Camera-equipped unmanned aerial vehicles (UAVs) and structure from motion (SfM) methods are increasingly being utilized as a low-cost method to conduct repeat topographic surveys in order to measure geomorphic change (Fonstad et al., 2013; Eltner et al., 2016; Anderson et al., 2019). To obtain high-quality 3-D models using SfM, precisely located ground control points (GCPs) are typically used (James and Robson, 2014; Carrivick et al., 2016) to both georeference the model and to improve the calculation of camera interior parameters and camera positions and orientations. This requires either the deployment of GCP targets prior to UAV flights or the identification of existing natural or artificial features that can be used as targets. In either case, the locations of the GCPs must be precisely measured, typically using a differential GNSS (dGNSS) or total station (James et al., 2017).

In the absence of GCPs, models can also be created using direct georeferencing, which requires GPS locations of the camera positions (Carbonneau and Dietrich, 2017). For highly accurate results, this relies on having very accurate camera locations, typically by using a UAV equipped with dGNSS (Turner et al., 2013; Hugenholtz et al., 2016). Direct georeferencing performed using only the GPS positions recorded by consumer-grade drones can lead to models that contain a range of errors and distortions (Carbonneau and Dietrich, 2017; James et al., 2017). Model errors can also be reduced by complementing nadir surveys with oblique images in a convergent geometry (James and Robson, 2014), but this is typically recommended in conjunction with GCPs or dGNSS based direct georeferencing. Peppa et al. (2019) presented a method for automatically generating pseudo-GCPs in stable areas using DEM (digital elevation model) curvature and openness, but this relies on using surface texture to estimate stability, which may not be reasonable in all settings. In addition, the generation of DEMs may not be appropriate for all terrain types, such as overhanging cliffs. Feurer and Vinatier (2018) introduce a method to process sets of archival aerial photographs in the same SfM block to achieve accurate change detection with only a small set of poorly constrained GCPs (accuracy ∼20 m) for scaling and georeferencing.

When considering accuracy in relation to change detection, we distinguish two different types: the real accuracy of an individual model and the comparative accuracy of a pair of models. Real accuracy includes both relative and absolute accuracy or the internal accuracy (distortion) of the model and the accuracy of the scaling and georeferencing of the model. We use the term comparative accuracy to describe the accuracy of the change measured between model pairs or to what degree the models are consistent with each other. High real accuracy should lead to high comparative accuracy and is the most desirable outcome, but it may be possible to achieve high comparative accuracy for model pairs with low real accuracy. For example, if two models are subjected to the same incorrect transformation or rescaling, their real accuracy will be affected while their comparative accuracy remains unchanged.

While high real accuracy is desirable, some settings of interest for change detection preclude the deployment or measurement of GCPs, and dGNSS-equipped UAVs may be prohibitively expensive. Therefore, an alternative method for achieving high comparative accuracy of survey pairs could open up new types of settings to event monitoring using low-cost UAVs. Here, we introduce a simple workflow involving the co-alignment of photographs from different surveys; our method is similar to that of Feurer and Vinatier (2018) but is generalized to any set of repeat SfM surveys and requires no GCPs. Using data from two contrasting study areas – a bedrock gorge in Taiwan and a steep cliff coast in northern Germany – we demonstrate that we can achieve high comparative survey accuracy and low limits of change detection using a low-cost off-the-shelf UAV without ground control points. Our workflow is extremely simple, can be performed entirely with the software Agisoft Photoscan Pro (now called Metashape Pro), and could be made fully automated.

We first present data from the Daan River gorge, a bedrock canyon in Taiwan. In this system, the river experiences large changes between survey periods, while the surrounding area has variable degrees of vegetation cover and remains stable aside from vegetation growth. The gorge also experiences localized erosion of its steep to vertical walls. An extensive description and analysis of survey accuracy in this setting can be found in Cook (2017), who estimated a level of detection of 10–30 cm (depending on surface characteristics) for GCP-constrained surveys. Because we have ground control information for these surveys, we can compare GCP-constrained changes to changes measured using our workflow without GCPs.

The primary study area is located in Jasmund National Park on the island of Rügen, Germany, where steep to overhanging coastal cliffs up to 118 m high are eroding rapidly (Schulz, 1998) (Fig. 1). Our study area comprises about 7 km of coastline, from the Königsstuhl in the north to the town of Sassnitz in the south. The cliffs, composed of chalk and glacial till, experience frequent rockfalls and collapses during the winter months. During our study period from 2017 to 2019, these failures varied in size from a few cubic meters to about 4000 m3. While rockfalls are relatively common, they affect a small proportion of the total cliff area, and the rest of the cliff face remains stable, with no discernable internal deformation.

This cliff coast presents a challenging environment for UAV-based surveying. The cliff sections are out of bounds, access to the base of the cliffs is limited and can be dangerous, the forest above the cliffs limits both ground visibility and the communication range of the UAV, and strong winds are common. In addition, the coast is a long linear feature that precludes complicated flight patterns, and flying close to the cliff is restricted to protect peregrine falcons nesting there. However, because cliff collapses can represent a significant hazard to National Park visitors, there is a strong interest in a rapid and easy to implement method of monitoring cliff activity. This combination of characteristics makes the cliff a good location for demonstrating the applicability of our workflow, as it is a setting in which conventional methods are unsatisfactory.

The Daan River surveys enable us to compare the effectiveness of the co-alignment workflow without GCPs to a traditional workflow using GCPs. We find that the co-alignment workflow results in a change map and density curve that are almost identical to those produced using the GCPs (Fig. 3). The only apparent differences between the two change maps occur on the left edge of the area, where the photograph coverage becomes marginal and errors occur in both the GCP-constrained and co-aligned comparisons. This provides evidence that co-alignment can be used for change detection with a level of detection comparable to that of a survey grade GCP-constrained pair of models.

For the Rügen data, we assessed the comparative accuracy of the resulting model pairs based on the measured change in stable areas of the cliff. Areas of poor fit can be distinguished from areas of real change by the spatial pattern of the differences, the sharpness of the boundary, and visual inspection of the before and after photographs (Figs. 1, 4–6).

Using the standard workflow, the point clouds from successive surveys each contain distinct errors and distortions. Because the error in each cloud is independent of the other cloud, the point clouds are distorted relative to each other and typically cannot be co-registered well, resulting in large errors in the change detection. The error varies throughout the model area, depending on the distortion of the individual models. The spatial pattern of error will also depend on the method used to co-register the two point clouds. For the example shown in Fig. 4, erroneous changes of up to 5 m are measured on the edges of the models and of up to 2.5 m in the center. Throughout the model area, up to 1–2 m of change are erroneously detected in many stable areas, indicating that real changes of this magnitude would be below the level of detection. For the Rügen study area, this level of detection would preclude the use of UAV surveys to monitor small cliff failures.

When the cameras from multiple surveys are co-aligned, the resulting point clouds still contain distortions, but if the procedure is successful, they have been fit to a common geometry and the distortions are consistent between the models. As a result, these errors do not influence comparisons between the models, comparative accuracy is much higher and robust change detection can be performed. We find that the measured change in stable areas is substantially less than in the control case, and therefore smaller amounts of real change can be detected (Fig. 4). For the examples shown here, the level of detection has been reduced from several meters to as low as 15–20 cm. Small cliff failures, bands of more diffusive erosion at the base of the cliff, and even the growth of individual bushes can be reliably detected (Figs. 4–6).

The increase in comparative accuracy is due to the generation of tie points between photographs from different surveys. These tie points, if they are well distributed, enforce a common geometry between the different surveys. We can evaluate the number of common tie points between surveys by comparing the number of points in each sparse cloud following chunk duplication and photo removal (Fig. 2) to the number of points in the sparse cloud generated during the combined alignment. Tie points generated using only photos from survey 1 will be removed when the photos from survey 1 are removed, while tie points generated using photos from both surveys will remain. If common tie points were generated, the two separated sparse clouds have more total points than the original, with the difference being the number of common points (Table 1). Note that this is distinct from the number of matches, as each tie point may be used in multiple matches.

Even when relatively few common tie points are generated, or when they are irregularly distributed, a successful alignment can be achieved. For example, Fig. 5 shows a section of the Rügen study area that is heavily vegetated, with only isolated patches of bare cliffs. While no common tie points can be generated in the vegetated areas, as long as there are common tie points distributed throughout the cliff sections, a relatively good comparative accuracy can be achieved, as illustrated for April 2018–May 2018 (Fig. 5b). However, if there are sections of the cliff where no matches can be made, then large comparative errors can result, as is shown in Fig. 5c for the survey pair October 2017–April 2018. This survey pair had both a low number (1355) and percentage (0.4 %) of common tie points compared to the April 2018–May 2018 pair, which had 3402, or 1 % common tie points. More importantly, there were no common tie points generated in a ∼350 m long stretch at one end of the model, leading to up to 1.5 m of comparative error in this section of the cliff. This illustrates that if common tie points are not distributed through the full extent of the model, edges of the models may not align well. The Daan River example further demonstrates that the distribution of tie points is more important than their number, as good alignment was achieved throughout most of the model despite the generation of only 900 common tie points (0.3 % of the total).

We show that for environments such as coastal cliffs, where the use of ground control points is not possible or not feasible, UAV-based change detection can still be performed with a high degree of confidence if there is sufficient stable area between successive surveys. The workflow we present is quite simple and involves performing image matching and bundle adjustment simultaneously on photographs from pairs or sets of different surveys. This technique may be particularly useful for monitoring processes such as rockfalls, which typically involve steep settings that are difficult to access and exhibit discrete regions of change set within large stable areas.