evolving temporal data, minimal preparation, and high adaptability and scalability. The software enhances both the efficiency and extendability of the co-alignment approach. Our initial goal for FACA was to support our research on the Baltic Sea coast, where we investigate mass movement potentials along cliffs in pilot regions, and aim to transfer this knowledge to other areas . Data collected during the multiple field campaigns required rapid analysis, interpretation, and dissemination to partners and decision-makers. Additionally, we sought a workflow that could easily be adapted to other coastal sections that was effectively applicable by non-photogrammetry experts. As a result, FACA emerged as a user-friendly software solution capable of producing reliable results quickly. Although initially developed for UAV imagery, FACA is suitable for aligning large datasets of images, regardless of scale or capture method, provided that adequate computing resources are available. The FACA code was written, commented, and licensed to allow for easy modification by other practitioners.

FACA, like all variations of co-alignment, leverages computer vision algorithms that underlie photogrammetry software. These techniques e.g., detect key points – distinctive, invariant locations in an image – and match them across all images. outlines four stages in this mode of operation: detecting, describing, and matching or tracking key points. Tracking is an alternative to matching limited to likely key points, such as those in similar areas across images. In the case of Agisoft Metashape, a proprietary variant of the broadly used SIFT algorithm is employed. Its default matching settings use a priori knowledge about image locations to preselect images for matching, along with other metadata information, e.g., camera parameters and light conditions. This differentiates the photogrammetry workflow from the core computer vision algorithms while enhancing performance and precision . Images from individual UAV surveys, or general field campaigns or sessions, are stored in camera groups; these are all stored in one unit, or “chunk” in Metashape. Camera groups allow for survey-specific camera parameter optimizations. Tie points are key points that the algorithm matches across two or more images, establishing spatial relationships when aligned. When tie points span images from different surveys they align these surveys together, the eponymous co-alignment. From this point, SfM algorithms then use these tie points to optimize camera position and orientation through bundle adjustment. Finally, tie points are employed in the 3D reconstruction process, generating a sparse point cloud which can be filtered and refined within the original chunk. Afterwards, the surveys are separated, creating individual chunks and discarding key points not matched with its images while preserving location information and camera group specific calibrations. These new sparse point clouds are then processed separately to generate final products such as point clouds in FACA’s case.

FACA introduces flexibility compared to the original co-alignment workflow described by , offering several optional steps that users can choose while maintaining the same foundational process. For instance, prior to key point matching and alignment, FACA allows users to adjust the accuracy of image location data, which can enhance co-alignment performance in cases where camera location information is uncertain across different surveys. This and other differences stem from the configuration possibilities FACA offers, see Table . FACA also differentiates itself from previous co-alignment works by not requiring human intervention, a path forward that already discussed.

FACA uses the Python API of Agisoft Metashape Professional version 2.1.0 to automate the co-alignment workflow and is compatible with versions 2.0.0 and later. FACA only requires a Python 3 installation with the standard library, an installed Metashape Python 3 Module, and a licensed copy of Agisoft Metashape Professional.

Every co-alignment relevant parameter that can be modified in the Metashape interface can also be altered via the Python API . Metashape also provides the ability to process batches, created using its main interface. This batch process is not a fully-fledged replacement for the Python API. In comparison it lacks many features, e.g., modifying existing chunks and labeling objects, in addition to losing the flexibility provided by Python, e.g., a custom interface and to dynamically name outputs. With the Metashape Java API , there exists a third option on par with the Python API. We use Python because of our familiarity with the language and its popularity for Metashape scripting.

FACA imposes minimal rules on the input data and its structure, ensuring a smooth, automated workflow while maintaining adaptability for various use cases. The input directory must contain at least two subdirectories, each with one surveys images as .jpg files, allowing for any number of subdirectories as required by the user, without limiting the complexity or organization of the data hierarchy. All images must contain location information in their metadata. The name of the first level subdirectories will be used to reference the images they contain, as well as the outputs generated from those images. We found it convenient to name them based on their survey date.

FACA allows the use of predefined parameters, which are stored as sections within a configuration file. The default configuration file included with FACA contains a section that defines overarching parameters, such as the image input directory, output directory, and EPSG code for the output point clouds. These default values can be overridden by redefining them in other sections.

There are three ways to interact with FACA. The graphical user interface (GUI) (Fig. ) uses the tkinter package and is the most accessible, but not as automatable as the alternatives. The interactive mode, where the software asks for each parameter, while providing the defaults as a suggestion and fallback. And, the script mode, where the user defines a configuration file and a section therein that FACA uses for the calculation. The user can optionally define parameters that differ from those specified in the configuration file section, for example, to apply established settings to an area with a different EPSG code or to modify input or output directories. If neither a configuration file nor a section is provided, but at least one calculation parameter is specified, the interactive mode will start, omitting prompts for the provided parameters. If all calculation parameters are provided as arguments, the calculation will start exactly as in script mode. Script mode allows FACA to be part of a larger automated workflow, e.g., applying filtering or color correction on the input images or removing vegetation and calculating statistics on the output PC. The modifiable configuration allows users to easily apply proven co-alignment workflows on their data, as well as find ideal workflows for their study areas with a sensitivity analysis. Configuration files can be created and modified manually or with the help of the GUI. For any other available setting, FACA uses the Agisoft default values, as specified in .

FACA records its steps and interim results, e.g., the number of detected surveys, corresponding images and tie point counts, the Metashape version used, and calculation parameters, in a time-stamped log file within the output directory. These small text files can be easily shared with peers to facilitate reproducibility, enhance workflow transparency, monitor performance, and assist in identifying and troubleshooting errors or issues. Additionally, the logs can be automatically parsed to generate figures for publications or to facilitate the automatic comparison of different parameterizations .

Parameters and possible values supported by FACA are listed in Table . Our streamlined approach allows us to quickly apply multiple co-alignment workflows discussed by other authors and examine their general transferability to other areas. FACA ships with multiple co-alignment parameters used previously in addition to its own defaults ready for use as part of its initial configuration file . FACA emulates the parameters of these co-alignment approaches as far as they do not require additional manual inputs from the user and do not use depreciated features of Agisoft Metashape.

FACA users can choose any settings they want, either from a modifiable list based on previous works or by selecting their own. FACA default parameters emphasize practical automated applications of co-alignment. Because of this, they are conservative, especially the tie point filtering, as manual fine-tuning is not expected. We see these parameters as reasonable suggestions and invite practitioners to expand upon them for their needs.

The Image Alignment Accuracy setting controls the level of image scaling applied during camera position estimation . Undersampling (Medium, Low, and Lowest Image Alignment Accuracy) the images normally offers faster processing at the cost of detail and accuracy. Oversampling uses more detailed images to provide precise results, but does so at the cost of roughly twice the computing time compared the unmodified images. Using the original image resolution (High Image Alignment Accuracy) is a reasonable compromise.

Camera accuracy defines the expected positional accuracy of each camera's coordinates . The default of 10 m maintains some flexibility when using tie point matching with posterior knowledge from image metadata. This is crucial, as the accuracy of image location data can vary between UAV flights, often exhibiting different systematic offsets across sessions.

Key points are significant features detected in an image and tie points are key points matched across images; their limits define the maximal amount of each to consider during processing . High or no key- and tie point limits can introduce noise into the sparse point cloud and increase processing time, while low values can lead to missing features. We use the default values of 40 000 key points and 4000 tie points per image , because we found larger values take more time to process with no noticeable quality differences.

Rule-based tie point filtering can automatically identify and remove inaccurate or erroneous points, thereby reducing noise, optimizing following bundle adjustments, and improving processing efficiency while enhancing the overall precision in the final product . Agisoft Metashape Professional offers four built-in tie point filtering criteria: “Image count” filters tie points based on the amount of image in which they are visible and “Reprojection error” uses the maximum difference between measured and parameter adjusted coordinates of a tie point, normalized by the scale used. “Reconstruction uncertainty” uses the ratio of the largest to the smallest semi-axis of the error ellipse of tie points. “Projection accuracy” uses the average image scale used when getting the tie point projection coordinates divided by the number of images containing the tie point. We filter tie points that are not visible in at least three images, have a reconstruction uncertainty over 50, and projection accuracy over 10.

A depth map is generated for each image, representing the relative distances of objects from the camera's viewpoint . The depth map quality settings defines its resolution, and thus the detail and accuracy of the output geometry. “Ultra high” quality uses the images original resolution and each quality step quartering its pixel count, all the way to 1/256 at lowest depth map quality. We found the medium quality (1/16 resolution) to be suitable for use with UAV imagery.

Filtering the depth map removes tie point outliers, and can reduce noise in the output, but smaller details can be lost. In their documentation, Agisoft suggests “Aggressive” for drone images , which is why we employ this setting.

Figure  shows the processing steps applied when running FACA with n different sessions. Every calculation expects all images taken during a flight session inside a directory, which, in turn, reside in a common main directory. The lower part of Fig. , below 'Copy Chunk n times', is simplified to show only one individual session i, but is applicable for all i from 1 to n.

All images are loaded into the original chunk, where each session is organized within its own camera group, to facilitate individual camera parameter optimization. Each camera group is named according to its survey's subdirectory. Subsequently the image location accuracy can be optionally adjusted. Options include using the default value of 10 m in each direction (longitude, latitude, height), setting individual values for each direction, or utilizing precise measurements, e.g., from RTK equipment, if available. Following this, all images get matched based on the specified alignment accuracy, key point limit, and tie point limit. This step aligns all pictures to generate an initial sparse point cloud. The sparse point cloud can then be optionally thinned using any combination of tie point filtering criteria available in Metashape, while optimizing the alignment. FACA creates a separate copy of the original chunk for each flight survey, removing other session images. Thus, each chunk contains only the images from the respective session, organized within a single camera group, and a thinned sparse point cloud that only contains tie points visible in that survey. FACA then generates a depth map from these tie points with the specified quality and filtering mode. This depth map is used to construct a dense point cloud. The dense point cloud is subsequently exported as a .las file in the coordinate system defined by an EPSG code. The name of the final PC output is derived from the original subdirectory name of the session images.

We applied co-alignment with FACA in two study areas along Germany's Baltic Sea coast, see Fig.  and Table . The Sellin study area is located northeast of the town, along the western coastline of Rügen island. Its cliff consists of thick glacial–fluvial to glacial–limnic sands covering small amounts of till at the foot . The bare, sandy plateau at the cliff top in the northern quarter contrasts with dense trees in the south. The study area spans approximately 600 meters of coastline, from coastal kilometer R089.030 to R089.630 .

The Wustrow study area lies west of Niehagen and Althagen on Fischland, along the western coastline. It is a cliff consisting predominantly of till, partly overlain by eolian sand and cliff-edge dunes . Landwards there is sparse forest in the southern quarter and agriculture in the north. The coastal retreat is progressing very quickly here due to wind exposure, wave erosion, weak silt material, and its shoreline parallel joint direction . North and south of the study area exist coastal protection structures, e.g., groins and breakwaters . Attempts to stabilize the coast line in the study area failed in the past . The study area extends over about 1750 m of coastline, from coastal kilometer F177.350 to F179.100. .

Both study areas experience natural coastal erosion and are affected by landslides, leading to rapid changes in the cliff faces with minimal changes inland.

We used a DJI Phantom 4 RTK in Terrain Awareness Mode (TAM). The drone is equipped with a 1^′′ CMOS 20 MP image sensor. RTK was enabled for all surveys. We took Nadir and oblique (45°) images along the same flight route. We combined the official German digital terrain model with 5 m resolution (DGM5) with the German Combined QuasiGeoid 2011 (GCG2011) to program the missions. TAM allows consistent ground sampling distance (1.1 cm per pixel at 40 m above ground for Nadir images) and image overlap. The automated flight plans were created using the DJI GS RTK App that came with the UAV remote controller. We planned for 70 % vertical and 80 % horizontal overlap of these images. Table  shows information concerning the eight field campaigns. Throughout most sessions, we supplemented the automatic pictures with ones taken manually. The were not primarily taken for Co-alignment, but for documentation purposes and to support other research areas in an overarching project . Decisions to capture additional images were based on identifying significant changes in the cliff that warranted further documentation, as well as the availability of time in our project schedule and suitable weather conditions for UAV operation. The flight time is taken from the metadata of the first and last image taken in the area during that field campaign. In the first field campaign we only flew the southernmost automatic flight mission in Wustrow (10 526 m²). Before Field Campaign 6, in October 2023, we received a second Phantom 4 RTK – allowing us to fly parallel sessions. This speeds up the survey time and can reduce lighting differences between photos of the same survey. Starting with Field Campaign 7 in November 2023, we supplemented our automated flights with manual RGB flights conducted using a DJI Mavic 3M. However, we utilized images from these manual flights exclusively for generating point clouds in the Wustrow bunker area, due to the complex structure following the bunker's topple. Images of Field Campaign 7 present an additional challenge for our workflow as the ground is covered by snow. Because of this we did not examine Sellin but chose to survey Wustrow, as its steeper cliffs remained exposed.

After each field campaign, we reviewed the images and compiled a list of unsatisfactory photos, e.g., poor focus, drone parts or only sea visible. We excluded these from further processing.

Our analysis focuses on the multi-temporal local precision and not the model's global accuracy. We analyzed the local precision of our results in two ways: firstly in an area without noticeable changes and secondly in areas with significant changes. For our stable area we used a 60.5 m² section of a rooftop unaffected by coastal erosion in the Sellin study area (54.3794° N, 13.7023° E). While the roof is in the border area (Fig. ) it is the only sufficiently sized stable and flat area in the study area. In the stable area, we measured the cloud-to-cloud (C2C) distance of the PCs in the z (height) dimension compared to the first survey. C2C was selected for its parameter-free nature. Its use was further justified by the high and uniform point density of the relatively flat roof, despite its location near the edge, and the focus on vertical differences. The roof's location at the edge of the study area could lead to slightly exaggerated C2C distances, as the image coverage is not ideal compared to the center. We estimated the model's uncertainty with the range of distances in stable areas across surveys. Those with the lowest differences are considered best. There are no suitable anthropogenic structures in the Wustrow study area, thus we did not measure unaffected areas there. Table  contains the area with noticeable changes. The best models would accurately represent the changes we saw in the field in high detail. The coastal retreat should be recognizable over several flights, without outliers that reverse the trend.

Besides these quality factors, we also measured the processing time, taken from the FACA log-file. We performed every calculation on the same workstation (CPU: Intel Xeon E-2286G; RAM: 128 GB DDR4 (2667 MHz); GPU: 16 GB Nvidia RTX 5000; HDD: HGST HUS726T4TALE6L4; OS: Microsoft Windows 10 Enterprise 22H2). The computation times should be regarded as relative references, not absolutes, as they can vary significantly depending on the available computing resources.

Figure  shows the location of the stable area (pink filled polygon in Fig. a), the area used to analyze changes in Sellin (green polygon in Fig. a), and the minimum bounding box of the images used to analyze the changes around the bunker (blue polygon in Fig. b) as well as the one used to quantify the coastal retreat (red polygon in Fig. b). The yellow polygons indicate the minimum bounding box of the automatic images taken during this field campaign. The stable area in Sellin lies just outside the minimum bounding box of images taken during automatic flight. The unstable areas in Wustrow (bunker in the south, coastal erosion in the north) are well covered by automatic flight images.

The bunker serves us to visually identify and describe coastal erosion processes, especially after its collapse on 18 February 2024 (between Field Campaigns 7 and 8).

We used subsets of the Wustrow pictures to analyze the bunker and coastal erosion areas, see Table . Both datasets are freely available under the CC-BY-SA 4.0 license . They are the images that were taken within the defined bounding boxes (blue and red polygon in Fig. b) of the unstable areas.

We quantified the coastal erosion in Wustrow by tracing the cliff top in the area of interest (see red polygon in Fig. b) with using the generated dense point clouds to create polylines. Then we exported them as Shapefiles and proceeded in QGIS . In the geographic information system, we generated a line grid with the extent set to be slightly larger than the traced cliff tops, 1 m horizontal and 200 m vertical spacing. We discarded all vertical lines and rotated the horizontal ones 20° clockwise, so they are perpendicular to the coastline. Finally, we split the rotated features with two cliff tops and removed the parts outside of this region. Doing this resulted in roughly 380 measurements of coastal erosion per parameterization. Knowing the days between the UAV surveys, we interpolated the yearly coastal erosion. We used the same rotated line grid for all measurements.

We analyzed the changes in the Sellin cliff by clipping the area of interest (green polygon in Fig. a) out of the remaining dense point cloud. The underlying Sellin image dataset (CC-BY-SA 4.0 license) is available . In this area, we calculated the M3C2 distance for each pair of sequential surveys. We used M3C2 here because it is robust and allows us to measure the direction of changes. C2C was deemed unsuitable here, because it does not account for surface orientation and provides only unsigned distance measures. We always used the older of the two compared PCs as the basis for sub-sampling to generate core points with 0.25 m spacing. We configured M3C2 to use multi-scale normals between 0.5 and 4.5 m with 1 m step size and a projection diameter of 0.5 m. The clipping and M3C2 calculation were done with .

FACA automatically logs the timestamps when generating a model. We used these logs to calculate time expenditures and compared them. The best model would both generate quickly and offer good local precision.

For an overview of parameters used in this example application of FACA see Table . We chose the parameters discussed in these works because of their range of values and good documentation in their respective articles. This work does not aim to evaluate the results of these previous studies but wants to show the capabilities of FACA across a wide range of parameters.