Introduction

The preservation and management of submerged cultural resources (SCRs), such as shipwrecks, is a difficult endeavor. Monitoring and site degradation studies necessarily must be conducted over a series of years. Common methodologies for monitoring sites usually involve a series of physical in situ surveys at regular intervals but these techniques can be costly, labor intensive and difficult to achieve when a site is difficult to reach either due to depth or location. These constraints may limit an organization’s likelihood of performing multiple surveys. In order to overcome many of the logistical and economic constraints imposed upon these types of surveys, a more rapid, quantitative method is greatly needed.

Structure from motion (SfM) offers a cost-effective alternative to extracting spatial information and/or creating 3D models of SCRs from diver acquired photographs or video stills (Mertes et al. 2014; McCarthy and Benjamin 2014). While there are a variety of commercial and open source software packages (e.g. Bundler, Python Photogrammetry Toolbox, 3DFlow), AgiSoft PhotoScan is the most widely used due to its intuitive Graphical User Interface. The process of extracting geometric information from imagery using the SfM method uses feature-matching algorithms to detect features in multiple images and generate a sparse point cloud. By doing so, it is possible to calculate not only the relative exterior orientation of the image planes (i.e. x, y, z, yaw, pitch, roll), but also the interior orientation of the camera itself (e.g. focal length, distortion coefficients). Having calculated the positions of the camera, using multiview stereo software, locations of more remaining overlapping pixels can be determined and a dense point cloud is created (Agarwal et al. 2011). By adding ground control points (GCPs) with X, Y, Z values at known points on the model, it is possible to scale the resulting model to a known reference system and create accurate three-dimensional digital elevation models with little residual error (e.g. Mertes et al. 2014).

Differencing multiple digital elevation models (DEM) for the purpose of detecting changes in shape and quantifying their magnitudes, is a method that is commonly used in the geosciences (Williams 2012). Two DEMs of the same surface are generated and aligned such that when elevation values for each aligned pixel from the latter DEM are subtracted from the earlier DEM, areas of deposition and erosion will result in positive and negative changes to elevation, respectively. DEM differencing is widely used to investigate changes in glacier volume (Bolch et al. 2011; Etzelmüller 2000; Kääb et al. 2012; Nuth et al. 2010; Rippin et al. 2003; Thompson et al. 2016), as well as river and coastal erosion rates (De Rose and Basher 2011; Lane et al. 2003; Williams et al. 2011; Gibbs et al. 2015; Obu et al. 2016) and even detection of volcanic deformation (Walder et al. 2004).

In this study, we investigate the potential of SfM and DEM differencing as a quantitative monitoring tool for mapping patterns of degradation in shipwrecks and other SCRs. Using SfM, we created two DEMs of a late nineteenth century shipwreck using still frames extracted from separate diver acquired video surveys. By differencing these DEMs, we were able to rapidly map and detect changes, such as large scale collapsing both inward and outward, the movement of objects or shifting of surfaces and, even assess changes in local flora and fauna.

Study Site

The S.S. Wisconsin

The remains of the iron package steamer S.S. Wisconsin rests in 38 m of water, 10.2 km southeast of Kenosha Wisconsin (Fig. 1). Built in 1881 at the Detroit Dry Dock Company in Wyandotte, Michigan, Wisconsin was originally part of the break bulk cargo trade, ferrying goods across Lake Michigan for the Goodrich Transportation Company. During its 48 years of service, Wisconsin served Lake Michigan’s cross lake freight trade under many different names and owners, before returning to the name Wisconsin. Wisconsin was lost 29 October 1929 due to an unknown leak during a violent storm while in route from Chicago to Milwaukee. Today Wisconsin lies upright and mostly intact on the lakebed. The lower portion of the wreck exhibits excellent preservation, while its upper decks have long since collapsed. The contents of Wisconsin’s final cargo remain intact within the vessel’s hull. Although the site retains extensive archaeological integrity, diver reports have indicated that the site has degraded rapidly in recent years.

Fig. 1
figure 1

Overview map showing location of SS Wisconsin shipwreck approximately 9 km SE of Kenosha Wisconsin

Full size image

Wisconsin is a rare example of a vessel type that was vital to Great Lake’s economy (Fig. 2). Year-round steamers like Wisconsin were an important link for railroad freight traffic connecting Wisconsin’s communities economically with wider regional and national markets. As one of the few iron package steamers in Wisconsin waters, the site is an archaeological rarity, and still has the potential to yield further archaeological information. Wisconsin was listed on the National Register of Historic Places on 7 October 2009 (Thomsen and Meverden 2010).

Fig. 2
figure 2

Archive photograph of the S.S. Wisconsin during operation (reproduced with permission from the Wisconsin Marine Historical Society)

Full size image

Methodology

Surveys

Divers conducted photographic and hand surveys of Wisconsin in July 2006, 2007 and again in July 2015. In 2007, divers completed a full Phase II archaeological survey of Wisconsin, recording detailed measurements of the ship for its National Register of Historic Places nomination. During the 2015 survey, fiberglass measuring tapes were used to determine the vessel’s overall length and gain detailed measurements of various sections of Wisconsin’s hull. Detailed annotated drawings were produced to indicate location of data collection and to note qualitative observations of change from the 2006 photomosaic of the site. The 2006 photomosaic was printed on sheets of Mylar and attached to plastic slates. Data and notes were recorded on these Mylar sheets with mechanical pencils. All measurements were recorded to a precision of 0.03 m in the X, Y plane to minimize recording errors, and all are assumed to have an accuracy of 0.03 m. These survey measurements were used for triangulating GCPs used for proper scaling of the 2015 3D model.

The 2006 video survey of Wisconsin was conducted to collect mosaic imagery in July of that year. Divers used a 2003 Sony Handycam video camera to collect all video. The camera was mounted in a down facing orientation on the front of a Silent Submersion UV-26 diver propulsion vehicle (DPV) and piloted approximately 20 feet above the shipwreck (Fig. 3). The video was shot using ambient lighting. Depth control of the UV-26 was accomplished using an OMS Bottom Timer as a depth gauge. A bubble level was mounted to the top of the DVP to ensure a down looking position. Overlapping passes of the wreck site and adjacent debris fields were made to collect a full data set. The survey took approximately 35 min with 8 passes. Still images were captured from the video every 10 s with no automation and cataloged by pass. The video was transformed into a collection of still images with the highest resolution at the time, of 72 dpi and dimensions of 640 × 480 pixels.

Fig. 3
figure 3

Diver operating the silent submersion diver propulsion vehicle

Full size image

The 2015 video survey of Wisconsin was conducted to collect comprehensive imagery of the site for 3D visualization purposes. Divers used a Canon 5D Mark II camera with an “Aquatica 9.25” mega glass dome port housing, to collect all video. The lens used was a Canon 16–35 mm II F2.8. The camera was mounted in a down facing orientation on the front of a Dive X DPV and piloted approximately 20 feet above the shipwreck. The video was shot using an array of four 50 W Light Monkey HIDs with reflectors, mounted to the DPV with depth compensating arms. The arms were adjusted at depth for appropriate light coverage. Depth control of the DPV was accomplished using a Hammerhead Rev. D dive computer as a depth gauge. A bubble level was mounted to the top of the DVP to assure a down looking position. Overlapping passes of the wreck site and adjacent debris fields were made to collect a full data set. The survey took approximately 1 h and 30 min with 12 passes.

Still images were extracted from each survey dataset at a frequency of every 10–15 frames using a custom MATLAB code that automates the process (Mertes et al. 2014). Stills from the 2015 survey were down sampled to roughly the same resolution as the 2006 survey. Downsampling was performed to reduce errors due to resolution differences as well as speed up processing time. Still image extraction resulted in 1754 images from the 2006 survey and 3580 images from the 2015 survey that were then used for reconstruction of the 3D shipwreck model.

Model Generation Using Agisoft Photoscan

Model reconstruction was performed using Agisoft Photoscan SfM software package version 1.1.6 build 2038–1.2.4 build 2399 (64 bit). Images were imported into Photoscan and processed following a similar approach as in Mertes et al. (2014). The processing was performed on a custom-built laptop running an Intel© Core™ i7-6700 4.00 GHz with eight cores and 64 Gb of memory. The 2006 model took roughly 2 h of processing time whereas the 2015 model took 16 h due to the larger image size and quality.

The initial alignment produced sparse point clouds of 548897 points and 452904 points for 2006 and 2015 respectively. These point clouds were next filtered using the Gradual Filter function and points with reprojection errors of greater than ~0.5 were removed along with points having reconstruction uncertainties of greater than 25. This step is recommended in the Agisoft manual to help improve the accuracy of resulting optimization steps and bundle adjustments by removing points with larger potential error (Agisoft 2014).

Secondly, a visual inspection of the point cloud was made to remove leftover points that appear to be erroneous and not conforming to the known shipwreck shape (e.g. far away points) which often result from low angle photographs. The final sparse point clouds of each survey after filtering had 406142 points and 156021 points for 2006 and 2015 respectively.

In order to apply proper scaling to the models, ground control points (GCPs) were placed on the 2006 model with the origin located at the tip of the bow increasing in distance towards the stern with negative values to starboard and positive to the port side. We used this method of relative positioning due to the fact that the original survey was not conducted with the intent of generating a 3D model and no ground truthing was performed. Future investigations of SCR sites would benefit from the implementation of permanent marker points around the site that could be accurately measured and visible in each video surveys. Stable, non-movable GCPs would allow rapid model scaling without the need for extensive measurements and triangulation of GCPs.

In Photoscan, markers were placed at their identified locations in a few photographs and assigned their corresponding triangulated X, Y values (Fig. 4). Agisoft recommends 10 GCPs as a minimum when georeferencing models (Agisoft 2014), yet for this study, as we are only interested in relative change between the two scenes, we used less than this. In addition, our GCPs are positioned on objects that we assumed have undergone little to no change, based on visual inspection of the model orthomosaic. Since the sparse alignment has previously been performed, Photoscan is then automatically able to determine in which other photos each specific GCPs occurs. Visual inspection of these automatically assigned markers must be done to assure correct placement. After every GCP had been located and adjusted in their corresponding images, we optimized the model a second time, thereby improving overall accuracy through a bundle adjustment.

Fig. 4
figure 4

Locations of GCPs and cameras for the 2006 and 2015 models, shown over photoscan shaded meshes (Color figure online)

Full size image

Generating dense point clouds took 18 min for the 2006 model and 4 h for the 2015 model. The resulting dense point clouds contained 7881099 points and 6472617 points for the 2006 and 2015 models, respectively. Outliers were again removed through visual inspection and then surface meshes were generated. The reconstruction parameters for the surface meshes were of type ‘Arbitrary’, using the dense point cloud, with interpolation disabled. Quality was set to ‘High’ and depth filtering (i.e. noise removal) was set to ‘Moderate’. This resulted in surface meshes of 19999999 faces and 11539095 vertices for the 2006 model and 17452030 faces and 9607632 vertices for the 2015 model. From these meshes we next generated DEMs with a resolution of 1 cm pixel−1 (Fig. 4).

Orthomosaics were created at a resolution of 0.005 cm pixel−1 using the settings ‘average’ for the color calculation of pixels based on multiple images. These products (i.e. DEM and Orthomosaic) were exported as geotiff files for further analysis using ESRI ArcMap GIS software. The 2006 model was completed first using four GCPs taken from known measurements. The GCPs for the 2015 model were then selected from the 2006 DEM and orthomosaic thereby allowing us to more accurately align the two models.

Results

Models

Creation of models from both surveys was successful for the majority of the upper deck from bow to stern. Considering that the purpose of the 2006 survey was to create a down-looking photomosaic, minimal coverage of the hull limited our ability to generate a complete 3D model. The 2015 survey included collection of video coverage oblique to the hull surface as well as inside the ship’s cargo hold, yet initial attempts to build a full model including interior spaces as well as the ship’s hull failed due mainly to poor lighting and resolution. In some areas, portions of the hull were successfully extracted, but not with sufficient spatial density to allow comparison between time periods. For this reason, we chose to focus on generating 3D models from a vertical perspective looking at changes visible on the upper deck as well as surrounding the shipwreck.

The resulting models have 3D RMSE, based on the GCPs, of 0.27 m and 0.30 m for the 2006 and 2015 models respectively. In the Z direction, the RMSE for 2006 and 2015 is 0.02 and 0.21 m respectively. Each model has photographic coverage of >9 image overlap over the majority of the area of interest (Fig. 5).

Fig. 5
figure 5

Locations of images (dots) and total imagery overlap color map, produced from Photoscan, showing coverage of >9 images over most of the shipwreck site for both 2006 and 2015

Full size image

DEM Differencing

By subtracting the height values of the 2006 model from the 2015 model, we created a difference map showing areas that have undergone some type of vertical change over the ~10-year period. We have excluded any changes greater than 5 m and less than −5 m as visual inspection revealed these regions were located in areas of poor model reconstruction. Assuming minimal changes to the surrounding lakebed area (e.g. low disturbance, slow deposition) we assess the differences in this area to quantify our model error. The middle 50% of our data range from 0.03 to −0.54 m, with ~99% of the data between 0.89 and −1.4 m. Normal and extreme outliers make up only 0.38% of our dataset (n = 3.7 × 106) (Fig. 6). We calculate an off ship standard error (σ) of ±0.36 m and an overall mean difference of −0.22 m. Obvious erroneous regions, identified as interpolated artifacts from the mesh generation, have also been clipped out to allow better visualization of changes within our desired range. These areas include a small section on the starboard side of the bow and a section near the stern starboard side (Fig. 7).

Fig. 6
figure 6

Box and whisker plot of off ship DEM differencing values. The gray dash line within the box represents the median while the upper and lower bounds of the box encloses the central 50% of data. The numbers to the left of the box are the upper adjacent the 75th percentile, median, 25th percentile and the lower adjacent. On the right are the percentages of the total data points (n = ~3.7 × 106)

Full size image
Fig. 7
figure 7

Difference map of the 2006 and 2015 models with specific areas of interest identified within labeled boxes. Bordering over interpolated areas that have been removed are indicated by dotted black outlines. Positive values of change indicate areas of decreased elevation whereas negative values indicate an increase in elevation from 2006 to 2015 (Color figure online)

Full size image

Our results show three large areas of high changes (i.e. approximately ±2–5 m) occurring mainly in the central region of the ship (Fig. 7a, d, e). Smaller areas with changes on the order of ~1 m appear in both the bow and stern regions indicating shifting of objects on the more stable deck (Fig. 7b). In the bow region, differences associated with planks extending from starboard to port, which were extracted from one set of imagery, yet not the other, clearly show up as changes of ~1 m (Fig. 7, Bow). Overall, however, the bow appears to have undergone little change as indicated by the minimal change of the capstan (Fig. 7). The remainder of the difference map appears to be within our calculated error.

Region A shows an increase in height (negative change) of 3–5 m off the starboard side, whereas along the starboard side from region A through E there is a decrease (positive change) of 0.2 to 4–5 m. Similarly, on the port side there is also a thin region (Fig. 7d) indicating a large increase in height bordered by an area of height decrease. These zones (A, D, E) are displaying large-scale collapse features and shifting of major sections of the hull. The central areas within regions A, D, and E have a pattern of increasing change from port to starboard indicating that the deck collapse outward has caused material to shift downward and slide from port to starboard.

Visible Changes

In addition to producing DEMs from underwater surveys, SfM can also produce high-resolution orthomosaics from the imagery and model, which can be used as a complementary source of data to inspect changes to SCRs. The large detected negative change along the aft section of the starboard hull can be attributed to the outward collapse of a 19.8 m section of the hull (Fig. 8, 1). This outward collapse of the hull appears to have also caused a destabilization to the deck as other material and objects can be seen to have shifted significantly between the two periods (Fig. 8, 2).

Fig. 8
figure 8

Subset A from Fig. 7 showing the outward collapse of the mid-aft section of the starboard hull (1). As well as observed rotation of a large piece of material (2)

Full size image

Small scale changes can also be visibly detected as shown in area B of Fig. 7. We can visually detect thickening changes on objects such as the steering quadrant and rudder post, as well as on the deck planks and structure around them (Fig. 9). Smaller magnitude changes detected in the aft section between the smokestack and the stern, which are just above our estimated error, can visibly be attributed to lateral shifting of large planar objects. A piece of the upper cabin works, roughly 10.6 m in length has shifted towards the port, and begun flexing downward over the aft cargo hold opening (Fig. 10, 1). In addition, a second large piece of cabin structure has also tipped inward and rotated slightly in an anti-clockwise fashion after it split into multiple pieces (Fig. 10, 3). Another object (Fig. 10, 4) can be observed to have fallen into a void created by the forward shifting of the smokestack (Fig. 10, 5).

Fig. 9
figure 9

Subset B from Fig. 7 showing the observed buildup of mussels on what is believed to be components of the steering quadrant (1) and the rudder post (2)

Full size image
Fig. 10
figure 10

Subset C of Fig. 7 showing shifting of a section of the upper cabin works from starboard to port (1) as well as rotation of another piece (2) caused by what appears to be sliding into the aft cargo hold. Objects 3, 4 and the smokestack (5) have all shifted forward. Object 4 has tipped upward and slid into the void between the deck and smokestack

Full size image

On the port side of the ship, where we detected larger changes, we can observe more outward collapse of the ship. A piece of the deck rail has peeled away from the upper hull and folded downward (Fig. 11, 1). Also, changes can be observed to have occurred to deck planks and other pieces of wood, that could possibly be attributed to the major starboard hull collapsing (Fig. 11, 2, 3). The collapse of the aft section of the starboard hull outward (Fig. 8, 1) and the inward collapse of the forward section (Fig. 12, 1) appears to have caused a decrease in underlying support as well as an increase in pressure from above, respectively. These changes appear to have caused shifting of material on the deck such as that seen in Fig. 11 as well as Fig. 12 (2, 3), downward and to the starboard, effectively increasing the slope of the deck in the port direction.

Fig. 11
figure 11

Subset D of Fig. 7 showing the outward collapse of the main railing on the port side (1) as well as degradation and movement of deck material (2, 3) which may have been caused by the outward and inward collapse of the starboard hull (Fig. 7a, e)

Full size image
Fig. 12
figure 12

Figure 7 subset E showing the two sections of the starboard hull which split and collapsed outward (black arrows, i.e. Figure A) and inward (1) onto the deck material. Changes in the orientation of deck material (e.g. planks) are observed (2) with the appearance of two metal radiator looking objects (3)

Full size image

Discussion

Our results demonstrate that using SfM and photo/video surveys of SCR sites can yield high-resolution models and orthomosaics, which can aid in management, monitoring, and preservation of SCRs. We have shown that the application of DEM differencing to a periodically obtained model of SCR sites is a powerful tool for detecting and quantifying spatial changes. Visual inspection of resulting orthomosaics can allow SCR managers to see changes on a wide variety of scales.

Model Generation

One of the largest difficulties encountered in constructing our models comes from image resolution and lighting. The 2006 video survey was shot at a relatively poor resolution compared to resolution available today. In order to make DEMs of comparable quality, and to speed up processing of large batches of images during DEM construction, we down sampled our 2015 imagery to match the resolution of the 2006 imagery. In many areas of the model, the relatively low image resolution does not seem to cause issues with model creation. In areas of low exposure, or in slightly blurred images, lower detail results in failure to match features across multiple images. The inability to construct a complete hull stems mainly from the fact that the camera angle during filming resulted in inclusion of the lake surface in the upper corner of each frame. The lake surface appears bright from sunlight, in contrast to the hull bottom, which was in shadow. The exposure of the camera adjusted for the brightness, rendering the hull bottom too dark to match features across images. This is also similar to imagery of the upper deck that includes large openings to the inner ship, where little light penetrates. These areas appear as depressions in the model. Keeping the camera with its upper edge at most parallel to the lakebed would likely have minimized this problem and produced more usable hull imagery.

Shooting video allows a large number of images to be collected quickly, but the slow shutter speeds and dim lighting characteristic of video result in many blurred images. Higher-quality imagery could be obtained with a still camera and underwater strobes, though using a still camera would result in an increase in overall survey time. For management applications where in-water survey time is not a critical deciding factor, the improvement in matched points, and ultimately model resolution, could make still photography preferable to video.

Despite image blur in some scenes, the two models constructed for this study display great resolution both geometrically as well as visually. The use of a minimal amount of GCPs for scaling doesn’t appear to have hindered our DEM differencing but it may be a contributing factor to the estimated error of ±0.36 m. The use of SfM in other areas of archaeology has demonstrated that with an organized network of accurately measured GCPs, it is possible to create DEMs from SfM with errors in the range of a few centimeters (Green et al. 2014). Another key factor contributing to error in our models is the limited distribution of GCPs across the scene, which leaves some regions prone to distortions due to lack of GCP constraints. In spite of our high degree of error, our models are sufficiently reliable to detect changes of >~1 m (i.e. % error of less than 36%), which is acceptable for a wide range of SCR management applications. Improving the method used for acquiring GCPs could significantly lessen this error.

Detected Changes

The results of the comparative analysis of the two models indicate that Wisconsin, similar to many metal shipwrecks in the Great Lakes, is collapsing. While many sites remain well-preserved after years of submersion, the majority of metal wrecks have begun to collapse inward, as opposed to outward (Zant et al. 2015; Thomsen et al. 2014; Meverden et al. 2016). Patterns of inward collapse are associated with ship construction techniques that relied on stanchions along the centerline of vessels that were made of narrow steel or iron I-beams, usually measuring 0.3 m2. Iron and steel ships were lighter, stronger and longer than the wooden vessels they replaced, which allowed them to carry higher tonnages of cargo. The I-beams used in their construction, however, have been a notorious weak point in metal shipwrecks located in Lake Michigan.

Initial archaeological investigations of Wisconsin in 2006, seemed to indicate that Wisconsin was following a similar trend of collapsing inward. The upper sections of Wisconsin’s outer hull had begun to lean inward as the central deck beams collapsed (Thomsen and Meverden 2010). Data reports from investigations in 2015 indicate that this pattern, however, is no longer evident. A 19.8 m section of Wisconsin’s starboard hull, which had been leaning inward significantly in 2006, has fallen away from the main body of the wreck, hinged at the aft end (Figs. 7a, 8). A 7.0 m long section of the hull, located just forward of this large section, has continued its collapse inward, creating a vertical crack along its forward edge, at the aft edge of the forecastle deck, and causing the collapse of previously intact forecastle deck beams (Figs. 7e, 12). Additionally, a section of the vessel’s main railing, which had remained intact in 2006, now hangs off the port side of the hull (Figs. 7d, 11).

Although there is no direct indication as to why this is occurring at this site and not others, one possible explanation is that site colonization by quagga mussels has altered the normal pattern of progressive inward collapse. During the sinking event, Wisconsin’s upper deck cabins blew off due to a sudden release of air trapped within its enclosed spaces. When this occurred, the integrity of the iron deck beams was compromised, and upon the vessel’s collision with the lakebed, these beams and their support stanchions, collapsed inward along the ship’s centerline (Thomsen and Meverden 2010). This trajectory of inward collapse should have continued as increasing weight from the buckled upper decks was placed on the remaining deck stanchions in the lower hold of the vessel. Numerous influencing factors may have contributed to the now outward collapse situation. These factors may include the colonization of mussels which not only add weight to the structure, but also metal corrosion caused by increases of local levels of carbonic acid through the introduction of carbon dioxide through mussel respiration (Hecky et al. 2004). Mussels also increase the overall surface roughness and, by doing so, increase drag forces caused by water currents. It is also important to consider that during late 2011, divers reported strong rip currents in the area extending to depths of 30–40 m. The combination of these factors could be contributing to the cracking of the thin iron covering the vessel’s upper deck above the bustle. The bustle refers to the 0.9 m sections on either side of the ship where it was widened in 1909. Historically, this was an area noted for leaking and weaknesses in the hull, and may be further indication of why the vessel is degrading more rapidly here than in other sections. Except for the small section of the vessel’s fantail stern, the bow and the stern of the vessel have experienced very little overall change between 2006 and 2015. This is likely because these two sections retain their integrity from the original build of the vessel in 1881.

Although DEM differencing indicates these are areas of little change, the remaining intact sections of the upper decks, in the bow and the stern, show some signs of degradation as well (Fig. 7). The most notable changes are located near the remains of Wisconsin’s smokestack. The 2015 data notes a significant shift in the angle of the smokestack (which is still attached to the boilers below decks) from a slight list to starboard to a significant list forward (Fig. 10, 5). This change is indicative of the collapse of the deck beams holding the smokestack in place. Though most of the upper deck beams were believed to be damaged during the original sinking, the visible continuation of this collapse progressively aft reveals a continued rapid degradation of the site. Similar deterioration has been noted in the vessel’s bow, with the forecastle deck beams beginning to collapse progressively forward.

During the 2015 archaeological investigations, divers qualitatively noted that the bulkheads and stanchions in the lowest deck of the cargo hold are significantly buckled as well, indicating further collapse since the 2006 survey. Despite this collapse below decks, most of the measured change in the vessel between 2006 and 2015 has occurred to the vessel’s upper decks, or what would have existed above the waterline when the vessel was afloat. This is likely occurring because these early iron and steel vessels were made with two different thicknesses of metal and using different sizes of plating. Built to break through thick Lake Michigan ice to operate in year-round service, Wisconsin’s lower hull plates were significantly stronger than those used on the upper decks. Below the bustle, the hull plating consists of a series of iron plates measuring significantly smaller in size than the hull plates used for the upper decks. These plates were also 0.03–0.06 m thicker than the upper deck hull plates as well (Zant et al. 2015; Thomsen and Meverden 2010).

In summary, the use of SfM modelling with repeat underwater video or still photographic surveys of SCRs can greatly aid in the detection of site degradation and environmental and anthropogenic disturbances. By constructing models of the shipwreck Wisconsin using video surveys performed in 2006 and 2015, we have calculated spatial changes, through the method of DEM differencing, of up to ~±5 m vertically. Visual inspection of these changes reveals that the ship has shifted from a state of inward collapse in 2006, to what now appears to be outward collapse as seen by the many large hulls sections falling away from the ship. Using the high-resolution orthoimages and DEMs from both time periods we have also been able to pinpoint numerous other objects on the upper deck that have undergone changes which can be associated with shifting of materials, degradation and mussel colonization.

Conclusion

The results of our study provide an example of a novel new approach to detecting and quantifying changes and degradation to SCRs through the use of diver-acquired photographic surveys and SfM model generation. The work demonstrates that, even with minimal GCPs from around a site, models can be easily co-registered to increase the overall accuracy of the model differencing. We have shown that the combination of both the DEM and high-resolution orthomosaic allow detailed inspection of changes in a spatial and visual manner. While previous repeat surveys of SCR rely on extensive and time-consuming observations and measurements, DEM differencing can more efficiently produce products to be used for comparison. The spatial resolution of detected changes in our survey is limited by the lack of stable, unmovable GCPs that were measured at high precision. By first establishing fixed GCPs around the SCR, model generation would be easier and less time consuming and accuracy would be improved relative to our study reported here. Future investigations using this type of method would benefit from using still photographs with powerful strobes to decrease image blur and increase image detail and quality. The knowledge gained from repeated surveys of SCR is of vital importance for monitoring and maintaining our historic sites.