Assessing 3D metric data of Digital Surface Models for extracting archaeological data from archive stereo-aerial 1 photographs 2 Heather Papworth (corresponding author)a, Andrew Fordb, Kate Welhama, and David Thackrayc 3 4 aDepartment of Archaeology, Anthropology & Forensic Science 5 bDepartment of Life & Environmental Sciences 6 cICOMOS UK 7 Bournemouth University 8 Faculty of Science and Technology 9 Talbot Campus, Fern Barrow 10 Poole, Dorset. BH12 5BB 11 Email: h.e.papworth@gmail.com 12 13 Abstract 14 Archaeological remains are under increasing threat of attrition from natural processes and the continued 15 mechanisation of anthropogenic activities. This research analyses the ability of digital photogrammetry software to 16 reconstruct extant, damaged, and destroyed archaeological earthworks from archive stereo-aerial photographs. Case 17 studies of Flower’s Barrow and Eggardon hillforts, both situated in Dorset, UK, are examined using a range of imagery 18 dating from the 1940s to 2010. Specialist photogrammetric software SocetGXP® is used to extract digital surface 19 models, and the results compared with airborne and terrestrial laser scanning data to assess their accuracy. Global 20 summary statistics and spatial autocorrelation techniques are used to examine error scales and distributions. 21 Extracted earthwork profiles are compared to both current and historical surveys of each study site. The results 22 demonstrate that metric information relating to earthwork form can be successfully obtained from archival 23 photography. In some instances, these data out-perform airborne laser scanning in the provision of digital surface 24 models with minimal error. The role of archival photography in regaining metric data from upstanding archaeology and 25 the consequent place for this approach to impact heritage management strategies is demonstrated. 26 27 Keywords 28 Digital photogrammetry; archaeology; archive stereo-photographs; earthworks; reconstruction; digital surface models; 29 laser scanning. 30 1. Introduction 31 32 mailto:h.e.papworth@gmail.com Archaeological sites are subject to substantive on going decay and damage caused by a variety of both natural and 33 human behaviours (Rowley and Wood 2008). Factors such as increased storm rates and sea-level rise, and the sharp 34 growth in the efficiency, rate and scale at which many anthropogenic activities occur within the UK landscape are a 35 pressing issue for many regions (Oxford Archaeology 2002; Murphy et al. 2009). Subsequently it has been estimated 36 that, of the 600,000 sites in England, one has been lost per day since 1945 (Darvill and Fulton 1998). This equates to 37 a projected disappearance rate of at least 25,000 sites over the past 70 years, and nowhere is this more apparent 38 than in the tangible loss and damage of earthwork features. 39 In an attempt to mitigate this loss a number of conservation charters exist that advocate recording archaeological sites 40 before they are destroyed (Bassegoda-Nonell et al. 1964; ICOMOS General Assembly 1996), but the reality of 41 achieving this ideal is an immense challenge. Within the UK alone the high density and sheer variety of monuments 42 mean that there are inevitably large numbers of sites where damage or destruction has occurred, and the level of 43 recording undertaken has been minimal, or in some extreme cases non-existent. The ability to utilise archive data 44 from the periods subsequent to the disappearance of such sites to reconstruct what previously existed would be of 45 significant benefit, not only for the understanding of these monuments but, just as importantly, to inform the 46 conservation, management and interpretation of our heritage assets in the future. 47 Fortunately, an archive containing these data does exist. In the UK, stereo-aerial photographs (SAPs) have been 48 gathered regularly, and on a large scale, since the 1940s. These images hold within them the potential for landscape 49 reconstruction. Three-dimensional (3D) data can be extracted using digital photogrammetry methods to create digital 50 surface models (DSMs); an approach that has already been successfully used in the geomorphology and surveying 51 disciplines to assess terrain change over time (Chandler 1989, Adams and Chandler 2002, Walstra et al. 2004, 52 Walstra 2006, Miller et al. 2008, Aguilar et al. 2013). In contrast, although archaeologists have been utilising aerial 53 photography for over a century and are experts in prospecting for and mapping earthwork features from them in two-54 dimensions (Wilson 2000, p.16; Barber 2011, p.215), very few have acknowledged the inherent 3D properties of SAPs 55 (Verhoeven et al. 2012) beyond the use of stereoscopes. 56 This research employs qualitative and quantitative methods to assess the ability of archive SAPs to reconstruct extant, 57 damaged and destroyed earthworks. Two case study sites are used to compare the results obtained from digital 58 surface models created from a range of SAPs to those achieved via modern metric survey techniques commonly used 59 in the archaeology and heritage sectors: global navigation satellite systems (GNSS), and terrestrial and airborne laser 60 scanners. The utility of these data when compared to both objective (metric) and interpretative (such as hachure 61 plans) survey is discussed (Bowden and McOmish 2012; Blake 2014). 62 2. Data and Methods 63 64 2.1 Field Site Selection 65 66 The field study sites of Flower’s Barrow and Eggardon Iron Age hillforts (Figures 1 and 2) were selected as they both 67 contain a mixture of subtle and pronounced earthworks, some of which are well preserved and stable whilst others 68 have been damaged or destroyed via natural or anthropogenic agents. 69 70 Figure 1: Location of Flower’s Barrow hillfort within the United Kingdom including an orthophotograph of the site and examples of earthworks within the hillfort (Bottom Left) Occupation Platforms and (Bottom Right) a linear annex. 71 Coverage of each area with archive SAPs devoid of cloud cover and with suitable stereo-overlap is available, with a 72 range of imagery for each decade from the 1940s to the present day. Flower’s Barrow is situated on Defence Estates 73 land and the site is only accessible to the public on weekends and during major school holidays, thus footfall is limited. 74 A condition assessment completed by Wessex Archaeology (2001) identified the hillfort and its environs as being in 75 good condition, although the southern ramparts have been lost to cliff erosion since construction. The terrestrial 76 hinterland is, however, stable. Eggardon is unique in that the northern half of the hillfort interior has been damaged by 77 Figure 2: Location of Eggardon within the United Kingdom including an orthophotograph of the site and examples of earthworks within the hillfort including (Bottom Left) linear features and pits, (Bottom Upper Right) damaged barrow and (Bottom Lower Right) henge monument. irregular ploughing since the 1940s, whilst the southern half is separated by a fence denoting the parish boundary and 78 has remained in the custody of the National Trust, facilitating its preservation. 79 2.1.1 Baseline Data Collection 80 A baseline reference metric survey was collected at each site using a Leica C10 terrestrial laser scanner (TLS) whose 81 station locations were ascertained using a Leica Viva GNSS. A sub-10cm point cloud was achieved within the hillfort 82 at each site, with the Mean Absolute Error for the dataset at Flowers Barrow calculated by Leica Cyclone as 0.011m 83 and at Eggardon Hillfort as 0.014m. The TLS point density as created at each field site is shown in Figure 3. 84 90 To identify systematic errors in the TLS dataset prior to undertaking analysis with it, a number of random points were 91 collected across each field site using a Leica Viva GNSS. The residual differences between TLS and GNSS elevation 92 values are illustrated in Figure 4. 93 94 Figure 3: Diagrams illustrating TLS point density created at Flowers Barrow (above) and Eggardon Hillfort (below). 20cm, with very few values exceeding this figure. The graphs in Figure 5 illustrate the lack of correlation between 100 residual elevation values between the TLS and GNSS in relation to the proximity of measurements to the scanner and 101 as a function of TLS point density. Subsequently it can be said that neither the proximity of the TLS data to the 102 scanner or the point density of the data influences error in the TLS data. 103 104 2.2. Archive Stereo-Aerial Photographs and Airborne Laser Scanning 105 106 Archive SAPs for both sites were obtained from the National Monuments Record (NMR) in Swindon, Bournemouth 107 University (BU), and Dorset County Council (DCC) (Table 1). 108 Figure 5: Scatter plots demonstrating the lack of a relationship between residual elevation values when examining these as a function of proximity to the location of the C10 TLS (above) and as a function of TLS point density (below). 109 The prints from BU and DCC were scanned using an A3 desktop scanner whilst the NMR scanned the requisite 110 negatives using a Vexcel Photogrammetric Scanner. Each image was scanned at a resolution of 2400 dots-per-inch 111 (dpi) and saved in the lossless TIFF file format. A recent set of commercially available, digital SAPs were obtained 112 from GetMapping Ltd, created using a Vexcel UltraCamX digital aerial camera and delivered in JPEG format from the 113 RGB (not panchromatic) sensors. 114 Archive airborne laser scanning (ALS) data from the EA was obtained for Flower’s Barrow in its raw format to ensure 115 that processing the data could be undertaken in a transparent way as the methods employed by the EA are not 116 disclosed. A parallel dataset was not available for Eggardon. 117 118 Flower’s Barrow Archive Stereo-Aerial Photographs Date Flown Archive/ Creator Scale Focal Length (mm) Flying Height (m) Ground Sample Distance (m) Image Type (Verticals) Format (cm) *Original Media March 1945 NMR/RAF 1:10500 203.2 2133.6 0.110 B&W 12.7x12.7 Negative August 1968 NMR/RAF 1:10000 152.4 1524 0.118 B&W 23x23 Negative June 1972 BU/J.A.Storey 1:12000 151.85 1822.2 0.132 B&W 23x23 Print April 1982 NMR/OS 1:8000 304.8 2438.4 0.081 B&W 23x23 Negative June 1986 DCC/ C.E.G.B.Winfrith 1:12000 153.05 1836.6 0.135 B&W 23x23 Print 1997 DCC/ C.E.G.B.Winfrith 1:10000 153.15 1531.5 0.114 Colour 23x23 Print Sept. 2009 GetMapping Ltd. - 100.5 - 0.150 Colour 10.39x6.78 Digital Eggardon hillfort Archive Stereo-Aerial Photographs Date Flown Archive/ Creator Scale Focal Length (mm) Flying Height (m) Ground Sample Distance (m) Image Type (Verticals) Format (cm) *Original Media January 1948 NMR/RAF 1:10000 508 5029.2 0.104 B&W 20.95x19.05 Negative March 1948 NMR/RAF 1:10000 508 5029.2 0.114 B&W 20.95x19.05 Negative April 1969 NMR/OS 1:7500 304.8 2286 0.072 B&W 23x23 Negative October 1972 DCC/J.A.Story 1:12000 151.85 1822.2 0.122 B&W 23x23 Print April 1984 NMR/OS 1:8000 304.8 2438.4 0.078 B&W 23x23 Negative July 1989 NMR/OS 1:8200 304.8 2499.4 0.084 B&W 23x23 Negative Sept. 1997 DCC/ NRSC Airphoto Group 1:10000 153.15 1531.5 0.104 Colour 23x23 Print 2010 GetMapping Ltd. - 100.5 - 0.150 Colour 10.39x6.78 Digital *Prior to purchase/digitisation, each image was examined as a photographic print to ensure suitable coverage and sufficient image quality Table 1: List of archive stereo-aerial photographs and their associated metadata for the field study sites Flower’s Barrow and Eggardon hillforts. 2.3. Photogrammetric Processing 119 120 High-end photogrammetric software was chosen for processing archive SAPs, despite the popularity of Structure-121 from-motion (SfM) software with the archaeological community (Hullo et al. 2009; Ducke et al. 2011; Verhoeven 2011; 122 Plets et al. 2012; Verhoeven et al. 2012a, 2012b; Koutsoudis et al. 2013; De Reu et al. 2013; Green et al. 2014; 123 McCarthy 2014). Although SfM was initially considered as a means of processing archive SAPs, it was disregarded for 124 its lack of optimisation for use with traditional, high resolution vertical stereo-aerial photographs. SfM is designed for 125 use with lower-resolution photographs that are taken with a suggested overlap of +80% forward and 60% side 126 (AgiSoft LLC 2014), not the 60% forward and 20-30% overlap present in SAPs. Whilst it is possible to obtain a DSM 127 from SfM using archive SAPs, any metric information extracted from them should be examined carefully. 128 DSMs were created from SAPs using the SocetGXP® Automatic Terrain Extraction algorithm (ATE), the settings for 129 which were determined via experimentation as described in Papworth (2014, p.153). SocetGXP® leads the user 130 through a step-by-step workflow, prompting the input of interior and exterior orientation parameters. As with other such 131 software packages, the best results are obtained if camera calibration data, namely fiducial coordinates, principle point 132 location and lens distortion parameters are provided (Figure 6). 133 134 However, many of these measures are not available for archive SAPs but, to account for these issues, the software 135 includes a self-calibrating bundle adjustment routine. With the exception of the 2009 and 2010 SAPs from GetMapping 136 Figure 6: Diagram illustrating the components of interior orientation (top) and the information strip often provided with an aerial photograph (below) (Papworth 2014, p59). Ltd, which were provided with camera calibration and exterior orientation information, the remaining SAP datasets 137 were processed using the self-calibrating bundle adjustment to obtain the missing camera parameters. To mitigate for 138 the lack of interior orientation data and exterior orientation information (i.e. GNSS camera positions at the time of 139 exposure and the associated rotation measures describing the attitude of the camera at this time) a large number of 140 ground control points (GCPs) were collected (Figure 7). The GCPs were gathered as close to the point of interest 141 (each hillfort) as possible. The objects used as GCPs were features identifiable in the archive SAPs such as gate 142 Figure 7: Location map showing the distribution of GNSS ground control points, shown in red, for (a.) Flower’s Barrow (GCPs from alternative mapping sources shown in green) and (b.) Eggardon hillfort (Papworth 2014). posts, road intersections, and the corner of structures such as buildings for example. Flower’s Barrow proved 143 challenging because of its location on a live firing range and the restricted access to the area surrounding the site. A 144 large number of GCPs were recorded using a Leica Viva GNSS that, when operating in Network RTK mode (i.e. 145 receives real-time positional corrections from static reference stations within the UK via a mobile phone), has a stated 146 accuracy of 8mm horizontally (+0.5 parts per million, or ppm) and 15mm vertically (+0.5ppm) (Leica Geosystems AG, 147 no date). However, many of the GCPs situated to the north of the hillfort within the fields and the farm area were 148 extracted from 1:1000 scale Ordnance Survey mapping and a 3rd party 2m DSM. The fields and roads surrounding 149 Eggardon were fully accessible, facilitating the collection of GCPs across the area of interest. Due to the problems 150 encountered with the lack of mobile phone reception in the area, Network RTK was not available and thus GCPs were 151 gathered using a Leica GS10 reference station in combination with the GS15 rover. This data was subsequently post-152 processed using Leica GeoOffice software, which resulted in a mean accuracy per GCP of 0.014m. 153 The workflow developed for this research utilised a small number of initial tie points to ensure an acceptable solution 154 was achieved for relative orientation between the SAPs when the bundle adjustment was first run. The overall root 155 mean square error (RMSE) was used to assess the orientation result, as it represents a global measure of the 156 accuracy with which the software has calculated the solution for the relationship between the SAPs and the GCPs. A 157 minimum RMSE value was sought by variously tightening and loosening the exterior orientation accuracy values, and 158 removing tie points with the largest errors. Subsequently, more tie points were added along with a small number of 159 GCPs. These were required to provide locational information in an appropriate coordinate system and strengthen the 160 relationship between the images. As each GCP recorded with the Viva GNSS is stored with data quality information, it 161 was possible to input these values into SocetGXP®, adding an extra 20cm to the x/y values, as suggested by Walstra 162 et al (2011), to account for potential offsets caused by GNSS errors and the observer identifying GCP locations within 163 the imagery. 164 The process was continued until no further decrease could be obtained in the root mean squared (RMS) value. The 165 accuracies achieved during triangulation of the imagery are shown in Table 2. A DSM of 1m resolution was extracted 166 from the data using the ATE adaptive algorithm and exported from SocetGXP® as a point cloud for interpolation in 167 ArcMap 10.1. As these data were to be validated using independent DSM datasets, described in Section 2.4.1, 168 standardising the interpolation algorithm used was necessary to limit the variables capable of influencing data quality 169 and subsequently its analysis. The ‘natural neighbour’ interpolator was employed because it has been identified as an 170 accurate method to apply to high resolution datasets (Abramov and McEwen 2004, Bater and Coops 2009) and is 171 stated by Maune et al. (2007) to work well with both regular and irregularly-spaced point cloud data and is not prone to 172 introducing artifacts. 173 174 175 2.4. Validation Methods 176 177 2.4.1. Quantitative Assessment 178 179 Objective assessment of error in the SAP DSMs was undertaken on their elevation values in comparison with those of 180 the TLS collected at each field site (see Section 2.1). This was achieved by subtracting each of the SAP DSMs from 181 the TLS DSM to create a DSM of Difference (DoD) for each SAP epoch that contained the residual values between 182 terrain models. These values were extracted from each DoD to create a table of residual values that are taken from 183 each cell of the raster, which in turn were used to create summary statistics as described in Section 2.4.2. The 184 desktop scanned SAPs for Flower’s Barrow were not assessed as the pilot study utilised only the NMR imagery to 185 determine the viability of the research. 186 Error assessment was enhanced by converting the elevation DSMs from both the SAPs and the TLS into first-order 187 derivatives, namely ‘slope’ and ‘aspect’, as per the approach advocated by Gallant and Wilson (2000). Whilst useful in 188 their own right as terrain attributes, the conversion of elevation data into first-order derivatives can enhance noise or 189 other errors contained in the original dataset, helping to identify problematic regions within a dataset. Therefore each 190 of the SAP slope DSMs were subtracted from the TLS slope model to create a slope DoD, and the same process was 191 repeated for the aspect datasets. The residual values from each of the slope and aspect DoDs were exported to 192 SPSS for statistical analysis (Section 2.4.2). 193 Flowers Barrow Triangulation Root Mean Square Residuals Date Flown Image (Pixels) X (m) Y (m) Z (m) Total RMS (m) Mar 1945 8.333 3.592 2.440 4.480 5.915 Aug 1968 1.004 2.776 2.120 6.494 7.374 Apr 1982 1.518 1.588 1.783 2.204 3.142 Sept 2009 1.136 8.326 9.699 3.157 1.317 Eggardon hillfort Triangulation Root Mean Square Residuals Date Flown Image (Pixels) X (m) Y (m) Z (m) Total RMS (m) Jan/Mar 1948 3.198 2.807 1.244 5.220 5.791 Apr 1969 1.142 3.797 8.092 4.085 8.948 Oct 1972 6.564 2.115 1.744 1.138 2.969 Apr 1984 6.243 5.900 3.164 1.815 1.934 Jul 1989 6.385 3.758 5.879 2.670 7.471 Sept 1997 2.466 7.829 1.452 5.256 1.731 May 2010 6.675 7.325 7.668 2.428 1.088 Table 2: Results of the image triangulation process in SocetGXP® 2.4.2. Summary Statistics 194 195 A number of statistical measures were used to assess whether systematic or random errors were present in the SAP 196 DSMs. Systematic errors are caused by a bias within the photogrammetric workflow, such as errors in pixel geometry 197 of a sensor (camera or scanner), or lens distortion for example (Mitchell 2007). These errors can be mitigated if they 198 have been measured, modelled and a correction for them applied (Wolf and Dewitt 2000). Random errors are 199 sometimes referred to as noise and are related to data quality, although tend to be difficult to predict. Within the SAPs, 200 these errors are caused by poor image geometry and image blur, for example, although other sources of random 201 error, such as image resolution and scale can, to some extent, be predicted. In TLS data, random errors are caused 202 by adverse influences on the laser beam, such as particulate matter in the air (i.e. water droplets), or strong winds 203 destabilising the instrument during data collection for example. Systematic errors tend to be a consistent value (i.e. 204 photographs taken with the same lens will all contain the same amount of lens distortion) and affect the accuracy of 205 the data by shifting calculated values by a known quantity away from the ‘true’ value (the latter of which can never 206 truly be determined). However, as systematic errors are consistent, they do not influence precision, which is related to 207 the repeatability of measurements. Precision is influenced by random errors that do not have consistent values and 208 are often difficult to account for. 209 Global indicators of DSM error, namely root mean square error (RMSE), mean error (ME) and standard deviation 210 (SD), were calculated to objectively compare the TLS DSM to each SAP DSM (Baily et al. 2003, Walstra et al. 2004, 211 Papasaika et al. 2008, Aguilar et al. 2009, Walstra et al. 2011, Perez Alvarez et al. 2013). RMSE was used as an 212 indicator of accuracy and to identify the presence of systematic errors within the SAP DSMs (Dowman and Muller 213 2011), therefore the larger the RMSE error, the poorer the accuracy of a dataset in comparison with the baseline TLS. 214 SD was used to indicate the precision of the SAP DSMs as this value is influenced by the presence of random errors 215 in a dataset. ME values reveal bias (Fisher and Tate 2006), which can be introduced by systematic errors that cause 216 under- or over-estimation of elevation values. With the exception of the RMSE, which was calculated in Microsoft 217 Excel, both the ME and SD from each DoD were created using SPSS. 218 Frequency histograms of the residual values were generated from the data to determine whether the residual 219 distribution between each SAP DSM and that of the TLS is normal. Such distributions are generally indicative of 220 random errors, with increasing histogram width illustrating a decrease in precision. A skewed distribution, which 221 contains the majority of residual values either in the left or right-hand section of the graph, may indicate the presence 222 of systematic bias in a dataset. In this particular instance, camera calibration information detailing the lens distortion 223 parameters, fiducial marks coordinates and the principle point were not available for any of the SAP datasets, with the 224 exception of the 2009 and 2010 images for Flowers Barrow and Eggardon Hillfort respectively, and thus it was 225 anticipated that the results would highlight the presence of systematic errors. 226 The peak of the distribution also provides an important indicator of difference. If all systematic errors were eliminated 227 via the use of an appropriate photogrammetric model, only the remaining random errors would influence this peak. 228 Subsequently, the size of these errors will be represented by the Kurtosis (spikiness) of the histogram. 229 Scatter plots were also created to identify whether a linear relationship exists between the SAP DSMs and the 230 independent dataset, namely the TLS. This process plots the elevation values of one dataset against another and can 231 reveal how similar or different these data are. A positive, linear relationship, or correlation, between the two datasets 232 would indicate their similarity. 233 2.4.3 Spatial Autocorrelation of Errors 234 235 In addition to the absolute (global) indicators of error calculated above, local Moran’s I analysis was undertaken to 236 determine how errors were spatially distributed across a DSM. This was important as, although the DoDs for 237 elevation, slope and aspect illustrate spatial distribution of residual values, they do not indicate whether these 238 differences are statistically significant. The results generate a diagram indicating where, across the area of interest, 239 clusters of statistically significant high (shown in red) or low (blue) residual values occur as well as regions with non-240 significant values (grey), an example is provided in Figure 8. Statistically significant values that are surrounded by 241 either much lower (light blue) or higher (orange) values are indicative of outliers. Together, both the clusters and 242 outliers are suggestive of values that exceed what would be expected from random error (ESRI 2014). 243 244 2.4.4. Qualitative Assessment 245 246 In addition to the metric performance of each dataset, the SAP DSMs were assessed against the interpretive surveys 247 produced by the Royal Commission for Historical Monuments of England (RCHME) for each site. Profile data were 248 extracted from each SAP DSM epoch and their form compared to that of the profile lines recorded by the RCHME on 249 *HH=Statistically Significant High Value, HL=High Value Surrounded by Low Values, LH=Low Value Surrounded by High Values, LL= Statistically Significant Low Value. The significance level is set at 95%. Figure 8: Example of a Moran’s I diagram illustrating the statistical significance of residual slope values between the 1984 SAP data and that of the TLS across Eggardon. their hachure plans from Eggardon (RCHME 1952) and Flower’s Barrow (RCHME 1970). Point data were generated 250 along the location of the profile line as shown on each hachure plan by importing the georeferenced RCHME data into 251 ESRI ArcMap v10. These were then extracted to provide locational information for the GNSS, and the stakeout 252 function used to locate the profile line in the field and re-measure it. This approach facilitated the direct comparison of 253 profile data as captured during a conventional archaeological survey (RCHME) with that created by mass-capture and 254 GNSS methods to assess how representative the latter were of archaeological earthworks. 255 256 3. Results 257 258 3.1 Quantitative Assessment of DSM Quality 259 260 3.1.1 Summary Statistics 261 262 The summary statistics from both Flower’s Barrow and Eggardon, based on the comparison of SAP DSM elevation, 263 slope and aspect values with those of the TLS data, are presented in Table 3. The general trend in the results 264 supports the observation that DSM quality increases as SAP age decreases. This is illustrated by the decrease in ME, 265 SD and RMSE values, although there are two important exceptions to this: the DSMs derived from the desktop 266 scanned prints within the Eggardon dataset and the 2009 digital photography obtained for Flower’s Barrow. The 267 differences between these datasets and the others are more easily discernible by examining the slope and aspect 268 results, which exhibit greater disparities due to the conversion of elevation values into first-order derivatives. 269 The residual histograms (Figures 9 and 10), display normal distributions for both field sites when comparing the slope 270 and aspect differences between the SAP and TLS values. 271 272 Flower’s Barrow TLS MINUS 1945 TLS MINUS 1968 TLS MINUS 1982 TLS MINUS 2009 TLS MINUS 2009 ALS Number of values 47361 51759 51759 51759 51759 Elevation Mean Error (m) 0.035 -0.163 -0.126 -0.097 -0.399 Std. Deviation (m) 1.204 0.741 0.213 0.619 0.394 RMSE (m) 1.205 0.759 0.248 0.626 0.561 Slope Mean Error (degrees) -2.210 1.630 0.213 0.584 -0.582 Std. Deviation (degrees) 8.914 8.250 3.532 7.830 6.059 RMSE (degrees) 9.184 8.410 3.538 7.852 6.087 Aspect Mean Error (degrees) -5.278 3.949 -0.093 -0.354 -1.920 Std. Deviation (degrees) 47.767 44.873 20.453 31.628 28.032 RMSE (degrees) 48.057 45.046 20.453 31.629 28.097 Eggardon hillfort TLS Minus 1948 TLS Minus 1969 TLS Minus 1972 (DTS)* TLS Minus 1984 TLS Minus 1989 TLS Minus 1997 (DTS)* TLS Minus 2010 Number of values 89021 89021 89021 89021 89021 89021 89021 Elevation Mean Error (m) 0.172 -0.044 -0.145 -0.386 0.076 0.171 0.282 Std. Deviation (m) 2.572 0.390 0.664 0.248 0.150 0.505 0.155 RMSE (m) 2.577 0.392 0.6796 0.459 0.168 0.533 0.321 Slope Mean Error (degrees) -14.410 -1.171 -3.540 -0.666 0.191 -3.166 0.481 Std. Deviation (degrees) 12.253 4.249 6.385 3.300 3.021 5.959 2.278 RMSE (degrees) 18.915 4.408 7.301 3.366 3.027 6.748 2.328 Aspect Mean Error (degrees) -13.663 -4.581 -7.232 -1.202 -0.681 -7.487 1.569 Std. Deviation (degrees) 68.852 52.521 61.173 48.557 44.069 62.097 39.191 RMSE (degrees) 70.195 52.721 61.599 48.572 44.074 62.546 39.223 *DTS = Desktop Scanned Prints Table 3: Summary statistics for Flower’s Barrow (top) and Eggardon hillfort (below) showing the global errors between the TLS DSM and the SAP DSMs and their derivatives 273 The majority of the histograms are leptokurtic i.e. demonstrates a high peak. Greater variation in histogram shape was 274 observed across the SAP DSM elevation values for both sites (Figures 9 and 10). The Flower’s Barrow elevation 275 residual histograms were positively skewed for the 1945 and 1968 SAPs whilst negatively skewed for the 2009 SAPs 276 and the 2009 ALS data, with the latter exhibiting a more leptokurtic peak. The 1982 SAP residual histogram was, 277 however, normally distributed and displayed a leptokurtic peak. Bimodal distributions were observed in the Eggardon 278 elevation residuals for the 1948, 1972 and 2010 SAP DSMs, indicating the presence of two modes (i.e. there are two 279 residual values that most commonly occur) within the data. The second, much smaller peak of residuals for the 1948 280 and 1972 datasets are comprised of negative values, whilst the second peak within the 2010 data has a very small 281 spread but a leptokurtic peak that occurs around 0m. The remaining epochs, namely the 1969, 1984, 1989 and 1997 282 SAP DSM residual histograms, are all normally distributed with leptokurtic peaks. 283 Figure 9: Residual Histograms of DSM difference for Flower’s Barrow. The normal distribution is represented by the bell-shaped line. Scatter plot results for discerning the relationship between the SAP DSM elevations and the TLS DSM (Figure 11 and 284 12), generally increase in positive linearity as SAP age decreases for both sites, although there are a number of 285 exceptions highlighted in the Eggardon dataset (Figure 12). 286 Figure 10: Residual Histograms of DSM difference for Eggardon. The normal distribution is represented by the bell-shaped line. 287 The 1972 and 1997 Eggardon scatter plots (Figure 12), whilst broadly linear, contain a large amount of noise, as 288 demonstrated by the spread of the values within the graph, which subsequently thickens the appearance of the linear 289 scatter. However, this does not greatly affect their correlation values, as shown in Table 4, which contains Pearson’s 290 ‘r’ values, where any value from 0.5 to 1.0 indicates a high, positive correlation (Field 2013, p.82). 291 Figure 11: Flower’s Barrow Elevation Scatter Plots Figure 12: Eggardon Elevation Scatter Plots 292 The 1948 elevation results from Eggardon exhibit the least significant relationship with an ‘r’ value of 0.717, although 293 this differs markedly to the results obtained for the 1945 SAPs at Flower’s Barrow, which obtained a value of 0.993. It 294 is also evident that, for the data at both field sites, the correlation between the TLS and SAP data decreases when 295 converted to a first-order derivative (Table 4). These statistical data also indicate that correlation improves as the age 296 of the SAPs decrease. 297 298 3.1.2 Spatial Autocorrelation of DSM Errors 299 300 The residual values from the elevation, slope and aspect Moran’s I maps all demonstrate that the residual distribution 301 is clustered for both sites, as illustrated in Figure 13 and 14. 302 Flower’s Barrow Correlations N* Elevation Correlation Slope Correlation Aspect Correlation TLS vs. 1945 47361 0.993 0.710 0.562 TLS vs. 1968 51759 0.997 0.700 0.595 TLS vs. 1982 51759 1.000 0.949 0.922 TLS vs. 2009 SAPs 51759 0.998 0.750 0.814 TLS vs. 2009 ALS 51759 0.999 0.856 0.856 Eggardon hillfort Correlations N* Elevation Correlation Slope Correlation Aspect Correlation TLS vs. 1948 89021 0.717 0.300 0.144 TLS vs. 1969 89021 0.981 0.725 0.510 TLS vs. 1972 89021 0.949 0.419 0.323 TLS vs. 1984 89021 0.992 0.832 0.572 TLS vs. 1989 89021 0.997 0.851 0.656 TLS vs. 1997 89021 0.967 0.423 0.308 TLS vs. 2010 89021 0.997 0.917 0.729 *N = number of elevation/slope/aspect values compared Table 4: Pearson’s ‘r’ correlation values obtained by comparing the TLS elevation and first-order derivative data to that of the SAPs at both Flower’s Barrow and Eggardon hillfort. 303 Figure 13: Flower’s Barrow Local Moran’s I Results 304 Figure 14: Eggardon Local Moran’s I Results An examination of the Flowers Barrow elevation datasets in Figure 13 indicates consistent clustering of values on the 305 rampart slopes for each epoch, including the ALS, with the exception of 1945. Within this diagram there appears to be 306 some distinction between clustered low values to the south-west and clustered high values to the north-east. This is 307 suggestive of a systematic error in the photogrammetric model created from the 1945 SAPs which, at first glance, may 308 indicate a tilt within the SAP DSM. However, if this were the case, it would be expected that a distinctive cluster of 309 high values would be followed by an area of non-significant values where the pivot point of the tilt would be, which 310 would then turn in to a region of low values. The pattern shown within the 1945 elevation Moran’s I diagram may show 311 large blocks of high and low values, but each is interspersed with clusters of values of the opposite sign. 312 Subsequently, it is unlikely that tilting of the SAP DSM is the cause of this difference. This same pattern is not 313 apparent in the slope and aspect derivatives from the 1945 DSM although, as derivative values of elevation, they are 314 independent of the factors affecting a tilt in elevation models. 315 A pattern of high and low value clustering along the rampart slopes and in the ditches of Flowers Barrow can be 316 identified in all of the Moran’s I diagrams, with the exception of the 1945 elevation map. Based upon the quality 317 assessment undertaken on the TLS data in Section 2.1.1 these errors are not caused by low point densities or 318 accuracies in this dataset. Whilst the largest errors between the TLS and GNSS datasets were located along the inner 319 most north-facing rampart slope, they do not occur sufficiently frequently elsewhere within the hillfort to suspect that 320 deficiencies in the TLS data influence the pattern of spatial clustering within the Moran’s I results. The pattern of 321 clustering suggests that, as slope angle increases, so to do the differences between the TLS and SAP datasets. 322 Within the Eggardon Hillfort results, as shown in Figure 14, there are different patterns of clustering. Most notable 323 within the elevation diagrams are the stripes of clustered values that are especially evident in the 1972 and 1997 DTS 324 datasets. These stripes do not necessarily coincide with features in the imagery, such as the information strips, the 325 edge of a frame or an artefact in the images themselves, although they have structure and are suggestive of a 326 systematic error. As this pattern is noted in the two datasets scanned from prints using a desktop scanner it is likely 327 that the source of error may be related to the use of print materials and an uncharacterised scanner, the issues of 328 which are discussed in Section 4.3. 329 A further systematic error appears to be present in the 2010 elevation diagram as shown by the clustering of 330 significant high values to the west and significant low values to the east. The transition between these significant 331 values occurs on the boundary between coverage of 6 stereo pairs (west) and 4 stereo pairs (east) in the image block. 332 It is possible that this difference in stereo coverage created differences in number and accuracy of tie points in the 333 block adjustment and also a difference in the number of stereo-matched points in the cloud used to interpolate the 334 DSM. Subsequently, this would result in different significant residuals across the DSM when compared with reference 335 data. Were the block to be extended to create a consistent stereo coverage across the site, such that 6 stereo pairs 336 covered the east of the hillfort, it is likely that these significant differences in residuals would be of the same sign, 337 minimised, or absent. 338 As with the Flowers Barrow datasets, there are noticeable clusters of significant values along the rampart slopes of 339 the hillfort in the majority of the elevation and slope diagrams for all SAP epochs. These are most pronounced along 340 the slope of the east-facing rampart, situated to the east of the hillfort. As with the Flowers Barrow results, an 341 examination of the TLS data density and accuracy in Section 2.1.1 does not depict any obvious deficit in the data 342 which may influence the outcome of the Moran’s I analysis. There is one exception to this observation and that is the 343 line of high values running from the north-west to the south-east of the hillfort, which is synonymous with the fence 344 that denotes the boundaries of ownership. This feature was not removed from the TLS dataset and thus indicates 345 elevation values that are expected to be much higher than those contained within the SAP DSMs. Overall, these 346 results support the same conclusion as Flowers Barrow, in that an increase in the slope angle of the ramparts 347 decreases the accuracy of its representation in comparison with the TLS data. 348 349 3.2 Qualitative Assessment of DSM Data 350 351 The profiles illustrated in Figure 15 show that the results for Eggardon appear to create a more reliable reconstruction 352 of the ramparts than those for Flower’s Barrow. There is close conformity with the RCHME survey of Flower’s Barrow 353 from the GNSS with the exception of the plateau in the middle of the profile. Aside from the 1945 and 1968 DSMs the 354 profiles from the remaining SAP DSMs are remarkably similar. The same can be said of the Eggardon profile whereby 355 SAP data from 1969 onwards conforms to the RCHME survey. In corroboration with the quantitative results, the DSMs 356 from newer SAPs create profiles that are more consistent with the shape of the rampart banks and ditches than those 357 from the older photography and the images scanned from prints using a desktop scanner. Subsequently, it can be 358 Figure 15: Profile Comparisons between the RCHME survey, GNSS and SAP DSM results. said that SAP DSMs can be used to provide archaeological survey data, although photography from the 1940s and 359 from desktop scanned photographic prints cannot. 360 361 4. Discussion 362 363 The results, both empirical and qualitative, demonstrate that archive SAPs can be used to generate data akin to that 364 of the RCHME surveys whilst highlighting a number of issues that should be considered when undertaking DSM 365 generation with these data. 366 367 4.1 SAP Age 368 369 It has been demonstrated that the quality of DSM output does generally improve with decreasing imagery age: a trend 370 which has also been identified by previous studies conducted by Walstra et al. (2004) and Aguilar et al. (2012). Both 371 authors provide tables showing a decrease in RMSE values as the age of SAPs also decrease. However the 372 relationship is complex and influenced by a variety of factors such as camera and film quality, suitable GCPs and the 373 state of preservation of photographic materials, for example. 374 The weakest performing DSMs were produced from the 1945 and 1948 photographs from Flower’s Barrow and 375 Eggardon respectively (see Section 3.1.1). The 1945 SAP DSM from Flower’s Barrow is missing the western tip of 376 the hillfort due to a lack of stereo-overlap in this region. However, when comparing the appearance of the 1945 and 377 1948 DSMs with one another (see Figure 13 and 14) the 1945 data has a less noisy appearance and the lack of 378 stereo-overlap does not appear to have hindered reconstruction of the hillfort to the centre and east of the earthwork. 379 This disparity between two SAP datasets of similar age illustrates the variation in image quality depending on location 380 for the period 1945-’51. During this period the RAF were responsible for the acquisition of aerial photography over the 381 United Kingdom on behalf of the OS. The cameras and aircraft used were not designed for photogrammetric use 382 (Owen and Pilbeam 1992). For instance, the most widely employed cameras (F24 & F52) were non-metric, relatively 383 small-format (5 x 5” and 8.5” x 7” respectively) and instead designed for aerial reconnaissance (Conyers Nesbit 2003). 384 A minority of sorties were flown at relatively low altitude with comparatively short focal lengths (2133.6m and 8” over 385 Flowers Barrow in 1945), but the majority, especially in rural areas, were flown at greater altitudes with longer focal 386 lengths (5029.2m and 20” over Eggardon in 1948). The resulting photo scales were very similar (1:10500 and 1:10000 387 respectively) but with markedly different baseline-to-altitude ratios, with the latter making photogrammetry very 388 difficult. These are the most likely explanations for the differences in photogrammetric results between 1945 and 1948 389 presented here. 390 These problems were further exacerbated by the use of “split-vertical” camera pairs. In the case of the Spitfire PR XIX, 391 as used over Eggardon in 1948, this arrangement had two F52 cameras angled 5° 20” from the vertical (Evidence in 392 Camera, March 1945). This resulted in low-oblique photographs, with twice the coverage of a single camera. This was 393 ideal for aerial reconnaissance but came at the expense of photogrammetry. With few photogrammetric plotters 394 available (Whitaker 2014) the emphasis was instead on the creation of photo-mosaics for reconstruction purposes. 395 The RAF were also unable to guarantee a supply of aircrew experienced in aerial survey, given problems of 396 recruitment and retention (Macdonald 1992). 397 From 1951 the OS resumed control of aerial photography acquisition, but aircraft and aircrew were provided by the 398 Ministry of Transport & Civil Aviation, who proved unsatisfactory for similar reasons as the RAF. For the majority of the 399 1950s and early 1960s aerial survey was limited in quantity and extent (Macdonald 1992) and thus archival 400 photography was not available for either Flowers Barrow or Eggardon. 401 It was only after the purchase of 9 x 9” format metric cameras (i.e. Zeiss RMK 30/23 & Wild RC8R) and the use of 402 civilian contractors for flying that matters finally improved from the mid-1960s (Survey Season, in FLIGHT 403 International, 4th January 1965. p. 49 & 50). Stricter quality-control requirements stipulating the timing and weather 404 conditions (including sun angle) were also used by the OS and contractors. Therefore DSMs from SAPs dating from 405 the mid-1960s onwards for both sites do not demonstrate as stark a difference as the data from the 1940s. 406 Further developments saw the RAF re-equipped with the Canberra PR 3, 7 & 9 aircraft and, whilst still employing split-407 vertical pairs of F52, they also carried the F49 9 x 9” format survey camera (Oakey 2008). These resulted in superior 408 photogrammetric products, as demonstrated with the SAP acquired over Flowers Barrow in 1968. 409 Based on the reasons stated above the authors recommend that archive SAPs of the UK dating from the 1945-‘65 410 period require independent consideration and verification, especially those acquired from 1945-’51. 411 In the 1980s the OS purchased cameras with forward motion compensation (i.e. Zeiss RMK 15/23 & Wild 412 RC10A/RC20), which further improved photogrammetric products (Macdonald 1992). However, it is arguable that the 413 introduction of colour film at the OS and commercial companies, in part to meet the growing requirement for colour 414 orthophotographs, may have negatively impacted photogrammetry in terms of the lower granularity with greater film 415 density (the opposite of black & white films) (Langford 1998). 416 417 4.2 Image Resolution 418 419 In comparison with the analogue archive SAPs, the most recent 2009 digital photography generated over Flower’s 420 Barrow by GetMapping Ltd. did not perform as favourably as the ALS and 1982 SAP datasets. This is illustrated by 421 the summary (global) statistics provided in Table 3. The elevation data from the 2010 digital photography of Eggardon 422 was also out-performed by 1989 OS SAPs on comparison of the RMSE values, which are 0.321m and 0.168m 423 respectively. The cause of this is postulated to be the reduced detail captured by the digital multispectral camera 424 system of the Vexcel UltraCam X from which the images were provided (unfortunately, the panchromatic images were 425 not available in the scope of this project). Unlike the 6µm pixel size of its panchromatic camera, the multispectral 426 imagery delivers a pixel size of 18µm (Microsoft Corp. 2008), providing an image with pixel dimensions of 5770x3770 427 pixels and a ground-sample distance (GSD) of 0.15m. Therefore, the newest SAP datasets for both field sites have 428 the lowest GSD of the group and contains a lesser amount of information per pixel than the film-based imagery of 429 earlier sorties. 430 It should be noted that image resolution for scanned photographs is dependent on the abilities of the scanner used to 431 digitise them. Whilst much of the issues surrounding scanners and the effect of photographic materials on resolution 432 will be discussed in Section 4.3, a brief appraisal of scan resolutions is given here. Walstra et al. (2011) state that a 433 scanned pixel size of 6 to 12µm would be required to preserve a resolution of 30 to 60 lines/mm as provided by 434 original film, based on a paper by Baltsavias (1999) who states that good DSM results can be acquired when using a 435 scan resolution of 25-30µm. The photographs for this research were scanned at a resolution of 2400 dots-per-inch or 436 10.6µm whilst those utilised by Walstra et al. (2011) and Aguilar et al. (2013) range between 15 - 42 µm. However, the 437 increased resolution of the scanned images used in this research will not necessarily increase the precision of the 438 resultant DSMs if the original film quality, atmospheric conditions at exposure and processing methods, for example, 439 are such that image quality is already degraded prior to scanning. Aguilar et al. (ibid.) compared the overall RMSE 440 results obtained from 1977 imagery, , with values obtained by Walstra et al. (2007) for 1971 (1.31m). 441 A further consideration regarding the GetMapping imagery is the provision of such in the compressed jpeg file format, 442 as was the case for this research, and whether this may have had a negative impact on DSM quality. Lam et al. 443 (2001) have shown that utilising jpeg files with a compression ratio of less than 10 should have no significant impact 444 upon DSM quality, unless the image is texturally rich. If so, there will be a more significant degradation of image 445 quality, which in turn affects the quality of the resultant DSM. As no information relating to the compression ratio was 446 provided with the GetMapping data for this research, the adverse influence of jpeg compression on the DSMs created 447 using this imagery cannot be dismissed. 448 449 4.3 Photographic Materials and Scanning Technologies 450 451 The positive influence of the use of well-defined photogrammetric scanning technology upon archive SAPs can be 452 seen by examining the Eggardon dataset. Tables 2 and 3 demonstrate the overall weak performance of the 1972 and 453 1997 SAP DSMs. This disparity is due to the scanning of these images using commercially available desktop 454 equipment that had not been characterised to provide error correction factors for any geometric or radiometric 455 distortions it may introduce to the digitised photography. Geometric errors (i.e. lens distortion, defective pixels and 456 CCD misalignments) and radiometric errors (i.e. illumination instabilities, stripes and electronic noise) are rarely, if 457 ever, accounted for by manufacturers of desktop scanners (Baltsavias and Waegli 1996; El-Ashmawy 2014), with the 458 radiometric errors in particular varying on a frequent basis (Baltsavias and Waegli ibid.), thus illustrating the instability 459 of this equipment. El-Ashmawy (ibid.) states that photogrammetric scanners are manufactured to robust standards to 460 ensure the accuracy of analogue to digital image production at every stage of the scanning process, which also 461 illustrates the inflated cost of these scanners. 462 All other SAPs from this site were scanned from original negatives using a photogrammetric scanner (see Section 463 2.2), and the DSMs produced from this imagery performed well. The 1948 SAPs form the only exception to this result 464 producing data seen to be inferior to both the 1972 and 1997 SAPs. Here age is thought likely to be the dominant 465 factor (see Section 4.1). A comparative study on the influence of scanning technologies upon archive SAPs and the 466 extraction of DSMs using SfM software has been investigated by Sevara (2013). The author also concluded that 467 DSMs obtained using photography scanned from negatives contained less distortion and performed better than scans 468 obtained from print materials. Print materials produce photographs with lower detail due to the larger size of silver 469 crystals used in their emulsions, as compared to that used in film (Walstra et al. 2011), and are known to be more 470 susceptible to degradations (Jacobson 2000, p373). Degradations can take a variety of forms such as image fading 471 and staining (caused by residual chemicals in photographic material or oxidisation of silver particles by atmospheric 472 gases, for example) or microbial attack on the gelatin contained with a negative or print, and on the paper substrate 473 within print materials. High humidity will cause gelatin in photographic materials to swell that, in a photographic print, 474 can also encourage some of the layers to detach, particularly around the edges, which gives rise to this ‘frilling’ as well 475 as blistering on the print surface (Weaver 2008, p13). Weaver (ibid., p14) also highlights further issues with humidity, 476 namely ‘cockling’ of print materials, which can occur differentially across a photograph as the gelatin and paper 477 expand at different rates, and the reduction of paper flexibility that enhances the likelihood of tearing and cracking in 478 this substrate. These factors highlight the challenges in removing geometric distortions from archive photographs, 479 particularly from prints. 480 Overall, the results from this section of our study reflect the findings from Walstra et al. (ibid.) in that the values 481 obtained from scanned prints are worse than those from scanned negatives. Walstra et al. (ibid.) state that the 482 accuracy values from scanned prints may be worse by a factor of 3.1 in comparison with scanned diapositives, 483 although the authors also note that this value is based upon limited data. It should also be remarked that different 484 photogrammetric software, namely Leica Photogrammetry Suite (LPS), was utilised by this study and thus the results 485 from an alternative package, such as SocetGXP, may be different. However, it is nevertheless a useful indication of 486 the quality of data achievable when utilising poorer-quality data. 487 488 4.4 Control and Tie Point Distribution 489 490 The distribution of tie points and GCPs across each set of imagery is suspected to be the cause of bimodal 491 distributions observed in the elevation residual histograms obtained for the 1948, 1972 and 2010 Eggardon datasets. 492 A lack of sufficient control close to the edges of the photographs, where lens distortion is at its peak, may cause the 493 triangulation routine to perform sub-optimally. Triangulation utilises both tie points and GCPs to calculate missing 494 exterior orientation and camera calibration information, including the lens distortion parameters. Subsequently, image 495 matching in these regions may perform sub-optimally, particularly if radial distortion has been modelled poorly, and the 496 triangulation solution may result in values that increase with distance from the principle point. This issue will propagate 497 into the terrain extraction process whereby unrepresentative elevation values may appear in areas where the SAPs 498 overlap, causing artifacts to appear in the DSM. These were observed in the 1972, 1997 and 2010 Eggardon DSMs, 499 manifesting as a stripe effect running north-south through the hillfort (Figure 11). However, the 1997 residual 500 histogram does not display a bimodal distribution, which may suggest that the elevation offsets in the SAP DSM 501 caused by the stripe may be negated by the other elevation values within the data. 502 Studies by Walstra (2006) and Aguilar et al. (2012) have referred to optimal GCP numbers when processing archive 503 SAPs, with Walstra (ibid.) using between 4 to 9 per stereopair and Aguilar et al. (ibid.) identifying similar triangulation 504 accuracies irrespective of whether 12 or 24 GCPs were utilised. Subsequently, the latter research recommended the 505 use of 6 to 9 GCPs per stereopair (14 when using very old photographs of lower geometric and radiometric quality 506 and/or smaller scale) if self-calibrating bundle adjustment was required, as was the requirement for this project due to 507 the deficit of camera calibration data. However, as Walstra et al. (2011) note, selecting suitable control in historical 508 imagery is often limited, relying instead on site accessibility and/or change since the photography was captured and 509 must thus be considered based upon the features visible in any one image set. Subsequently, the number of GCPs 510 identified per stereopair, per epoch for this research were akin to those used by Walstra (2006), namely between 4 to 511 9 GCPs. It should also be noted that a number of natural features (i.e. fence and gate posts, road intersections) were 512 used as GCP locations, which have been identified by Walstra et al. (2011) as a potentially large source of 513 uncertainty, particularly when compared with artificial targets deployed as a control network during contemporary a 514 photogrammetric survey. 515 516 4.5 Camera Calibration Information 517 518 Camera calibration information was only available for the most recent digital photography, namely the 2009 Flower’s 519 Barrow and 2010 Eggardon datasets from GetMapping Ltd. However, this did not appear to influence the quality of the 520 DSM results, particularly for the Flower’s Barrow field site over which the 1982 SAP DSM performed most favourably 521 in comparison with the TLS, as highlighted by the summary statistics provided in Table 3. In addition, the inclusion of 522 camera calibration information for the 2009 and 2010 digital photography did not prevent the clustering of residual 523 values (Table 4). These results suggest that calibration information does not greatly influence the occurrence of 524 systematic errors. Therefore, whilst it is good practice to work with camera calibration information where possible, for 525 archival imagery their absence should not significantly negatively impact on the result obtained. 526 527 4.6 Archaeological Implications 528 529 The recording of archaeological earthworks, particularly those at threat from attrition, should be regularly recorded for 530 management and monitoring purposes (ICOMOS General Assembly 1996). Metric information is crucial for these 531 purposes, playing an important role in the analytical process of understanding a monument and mapping changes that 532 are occurring or may occur in due course. However, as demonstrated by the field sites utilised for this study, the only 533 known survey data of each consists of an interpretive hachure plan and selected profiles, which depict the form of the 534 earthworks but do not facilitate future mapping and monitoring practices. Subsequently, the regaining of metric data 535 for such purposes in relation to these sites is only possible via the use of SAPs. This research has demonstrated that 536 many of the photogrammetrically scanned SAPs and the DSMs created from them have been able to provide metric 537 information akin to that recorded using a TLS. This has been demonstrated by high correlation values between the 538 elevation data produced by photogrammetry and laser scanning. It has also been shown that profiles similar to the 539 RCHME interpretive surveys can be extracted from SAP DSMs, thus illustrating the capability to extract archaeological 540 information from archive SAPs. 541 542 5. Conclusion 543 544 In conclusion, this research has demonstrated that archive SAPs can produce data of archaeological utility. 545 Archaeologists wishing to extract DSMs from these data are advised to consider a number of factors when pursuing 546 their own SAP datasets. Firstly, SAP age influences DSM quality, with photography from the 1940s generating poorer 547 results than DSMs generated using desktop scanned prints. However, the 1945 SAPs processed for Flower’s Barrow 548 did show promise, particularly when reconstructing the rampart slopes as illustrated by its profile. The digitisation 549 method was also a key factor in producing high-quality DSMs and thus obtaining digital images scanned from 550 negatives using a photogrammetric scanner will provide the best results. 551 Modern aerial photography providers should be asked to provide data from the panchromatic sensors in their digital 552 cameras in a lossless (i.e. TIF format) as the RGB imagery processed by this research provided results that were, in 553 the case of the Flower’s Barrow study, poorer than imagery captured in 1982. Whilst camera calibration information 554 did not appear to greatly influence the quality of the DSM results, GCPs should be gathered using GNSS equipment 555 and, if possible, well distributed across the area of interest and throughout the photographic pair, strip or block. 556 The results presented by this research demonstrate that archaeologists must not solely rely on empirical analysis to 557 assess DSM quality, but continue to conduct visual assessments of the data. This approach has been advocated 558 previously by the archaeological community with reference to ALS DSMs (Doneus et al. 2008; Corns and Shaw 2009). 559 The implications the success of this method has upon archaeological research management and the mitigation of 560 earthwork loss is significant. It provides a means by which to recover metric information lost through attrition, 561 particularly if no prior record has been created. The reconstruction of larger earthworks has been demonstrated, and 562 further work is required to assess this method for smaller earthwork features, such as pits and housing platforms. 563 564 6. Acknowledgements 565 566 This research was supported by a doctoral research studentship jointly funded by Bournemouth University (BU) and 567 the National Trust. The authors would like to thank the technical staff at BAE Systems UK (Rick Mort, Steve Foster 568 and Rut Gallmeier) for their help and advice regarding photogrammetric workflows and Professor Stuart Robson and 569 Dietmar Backes (University College London) for training and assistance with SocetGXP. Thanks are also due to 570 Martin Schaefer (University of Portsmouth) for initially providing photogrammetric scans of BU imagery and members 571 of staff at the National Trust (Dr Martin Papworth, Guy Salkeld and Alison Lane) for providing help and access to 572 National Trust resources. Grateful thanks are also extended to the two anonymous reviewers who took the time and 573 trouble to make valuable comments and suggestions regarding the content of this paper. 574 575 7. References 576 577 Abramov, O. and McEwen, A. 2004. An evaluation of interpolation methods for Mars Orbiter Laser Altimeter (MOLA) 578 data. International Journal of Remote Sensing, 25(3), 669-676. 579 Adams, J. and Chandler, J., 2002. Evaluation of Lidar and Medium Scale Photogrammetry for Detecting Soft-Cliff 580 Coastal Change. The Photogrammetric Record, 17 (99), 405-418. 581 AgiSoft LLC, 2014. Agisoft PhotoScan User Manual: Professional Edition, Version 1.1. AgiSoft LLC. 582 Aguilar, M., Aguilar, F. and Negreiros, J., 2009. Self-Calibration Methods for Using Historical Aerial Photographs with 583 Photogrammetric Purposes. In: Ingegraf XXI Lugo. 584 Aguilar, M. A., Aguilar, F. J., Fernández, I. and Mills, J. P., 2013. Accuracy Assessment of Commercial Self-585 Calibrating Bundle Adjustment Routines Applied to Archival Aerial Photography. The Photogrammetric Record, 28 586 (141), 96-114. 587 Baily, B., Collier, P., Farres, P., Inkpen, R. and Pearson, A., 2003. Comparative assessment of analytical and digital 588 photogrammetric methods in the construction of DSMs of geomorphological forms. Earth Surface Processes and 589 Landforms, 28 (3), 307-320. 590 Baltsavias, E.P., Waegli, B., 1996: Quality Analysis and Calibration of DTP Scanners. Paper presented at the 18th 591 ISPRS Congress in Vienna. In 'International Archives of Photogrammetry and Remote Sensing', Vol. XXXI/B1, pp. 13-592 19. 593 Baltsavias EP. 1999. On the performance of photogrammetric scanners. Photogrammetric Week ‘99: 155–173. 594 Barber, M., 2011. A History of Aerial Photography and Archaeology: Mata Hari's glass eye and other stories. Swindon: 595 English Heritage. 596 Bassegoda-Nonell, J., Benavente, L., Boskovic, D., Daifuku, H., Flores, C., Gazzola, P., Langberg, H., Lemaire, R., 597 Matteucci, M., Merlet, J., Pane, R., Pavel, S. C. J., Philippot, P., Pimentel, V., Plenderleith, H., Redig de Campos, D., 598 Sonnier, J., Sorlin, F., Stikas, E., Tripp, G., de Vrieze, P. L., Zachwatovicz, J. and Zbiss, M. S., 1964. The Venice 599 Charter. In ICOMOS (Ed.). Venice. 600 Bater, C. W. and Coops, N. C., 2009. Evaluating error associated with lidar-derived DSM interpolation. Computers and 601 Geosciences, 35 (2), 289-300. 602 Blake, B., 2014. The slippery slope: CAD hachures. billboyheritagesurvey. 13th March. Available from: 603 http://wp.me/pYvo7-1ZR [Accessed 14th March 2014] 604 Bowden, M. and McOmish, D., 2012. A British Tradition? Mapping the Archaeological Landscape. Landscapes, 12 (2), 605 20-40. 606 Chandler, J., 1989. The Acquisition of Spatial Data from Archival Photographs and Their Application to 607 Geomorphology. (PhD). City University, London. 608 Conyers Nesbit, R. 2003. Eyes of the RAF: A History of Photo-reconnaissance. Stroud: Sutton Publishing Ltd. 609 Darvill, T. and Fulton, A., 1998. Monuments at Risk Survey 1995. Bournemouth and London: Bournemouth University 610 and English Heritage. 611 De Reu, J., Plets, G., Verhoeven, G., De Smedt, P., Bats, M., Cherretté, B., De Maeyer, W., Deconynck, J., 612 Herremans, D., Laloo, P., Van Meirvenne, M. and De Clercq, W., 2013. Towards a three-dimensional cost-effective 613 registration of the archaeological heritage. Journal of Archaeological Science, 40 (2), 1108-1121. 614 Dowman, I.J. and Muller, J.P. 2011. Evaluation of Spaceborne DSMs: A Guide to Methodology. Committee on Earth 615 Observation Satellites, Terrain Mapping Subgroup. 616 Ducke, B., Score, D. and Reeves, J., 2011. Multiview 3D reconstruction of the archaeological site at Weymouth from 617 image series. Computers & Graphics, 35 (2), 375-382. 618 ESRI, 2014. ArcGIS 10.1 Desktop Help. Cluster and Outlier Analysis (Anselin Local Moran’s I) (Spatial Statistics). 619 Field, A. 2013. Discovering Statistics Using IBM SPSS Statistics. London: Sage. 620 Fisher, P. F. and Tate, N. J., 2006. Causes and consequences of error in digital elevation models. Progress in 621 Physical Geography, 30 (4), 467-489. 622 Gallant, J. C. and Wilson, J. P., 2000. Primary Topographic Attributes. In: Wilson, J. P., and Gallant, J. C., eds. 623 Terrain Analysis. Principles and Practice. John Wiley & Sons Inc. 624 Green, S., Bevan, A. and Shapland, M., 2014. A comparative assessment of structure from motion methods for 625 archaeological research. Journal of Archaeological Science, 46 (0), 173-181. 626 Hullo, F., Grussenmeyer, P. and Fares, S., 2009. Photogrammetry and dense stereo matching approach applied to 627 the documentation of the cultural heritage site of Kilwa (Saudi Arabia). International Archives of Photogrammetry, 628 Remote Sensing and Spatial Information Sciences, XXXVIII-3. 629 ICOMOS General Assembly, 1996. Principles for the Recording of Monuments, Groups of Buildings and Sites, Sofia, 630 Bulgaria 5-9 October. 631 Jacobson, R.E., 2000. Life expectancy of imaging media. In: Jacobson, R. E., Ray, S. F., Attridge, G. G. and Axford, 632 N., eds. The Manual of Photography. Ninth Edition. Focal Press. 633 Koutsoudis, A., Vidmar, B. and Arnaoutoglou, F., 2013. Performance evaluation of a multi-image 3D reconstruction 634 software on a low-feature artefact. Journal of Archaeological Science, 40 (12), 4450-4456. 635 Lam, K. W. K., Li, Z. and Yuan, X., 2001. Effects of JPEG compression on the accuracy of digital terrain models 636 automatically derived from digital aerial images. Photogrammetric Record, 17(98): 331–342. 637 Langford, M. 1998. Advanced Photography, 6th Ed. Focal Press. 638 Leica Geosystems AG (no date). Leica Viva GS15 Data Sheet [online]. Available from: http://www.leica-639 geosystems.co.uk/downloads123/zz/gpsgis/Viva%20GS15/brochures-datasheet/Leica_Viva_GS15_DS_en.pdf 640 [Accessed: 22nd November 2015]. 641 Macdonald, A.S. 1992. Air Photography at the Ordnance Survey from 1919 to 1991. The Photogrammetric Record. 642 14(80): 249-260 643 Maune, D. F., Kopp, S. M., Crawford, C. A. and Zervas, C. E., 2007. Introduction. In: Maune, D. F., ed. Digital 644 Elevation Model Technologies and Applications: The DSM Users Manual, 2nd Edition. Bethesda, Maryland: The 645 American Society for Photogrammetry and Remote Sensing. 646 McCarthy, J., 2014. Multi-image photogrammetry as a practical tool for cultural heritage survey and community 647 engagement. Journal of Archaeological Science, 43 (0), 175-185. 648 Miller, P., Mills, J., Edwards, S., Bryan, P., Marsh, S., Mitchell, H. and Hobbs, P., 2008. A robust surface matching 649 technique for coastal geohazard assessment and management. ISPRS Journal of Photogrammetry and Remote 650 Sensing, 63 (5), 529-542. 651 Mitchell, H., 2007. Fundamentals of photogrammetry. In: Fryer, J., Mitchell, H., and Chandler, J., eds. Applications of 652 3D Measurement from Images. Caithness, Scotland: Whittles Publishing. 653 Murphy, P., Thackray, D. and Wilson, E., 2009. Coastal Heritage and Climate Change in England: Assessing Threats 654 and Priorities. Conservation and Management of Archaeological Sites, 11 (1), 9-15. 655 Oakey, T. 2008. Candid Canberra. Aeroplane Monthly. 36(2): 34-38 656 Odle, J. E. 1965. The Williamson F49 Mark 4 Air Survey Camera. The Photogrammetric Record. 5 (25): 37–49 657 Owen, T. and Pilbeam, E., 1992. Ordnance Survey: map makers to Britain since 1791. HMSO: London. 658 Oxford Archaeology, 2002. Management of Archaeological Sites in Arable Landscapes. London: DEFRA. 659 Papasaika, H., Poli, D. and Baltsavias, E., 2008. A Framework for the Fusion of Digital Elevation Models. In: XXI 660 ISPRS Conference "Silk Road for Information From Imagery" Beijing, China. 661 Papworth, H., 2014. An assessment of archive stereo-aerial photographs for 3-dimensional reconstruction of damaged 662 and destroyed archaeological earthworks [online]. (PhD). Bournemouth University. Available from: 663 http://eprints.bournemouth.ac.uk/21780/ [Accessed: 6th January 2016]. 664 Pérez Álvarez, J. A., Herrera, V. M., Martínez del Pozo, J. Á. and de Tena, M. T., 2013. Multi-temporal archaeological 665 analyses of alluvial landscapes using the photogrammetric restitution of historical flights: a case study of Medellin 666 (Badajoz, Spain). Journal of Archaeological Science, 40 (1), 349-364. 667 Plets, G., Verhoeven, G., Cheremisin, D., Plets, R., Bourgeois, J., Stichelbaut, B., Gheyle, W. and De Reu, J., 2012. 668 The Deteriorating Preservation of the Altai Rock Art: Assessing Three-Dimensional Image-Based Modelling in Rock 669 Art Research and Management. Rock Art Research, 29 (2), 139-156. 670 Royal Commission on Historic Monuments, 1970. An Inventory of the Historical Monuments of the County of Dorset: 671 Volume II, South-East Part 3. 672 Royal Commission on Historical Monuments, 1952. An Inventory of the Historical Monuments in Dorset: Volume I, 673 West Part 1. 674 Rowley, T. and Wood, J., 2008. Deserted Villages. Shire Publications Ltd. 675 Sevara, C., 2013. Top Secret Topographies: Recovering Two and Three-Dimensional Archaeological Information from 676 Historic Reconnaissance Datasets Using Image-Based Modelling Techniques. International Journal of Heritage in the 677 Digital Era, 2 (3), 395-418. 678 Walstra, J., Chandler, J. H., Dixon, N. and Dijkstra, T. A., 2004. Time for change - quantifying landslide evolution using 679 historical aerial photographs and modern photogrammetric methods. In: Proceedings of the 20th ISPRS Congress. 680 Istanbul, Turkey, 12-23 July 2004. 681 Walstra, J., 2006. Historical Aerial Photographs and Digital Photogrammetry for Landslide Assessment. (PhD). 682 Loughborough University. 683 Walstra, J., Chandler, J. H., Dixon, N. and Wackrow, R., 2011. Evaluation of the controls affecting the quality of spatial 684 data derived from historical aerial photographs. Earth Surface Processes and Landforms, 36 (7), 853-863. 685 Wessex Archaeology, 2001. Lulworth Ranges, South Dorset. Archaeological Integrated Land Management Plan 686 (ILMP) Baseline Study. Desk-Based Archaeological Assessment and Monument Condition Survey. 687 Whitaker, K.A. 2014. The Wild Heerbrugg A5 in Britain in 2014. The Photogrammetric Record. 29(148): 456–462 688 Wilson, D. R., 2000. Air Photo Interpretation for Archaeologists. Stroud: Tempus Publishing Ltd. 689 Wolf, P. R. and Dewitt, B. A., 2000. Elements of Photogrammetry: with Applications in GIS. 3rd. Boston, USA and 690 London, UK: McGraw-Hill. 691 Verhoeven, G., 2011. Taking computer vision aloft – archaeological three-dimensional reconstructions from aerial 692 photographs with photoscan. Archaeological Prospection, 18 (1), 67-73. 693 Verhoeven, G., Doneus, M., Briese, C. and Vermeulen, F., 2012a. Mapping by matching: a computer vision-based 694 approach to fast and accurate georeferencing of archaeological aerial photographs. Journal of Archaeological 695 Science, 39 (7), 2060-2070. 696 Verhoeven, G., Taelman, D. and Vermeulen, F., 2012b. Computer Vision-Based Orthophoto Mapping of Complex 697 Archaeological Sites: The Ancient Quarry of Pitaranha (Portugal-Spain). Archaeometry, 54 (6), 1114-1129. 698 1. Introduction 2. Data and Methods 2.1 Field Site Selection 2.1.1 Baseline Data Collection 2.2. Archive Stereo-Aerial Photographs and Airborne Laser Scanning 2.3. Photogrammetric Processing 2.4. Validation Methods 2.4.1. Quantitative Assessment 2.4.2. Summary Statistics 2.4.3 Spatial Autocorrelation of Errors 2.4.4. Qualitative Assessment 3. Results 3.1 Quantitative Assessment of DSM Quality 3.1.1 Summary Statistics 3.1.2 Spatial Autocorrelation of DSM Errors 3.2 Qualitative Assessment of DSM Data 4. Discussion 4.1 SAP Age 4.2 Image Resolution 4.3 Photographic Materials and Scanning Technologies 4.4 Control and Tie Point Distribution 4.5 Camera Calibration Information 4.6 Archaeological Implications 5. Conclusion 6. Acknowledgements 7. References