Re-Using Historical Aerial Imagery for Obtaining 3D Data of Beach-Dune Systems: A Novel Refinement Method for Producing Precise and Comparable DSMs

Abstract

This study explores the potential of repurposing historical aerial photographs to produce high-accuracy digital surface models (DSMs) at regional scales. A novel methodology is introduced, incorporating road points for quality control and refinement to enhance the precision and comparability of multitemporal DSMs. The method consists of two phases. The first is the photogrammetric phase, where DSMs are generated using photogrammetric and structure from motion (SfM) techniques. The second is the refinement phase, which uses a large number (millions) of points extracted from road centrelines to evaluate altimetric residuals—defined as the differences between photogrammetric DSMs and a reference DSM. These points are filtered to ensure that they represent stable positions. The analysis shows that the initial residuals exhibit geographical trends, rather than random behaviour, that are removed after the refinement. An application example covering the whole coast of the Valencian region (Eastern Spain, 518 km of coastline) shows the obtention of a series composed of six DSMs. The method achieves levels of accuracy (0.15–0.20 m) comparable to modern LiDAR techniques, offering a cost-effective alternative for three-dimensional characterisation. The application to the foredune and coastal environment demonstrated the method’s effectiveness in quantifying sand volumetric changes through comparison with a reference DSM. The achieved accuracy is crucial for establishing precise sedimentary balances, essential for coastal management. At the same time, this method shows significant potential for its application in other dynamic landscapes, as well as urban or agricultural monitoring.

1. Introduction

Beach dune systems have a highly dynamic morphology as they are closely interrelated elements conditioned by hydrodynamic processes [1]. Dunes grow from sand accumulated on the emerged beach and can return sediment to the latter after coastal storms, therefore playing an essential role in beach recovery and coastal evolution [2]. Due to the erosion problems detected on many of the world’s beaches and the strong environmental, social and economic repercussions associated [3] monitoring actions are a growing need for the management of the coast [4]. The integrated assessment of the coastal spaces is essential for an efficient management that ensures the physical sustainability of beaches. With this regard, quantifying volumetric changes is crucial for calculating sediment budgets within coastal sediment cells [5]. While pluriannual topo-bathymetric surveys allow the accurate characterisation of the beach dune system behaviour, they constitute an in situ and time-consuming task that has been limited to a relatively small number of coastal stretches (e.g., [6,7,8,9]). This impedes its worldwide application for coastal monitoring.

The last decades have seen the emergence of new techniques in coastal monitoring capabilities, mainly driven by new remote sensing techniques [10]. These advancements enable extracting detailed information over relatively large areas, facilitating regional-scale and long-term morphological characterisation. This is the case of the topographic mapping based on tidal inundation frequency [11,12,13] with applications for intertidal spaces, as well as the satellite-derived shorelines obtained from optical imagery as a bi-dimensional descriptor of the beach morphology [14]. This latter technique makes it possible to quantify the beach responses to disruptive events (e.g., [15,16]), as well as multidecadal trends (e.g., [17,18,19]). However, morphological changes occur along the entire dynamic part of the beach, including the foredune. For a holistic characterisation of the beach state, the analysis of sediment volume needs to consider both the emerged beach and the dune. Unfortunately, approaches based on optical satellite imagery and shoreline definition fall short of quantifying volumetric changes. Contrary to 2D techniques, quantitative beach assessment in 3D is still relatively rare [20] as the processes to derive altimetry are more complex and/or expensive.

The most widely developed and adopted remote technique for obtaining 3D morphological data is LiDAR [21]. Many countries have carried out LiDAR surveys of their territories, allowing them to characterise the state of their coasts at the national scale. A recent review carried out in Europe [22] has shown that of the 32 countries analysed (EU27+, Norway, Switzerland, UK, Serbia), 18 of them were fully covered by LiDAR data and nine were partially covered, while no data were available for the remaining five. The elaboration of large-scale LiDAR point cloud data collection is time-intensive and data-heavy. Nevertheless, in some countries, such as Spain, two complete flights covering the whole territory were available and a third one is currently in progress (https://pnoa.ign.es/web/portal/pnoa-lidar/presentacion, accessed on 6 February 2025). Due to the high accuracy (tens of centimetres) and large spatial coverage, this source of information is sufficiently precise to quantify the volumetric changes in the beach-dune system (e.g., [23,24]. However, using LiDAR data to analyse the beach dune system’s response to specific events (e.g., coastal storms or large human interventions) or its long-term evolution faces several challenges, including limited historical depth (as the commercialisation of LiDAR data accelerated only at the start of the 21st century), the low frequency of surveys and, in most cases, reduced areas.

LiDAR is not the first aerial technique used for 3D information acquisition. Photogrammetry has been widely applied since the 20th century, particularly from the 1950s. Although initially analogue, its foundations were well-established early on. Terrain relief was traditionally extracted from aerial surveys using stereoplotters and later adapted to digital workstations [25]. Today, many cartographic institutions use pre-developed DSMs to project aerial photographs, prioritising efficiency and visual quality but often underutilising the 3D data’s full potential. The same stereoscopic mathematical background has undergone a profound development from the digital perspective with computer vision, stereo vision, multi-view stereo (MVS) and structure from motion (SfM). These techniques work in the image space, making possible the relative location of the photographs and 3D objects, and the reconstruction through the bundle adjustment of massively correlated points [26]. When these technological developments are also linked to terrain coordinates, they constitute a complete renewal of the photogrammetric paradigm. The application of these techniques, based on computer vision, employs large collections of photographs of the same scene or location. In geoinformatics, this scenario takes place when working with UAVs and its use is expanding into various fields such as archaeology [27], geomorphology [28,29] and forestry studies [30], already applying these techniques to aerial images. This trend is increasingly evident in coastal research. As reviewed by Casella et al. [31], 48 studies were conducted between 2015 and 2020 using UAV surveys over limited coastal areas to quantify volumetric changes over short timeframes. These surveys often focus on assessing the impacts of specific events, such as storms [32,33,34] or seasonal changes [35], though studies extending beyond a year remain rare.

The new scenario offers great potential for re-using the aerial imagery acquired by institutional mapping services for other purposes (i.e., aerial photographs for the elaboration of orthophotographs) also to provide 3D information and, particularly, to monitor volumetric evolution over beach and foredunes zones. The potential of using old aerial images has been proved in coastal areas [36] and to make sediment budgets in certain coastal stretches [37,38,39] though its usage also faces challenges, such as processing becoming increasingly difficult for older flights due to lower image quality [39,40] and the lack of Ground Control Points (GCPs) because of landscape changes.

The main objective of this work is to demonstrate that it is possible to obtain detailed three-dimensional models of large territories with a sufficient level of accuracy to be useful for quantifying sediment volumetric changes from the series of aerial photographs originally acquired to generate orthophotographs and other auxiliary cartographic information. The application for monitoring beach-dune systems aims (i) to demonstrate the coverage potential by creating six annual photogrammetric DSMs over a large area (more than 8000 km²), (ii) to propose a novel method for evaluating DSM quality using road points alongside traditional photogrammetric metrics, (iii) to highlight potential improvements achievable with these road points and (iv) to present an example of volumetric assessment by using a pre-existent LiDAR point cloud.

2. Materials and Methods

2.1. Materials and Study Site

The work developed in this paper has taken place in the Valencian region (eastern coast of Spain, Figure 1). This paper employs the aerial photographs produced by the cartographic institution of the Valencian region, the Institut Cartogràfic Valencià (ICV, https://icv.gva.es/es/, accessed on 6 February 2025). Since 2017, the ICV has conducted annual flights over the entire territory of Valencia to obtain annual orthophotographs with a spatial resolution of 25 cm/pixel. These images are projected onto a pre-existing and refined digital surface model (DSM) rather than that obtained from each photogrammetric flight.

The orthophotographs developed by the ICV are publicly available. The datasets cover the period from 2017 to 2022, with some additional years, such as 2015, that will also be used. This information has served as the planimetric reference for this project as it will be seen afterwards. For each year, the ICV conducted photogrammetric flights between April and June. The original photographs presented four channels (red, green, blue and infrared) and a GSD (Ground Sampling Distance, average resolution of the pixels projected on the ground) of 22 cm. The longitudinal and transversal overlaps are typically over 60% and 30%, respectively.

For this project, aerial photographs covering the coastal zone and dry riverbeds over the Valencian territory were provided by the ICV covering more than 8000 km² (Figure 1). Given that the flights follow the East–West direction, whilst the coast follows the North–South direction, the narrowest coastal segments were normally covered by four images.

The ICV supplied 13,997 images (

Table 1. Number of photographs per year.

Year	2017	2018	2019	2020	2021	2022
no. photographs	2418	2264	2271	2498	2297	2249

) together with the camera calibration (internal orientation), projection centre coordinates and rotation angles (external orientation).

In addition to these datasets, the Spanish National Territory Observation Plan (PNOT, using its Spanish acronym) provides LiDAR data for the entire country, with different regions surveyed over the years. For the Valencian region, central and southern areas (Valencia and Alicante provinces) were surveyed in 2015 with a density of 0.5 points/m², while the northern area (Castellón province) was surveyed in 2017 at a density of 1 points/m². This LiDAR data are freely available through the National Centre of Geographical Information (CNIG, https://centrodedescargas.cnig.es/CentroDescargas/index.jsp, accessed on 6 February 2025). The planimetric and altimetric precisions of the LiDAR point clouds are expected to be 0.2 m and 0.3 m RMSE (Root Mean Squared Error), respectively (https://pnoa.ign.es/pnoa-lidar/especificaciones-tecnicas, accessed on 6 February 2025). For this project, both LiDAR point clouds were downloaded and processed to create the 2015 and the 2017 LiDAR reference DSM, hereafter the reference DSM, with a resolution of 1 m/pixel, which will serve as an altimetric reference.

The road centrelines were extracted from the Valencian 1:5000 cartography, freely available in vector format. As most roads have remained unchanged over the study period, they can be considered suitable for comparisons and evaluations. Points were defined at 3 metre intervals along the road centrelines, and the altimetric differences at these points, between each photogrammetric DSM and a reference DSM, were used to assess and calibrate their similarity and comparability.

In summary, the datasets used in this project include the official orthophotos from 2015 to 2022, aerial photographs and orientation metadata from 2017 to 2022, the LiDAR reference DSM and road centreline points. These inputs were utilised in the subsequent analyses.

2.2. Methodology

This paper is divided into two distinguishable parts. On the one hand, it shows the procedure used to provide the best accuracy and interoperability among a time series of DSMs, which is divided into a photogrammetric phase followed by a refinement phase (Figure 2). On the other hand, an example of applying a volumetric time series based on previous DSMs to show the reliability potential of the methodology is presented. In the first part, six datasets of aerial photographs were used to generate their respective photogrammetric DSMs. Simultaneously, and assuming their stability over time, roads served as a vertical reference. Their positions were extracted from institutional cartography and converted into points, then used to model vertical errors across the study area and produce residual raster models (RRMs). These RRMs were subsequently applied to refine each photogrammetric DSM.

2.2.1. DSM: Photogrammetric and Refinement Processes

Obtaining the annual DSMs followed two main steps: the photogrammetric phase and the refinement (see Figure 2). The first phase aimed to produce the most precise DSMs from the photogrammetric point of view. The second phase allowed us to model the slight deformations remaining between each of those DSMs and the reference DSM, and using this modelling to improve the final DSM. As it can be seen, the process aimed to provide not only the best accuracy but also the best comparability among DSMs for the time-series analyses.

Photogrammetric Process

The software Agisoft Metashape 2.2.0 was used for the photogrammetric processing. This software leverages SfM techniques, which identify and correlate large numbers of homologous points between images to resolve the scene’s relative alignment.

In an initial approximation, the process benefited from the aerial photographs and the exterior orientation provided for each photograph. Although the resulting DSMs were georeferenced, they showed coarse altimetric errors, frequently higher than 1 m. To solve this, GCPs were extracted from existing cartographic datasets (see Figure 1B) as follows: (i) the set of orthophotographs of 2015 and between 2017 and 2022; and (ii) the reference DSM. The GCPs had to meet two conditions: (i) they must be present throughout the study period, and (ii) they must be located in flat areas. Ideal examples of GCPs included those located on tennis courts or parking lots. It is worth noting that the historical orthoimages, available via WMS, were generated from the same photographs reprocessed in this study, ensuring the consistent presence of these points. Once the GCPs were planimetrically identified, their elevations were derived from the reference DSM. In any case, as a general rule, an RMSE of up to two times the spatial resolution is expected. We were conscious of this, which is why only one year of data was used to map the location of the GCPs, while other years were employed to ensure the consistency of these recognisable points over time. The confidence among points might vary, particularly between urban and rural areas. To account for these differences, weights were applied to each selected GCP. The first condition ensured a consistent multitemporal dataset of GCPs, while the second condition minimised the risk of significant altimetric errors caused by point selection inaccuracies.

In some regions, identifying suitable GCPs posed significant challenges due to the lack of distinguishable features within the landscape. Furthermore, during the post-processing phase, it was occasionally discovered that certain GCPs were either inaccurately identified or exhibited instability, necessitating their exclusion from the dataset. Such adjustments required reprocessing the model, which, given the substantial volume of data, was computationally intensive and time-consuming. To address this limitation and enhance processing efficiency, we established a computing cluster comprising multiple machines equipped with processors ranging from 8 to 20 cores and GPUs spanning from RTX 2060 to RTX 4070 Ti. This configuration provided an aggregate computational capacity of 48 cores and 24 GB of VRAM, significantly accelerating the photogrammetric workflow.

DSM Refinement

In addition to assessing GCP errors, road points were used for a more detailed evaluation of the resulting DSMs. Altimetric values for each point composing the road centrelines were extracted from both the annual photogrammetric DSMs and the reference DSM (i.e., from the 2015/2017 LiDAR), and their differences—residuals—were used as error measurements. As presented in the results section, these errors followed a geographically continuous pattern rather than behaving randomly, suggesting the possibility of modelling and reducing them with a refinement process.

In this process, not all road points were suitable for this analysis as some points were located near buildings or trees, which may lead to large altimetric discrepancies. In order to select the suitable road points to describe potential geographical trends, two values were used, residual and incoherence, which will be defined specifically for this paper. Residuals, as typically used, were defined as the difference between each photogrammetric DSM and the reference DSM. In the first step, any point with residuals greater than 5 metres was removed. This threshold was chosen as it is large enough to capture potential trends but small enough to exclude extreme outliers that could significantly impact the statistical analyses. On the other hand, incoherence was defined as the difference between a point’s specific residual and the median residual of the neighbouring points within a defined radius. This concept of incoherence assumed that residuals exhibited consistent geographical trends among nearby points, and therefore, points deviating beyond a specified threshold could be filtered. For instance, a hypothetical region could have homogeneous residuals of around 1.5 m. Although this is a coarse residual average, if the area is homogeneous, the incoherence would remain near zero allowing these points to contribute to modelling geographical trends, so as to remove those aprioristic 1.5 m residuals. This procedure was repeated independently for each of the six photogrammetric DSMs. The incoherence threshold was set to 0.1 m considering the specific results of this work. In this case, the average incoherence reached was 0.1 m. Therefore, 1 m was considered as a proper threshold, with considerable margin, to remove points that were supposed to remain constant.

The remaining road points were used to model the undulation of residuals in annual residuals raster models (RRMs). From these remaining points, one part (70%) was used to create (to train) the RRMs whilst the other part (30%) was used for assessing (to test) the resulting correction. The RRMs were created through ordinary kriging interpolations. Each corrective DSM was applied to the corresponding photogrammetric DSM to remove geographical trends. The final assessment was based on these refined DSMs, resulting in updated residuals for both the 70% of road points used for training and for the 30% reserved for testing.

2.2.2. Volumetric Change Measurement: Example of Application for Quantifying Sediment Changes on the Valencian Coast

To provide a demonstration of the application of the methodology, a coastal segment between the Port of Valencia and the Cape of Cullera, in the centre of Valencian region (see Figure 1B) was used to illustrate the quantification of the volumetric changes over time. This section was selected due to the presence of a well-defined foredune susceptible to showing important volumetric changes associated with Storm Gloria in January 2020, as it occurred on many beaches along the Western Mediterranean (e.g., [16,41,42]). The previously exposed refinement improved the comparability between DSMs and enabled the calculation of the volumetric changes with better precision. The volumetric change time series was derived by calculating the difference between each photogrammetric DSM and the 2015 reference DSM, hereafter referred to as DoD (Difference of DSMs), following the methods outlined by James et al. [43], Wheaton et al. [44] and Milan et al. [45]. A set of six DoDs (one per year) was elaborated, showing the volumetric evolution starting from the reference DSM. Both the beach and foredune areas in this zone appeared largely free from vegetation, so no extra processing was needed.

Nevertheless, not all detected changes had the same level of reliability. The limit of detection (LoD) represents the smallest detectable change that can be considered reliable. The value of LoD depends on the accuracy of the products being compared. For example, a 0.3 metre change at a pixel between two DTMs would be significant if their accuracy is 0.05 metres, but not if it is 0.5 metres. In this study, the LoD reference value was derived from the statistical data of road points. Here, the standard deviation of the residuals—one per year—at these points reflected the expected uncertainty when comparing volumetric changes between DSMs. The differences between models were only taken into account if their magnitude were greater than these thresholds.

3. Results

3.1. DSM: Photogrammetry and Refinement

After the photogrammetric alignments, the GCP residuals for the XYZ coordinates were calculated. As expected, following a least squares alignment, the average residuals approached zero, with the standard deviation representing the error. This error (

Table 2. Error statistics at the GCPs after the photogrammetric alignment process.

Year	Number of GCPs	Total Error (m)	XY_Error	X_Error	Y_Error	Z_Error
2017	439	0.289	0.249	0.171	0.181	0.148
2018	542	0.283	0.256	0.179	0.184	0.121
2019	542	0.361	0.294	0.195	0.219	0.210
2020	550	0.288	0.250	0.159	0.193	0.142
2021	480	0.287	0.251	0.164	0.190	0.140
2022	445	0.302	0.272	0.181	0.204	0.131

) is further divided into planimetric and altimetric components for a detailed analysis.

The error of each component is the standard deviation of the residuals. The XY deviations move around 25 cm, coherent with the original geometric resolution, 25 cm/pixel, of the orthophotographs from which the GCPs were photointerpreted. The altimetric deviations are below 20 cm, which is also coherent with the technical requirements of LiDAR data acquisition. At this point, the GCP errors are coherent with the accuracy of the reference data.

Further evaluation involved calculating residuals (the differences between the photogrammetric DSMs and the reference DSM) at the road points. In non-changing zones, the differences between the DSMs are expected to be highly consistent. Based on this idea, we define incoherence as the deviation of a point’s residual from the residuals of its neighbouring points. The larger this deviation, the more incoherent the point is relative to the expected consistency of the DSM differences.

A hexagonal grid with a radius of 1.2 km was used to cover the entire area. For each hexagon, the median of the annual residuals was computed. Points with residuals differing by more than 1 m from their hexagon’s median were identified as outliers and removed. The hexagon aims to be a simplification of the ideal circular form.

As a result, only points whose residuals are coherent with the general trend of their surrounding area were retained, regardless of their absolute residual values. The remaining residuals can be visualised as individual points (Figure 3A), and their median values can be displayed for each hexagon (Figure 3B).

As shown in Figure 3, the residuals (and median of hexagons) do not behave randomly but instead exhibit geographical trends. The key observation is that, while adding GCPs improves photogrammetric alignment, non-stochastic errors may still arise in large areas. The presented technique effectively identifies the specific locations where additional GCPs should be added if local residuals exceed acceptable limits. However, repeating this process across the entire area may be impractical due to its size. Additionally, most observed trends fall within the project’s tolerance (20 cm as defined for LiDAR projects), making adding more GCPs potentially unnecessary.

To refine the DSMs, 70% of the road points were used to create residual raster models (RRMs) at the same resolution as the originals. By adding these RRMs from their corresponding photogrammetric DSMs, a set of refined DSMs was obtained. New residuals were then calculated, showing (Figure 4A,B) that the geographical trends have been removed.

Figure 5 shows the overlap between histograms of the residuals before and after the refinement process. It is worth noting that the original data should not be considered a poor result, as the photogrammetric process is generally accurate. However, the refinement process centres the error distribution and ensures a normal distribution, which is not achieved with the primary results from 2017.

The errors appeared below 20 cm, and, importantly, Figure 6 demonstrates that the residuals now behave randomly across the entire zone. This is the ideal outcome expected after performing the photogrammetric least squares adjustment over the whole area.

The numerical results of the shown maps are summarised in

Table 3. Average and standard deviations (both in m) of 6.9 million road points—training and testing points—before and after the refinement.

	4.5 Million Training Road Points		1.6 Million Testing Road Points
	Pre	Post	Pre	Post
2017	0.0574 ± 0.437	0.0006 ± 0.206	0.0701 ± 0.474	−0.0006 ± 0.193
2018	0.0146 ± 0.351	0.0003 ± 0.163	0.0522 ± 0.282	−0.0010 ± 0.157
2019	0.0429 ± 0.319	−0.0017 ± 0.163	0.0522 ± 0.293	−0.0022 ± 0.156
2020	0.0123 ± 0.258	−0.0010 ± 0.163	0.0292 ± 0.238	−0.0011 ± 0.158
2021	0.0417 ± 0.274	−0.0005 ± 0.184	0.0482 ± 0.248	−0.0006 ± 0.178
2022	0.0477 ± 0.250	0.0001 ± 0.183	0.0678 ± 0.249	−0.0008 ± 0.177

. The residuals before and after the refinement phase were equivalent both at the training (those implied at the kriging process) and test points (those not included at the kriging process but used to evaluate the result). The average residuals before the refinement are below 7 cm, which, in this case, would be considered negligible in agreement with the least squares criteria. However, their standard deviations range up to 0.5 m, nearly double the standard deviation of the GCPs.

These results must be reinterpreted with caution, as standard deviation typically assumes random error behaviour, which—as shown before—is not the case here. After the refinement, both the training and test points show millimetric average residuals, with new standard deviations reduced to a maximum of 20 cm. Figure 6 shows a local (35 km long) detail of the road points before and after the refinement process, showing how the main geographical trends have disappeared, and the remaining residuals behave randomly, as desired.

3.2. Volume Change Assessment

Using the DSMs obtained as described and after establishing the level of precision, it was possible to calculate the volume changes in a coastal area of 182.87 hectares stretching from the Port of Valencia to the Cape of Cullera (28 km length, see Figure 1B). For the analysis, the available series of DSMs were used (Figure 7A) and the resulting volume changes in the foredune were obtained by comparing each DSM against the reference DSM, which, in this zone, made use of 2015 LiDAR data. However, not every difference can be considered relevant as the precision of the DSMs limits the capability to detect whether a change is meaningful. As mentioned in the methodology, this is described as the limit of detection (LoD) which means that, depending on the accuracy of the compared DSMs, certain difference values will be reliable or not. In this case, the LoD was considered from the standard deviations obtained after the DSM refinement (Table 3 and

Table 4. LoD applied to each pair of years compared.

Difference Between Models	LoD Applied (In Metres)
2017–2015	0.206
2018–2015	0.163
2019–2015	0.163
2020–2015	0.163
2021–2015	0.184
2022–2015	0.183

). Every pixel with differences below the LoD was removed from the analyses (Figure 7B,C) and was not taken into account for the volumetric evolution.

Only significant pixels were used to calculate the DSM evolution (

Table 5. Volume changes against the DSM of 2015 obtained from LiDAR.

Year	Total Volume (m3)	Differences with Respect to 2015 (m3)	Percentage of Change (%)
2015	4,024,486.6	0.0	0.0
2017	3,521,001.5	−503,485.1	−12.5
2018	3,695,544.6	−328,942.0	−8.2
2019	3,640,938.1	−383,548.4	−9.5
2020	3,061,109.8	−963,376.7	−23.9
2021	3,175,796.8	−848,689.8	−21.1
2022	3,210,691.0	−813,795.6	−20.2

, Figure 8). As can be seen, the zone shows an erosive trend, with a clear loss of sediment exacerbated between the DSMs for 2019 and 2020. It is worth noticing that this phenomenon coincides with the impact of Storm Gloria in January 2020.

The results (Table 5) show that the beach dune system has undergone two phases of significant losses, separated by periods of slight recovery. In 2017, a loss of more than 503 thousand cubic metres of sand was quantified compared to 2015, partially recovered in 2018 with an increase of 174,543 m³. Between 2018 and 2019, the changes were minimal, with a loss of 54,606 m³. However, the 2020 model evidenced a significant loss again, with a reduction of 579,828 m³ compared to 2019. During the following two years, sediment recovery was observed, with increases of 114,686 m³ in 2021 and 34,894 m³ in 2022.

4. Discussion

The methodology for deriving 3D data at a regional scale presented in this paper aligns with the ongoing advancement of photogrammetric techniques in 3D coastal monitoring. They underscore the significant potential of modern photogrammetric methods to obtain 3D information by leveraging historical aerial photographs. These photographs mainly refer to those taken from the beginning of the 21st century onwards (originally used to generate orthophotos by projecting them onto pre-developed DSMs) and older aerial photographs with a retrospective view from the mid-20th century originally used to map the relief. The present work shows that it is possible to produce digital surface models (DSMs) with precision levels similar to those of more recent methods, such as LiDAR. The datasets of aerial photographs cover hundreds of kilometres, and unlocking their historical 3D information might be crucial for coastal monitoring, particularly sediment quantification on beaches and foredunes.

This new scenario has become possible thanks to computer vision techniques such as SfM, dense point cloud correlation and bundle adjustment of camera and object geometry [26]. While these techniques are not the primary focus of this study, achieving a reliable reconstruction of DSMs is neither straightforward nor automatic. It requires empirical testing of the hyperparameters available in the photogrammetric software—Metashape in this case—which are likely to vary depending on the type of terrain or object being reconstructed, as well as the specific characteristics of the images used. This consideration becomes especially critical when working with older aerial photographs [36]. Indeed, the study of these hyperparameters could itself be considered a distinct field of research [39].

The mentioned techniques are experiencing significant growth in their application to UAV-based projects [31]. UAV flights typically cover small areas with a high degree of overlap, greatly facilitating object reconstruction. However, applying these techniques to traditional photogrammetric flights presents a considerable challenge: such flights often span much larger areas—tens or even hundreds of kilometres—and are characterised by limited overlaps, typically around 60% longitudinally and 30% transversally.

This study introduces two key methodological innovations, both related to the use of road points for assessing and refining photogrammetric DSMs. Over a six-year analysis period, the assumption that most roads maintained consistent positions and altimetry allowed these points to serve as stable references for evaluating residuals between annual DSMs and a reference DSM. This novel use of road points provided a robust framework for quality control and improved the precision and comparability of the models. Traditionally, GCPs are employed to validate photogrammetric projects by comparing reference and calculated coordinates. In this study, the GCP accuracy is inside the range of the reference data resolution (both planimetric and altimetric), offering limited room for improvement. The first methodological key point was given by using road points provided as an alternative method, independent of the photogrammetric process, to assess errors across the entire study area. While the residuals along the roads were expected to behave randomly, the analysis revealed geographical trends or undulations. This finding showed the limitations of relying only on GCPs, which minimise residuals at specific locations, but leave gaps in accuracy between them. This constitutes one of the main differences in employing aerial photography covering large areas instead of UAVs.

It may be noticed that the methodology is suitable for being employed in a wide range of contexts in which DSMs may be applied. Thus, the methodology shows potential for its application in other dynamic landscapes and in urban or agricultural monitoring. In the case of the beach dunes systems, the highly homogeneous textures of the images could lead to inaccurate DSMs. This has not been an issue in this study case, and it does not constitute the focus of this research, although it could be a field of study itself.

As a second key issue, the road points were used to generate the correction rasters that mapped vertical residuals associated with each DSM. The novelty of the proposed methodology also lies in emphasising the importance of filtering reliable road points. In small study areas such as those covered by UAVs, it might not be necessary or, if applied, it could be done manually. On the contrary, the automation of this process has become essential for larger regions. The proposed methodology automates the filtering process by considering neighbouring points, ensuring that residual trends are preserved regardless of their magnitude. This approach enhances the reliability of the correction process and demonstrates the potential of road points as a scalable tool for DSM refinement in large-scale photogrammetric projects.

The correction rasters facilitate the refinement of the DSMs, bringing them closer to the reference DSM. The analysis also highlighted the difference between accuracy and precision: while the LiDAR DSM served as a reference, its absolute accuracy relative to the photogrammetric DSMs remained uncertain. However, the refinements significantly improved model consistency and intercomparability. In a practical sense, the difference between the two DSMs has a credibility limit, i.e., a limit of detection (LoD), as described by James et al. [43]. Considering the LoD uncertainty as the deviation of residuals expected in constant zones (i.e., road points), it was reduced from up to 40 cm before refinement to up to 20 cm after refinement (Table 2).

It is valuable to compare the accuracy achieved in this study with that reported in previous work using aerial photo series. Grottoli et al. [39] reported an altimetric RMSE of 0.93 m when working with digitised historical photographs from 1963, compared to a 2014 LiDAR survey. Similarly, Carvalho et al. [37] analysed coastal dunes and beaches in a small area (less than 6 km long) of Australia using aerial images (so not originally digital) from 1969, 1977 and 1986, and achieved RMSE values between 0.5 and 0.99 m. Carvalho and Reef [38] obtained more accurate results, with an RMSE of 0.48 m by comparing 1977 aerial photographs with 2007 LiDAR data. However, these studies covered much smaller areas and relied on much older imagery than the present study. In addition, these authors consistently pointed out the challenge of identifying reliable control points, a critical factor in achieving higher accuracy. In contrast, after refining the DSMs, the present study achieved an accuracy of 20 cm, representing a significant improvement. Nevertheless, its application over older images is to be tested.

The ability to reconstruct 3D models from aerial photographs on a large scale represents a paradigm shift in coastal change monitoring. This approach complements recent advancements in coastal characterisation derived from satellite imagery, which typically provides a two-dimensional perspective, such as shoreline extraction (e.g., [14,16]). However, significant changes in the volume of the beach and, particularly, the foredune are not always reflected in shoreline position shifts. By reusing aerial datasets, it is possible to achieve a volumetric characterisation of the entire beach-dune system with detailed altimetric data. This method of providing high-resolution altimetric information at a regional scale marks a turning point in remote morphological characterisation, expanding the scope of coastal monitoring to the scale of sedimentary cells, and enabling progress toward the goal of quantifying sediment budgets. Future research is envisioned to focus on the ability to characterise beach changes on diverse spatial and temporal scales, taking advantage of the complementarity between shoreline datasets and 3D data sources.

The results presented here, focused on a specific coastal segment, highlight how the dunes experience significant sediment losses during certain periods, such as those recorded between 2015 and 2017 and, most notably, between 2019 and 2020. These losses contrast with positive but less substantial changes observed during other years. The ability to discretise sediment budgets into short timeframes (almost annually) is extremely valuable for identifying the drivers of the observed changes. In this case, it is evident that major coastal storms, such as those that occurred in January 2017 and 2020, are the primary factors explaining the recorded alterations.

5. Conclusions

This study demonstrates the potential of using historical aerial photographs to generate regional-scale 3D data with accuracies comparable to modern LiDAR techniques. The methodology developed, which combines photogrammetric processing with a novel refinement approach based on road points, allows for systematic and efficient monitoring of volumetric changes across diverse landscapes. Although this work’s justification and application focus on the beach and the foredune, the approach is not limited to these environments and could be adapted to different landforms.

The use of road points for error assessment and model refinement constitutes a significant innovation by providing an automated and scalable solution for large-scale photogrammetric projects. By incorporating spatial trends and eliminating residual inconsistencies, the methodology improves the precision and comparability of the resulting DSMs, ensuring a reliable analysis over large and heterogeneous areas.

The ability to extract meaningful altimetric information from aerial imagery fills the gap left by the absence or limited availability of 3D data. By achieving levels of accuracy comparable to LiDAR through the re-use of historical photogrammetric flights—an economically viable solution—this methodology emerges as a practical tool for three-dimensional monitoring as demonstrated in this study through its particular application to a beach and foredune system.

In coastal monitoring, the regular and medium-term availability of high-precision volumetric data on the emerged beach and foredune is highly valuable. This data can complement and validate the trends identified through short-term two-dimensional analyses, enhancing the accuracy and reliability of coastal assessments.