Dynamically documenting archaeological excavations based on 3D modeling: a case study of the excavation of the #3 fossil of hominin cranium from Yunxian, Hubei, China

Abstract

Documenting tangible cultural heritage using 3D modeling techniques is gradually becoming an indispensable component of archaeological practice. The 3D modeling techniques based on photogrammetry and LiDAR scanning enable high-accuracy and high-realistic reconstruction of archaeological sites, and have been proven a powerful tool for documenting archaeological excavations. However, dynamically documenting an ongoing excavation using these techniques is still considered tedious, time-consuming, expensive, and dependent on expertise. Moreover, the application of 3D modeling techniques in archaeological excavations still faces some technical challenges, such as modeling with multi-source and multi-scale data, fusing local models at different times into a whole, achieving fast modeling while GPU workstations are not available in the field, and evaluating the quality of 3D models. As a result, there are still very few archaeological teams deeply engaged in dynamic documentation with 3D modeling techniques, and traditional drawing sketches and taking photographs still dominate. In these senses, documenting the archaeological excavation at the Yunxian Man site (located in Hubei, China) is an invaluable opportunity for exploration and practice. Archaeologists determined to conduct dynamically documenting at the beginning of the 6th excavation project for the site, and established a rotation system to reconcile physical excavation with digital preservation. Through repeated practice and communication, we proposed a workflow and pursued several new methods to enhance the feasibility of dynamically documenting, and obtained 4D models of the ongoing archaeological excavations. In 2022, the Yunxian Man site unearthed the most intact fossil of hominin cranium from about one million years ago in the Eurasian continent, preserving important and scarce anatomical features of early humans in Asia. As the original taphonomic context of the fossil corroded away during physical excavations, the digital documentation consisting of 4D models serves as permanent original data source in subsequent archaeological research. Moreover, we obtained cross-scale 3D models from geographical environment to archaeological site, excavation area, and cultural remains, and all of these 3D models are in an actual, unified coordinate framework. Thus, we can contribute to multidisciplinary cross-collaborative research through data sharing. Considering that digital documentations serve a great value in archaeological research, this paper focuses on sharing the workflow and methods to facilitate digital preservation for more archaeological projects.

Introduction

The irreversible nature and inevitable destructiveness of archaeological excavations make archaeologists place much importance on documenting and archiving. The 3D modeling techniques enable us to document high-accuracy and high-realistic 3D models for archaeological excavations. Performing multiple 3D modeling to document various momentary states of the excavation area can be formulated as dynamically documenting. Dynamically documenting enables comprehensive recording the occurrence and taphonomic context of the archaeological sites in their original stratigraphy and can provide a wealth of details for post-excavation research. As the original taphonomic context corroded away during physical excavations, the digital documentation serves as a permanent 3D data source. This data source supports unlimited observations, measurements, inferential analyses, and multidisciplinary and multi-institutional collaborative research through data sharing.

A brief review of the application of 3D techniques in archaeology and cultural heritage.

In the last 20 years, the application of 3D techniques in archaeology and cultural heritage studies has increased exponentially and has become very common in archaeological practice and cultural heritage conservation. Laser scanning and photogrammetric reconstruction are widely used 3D modeling techniques by archaeologists. Due to the ability to obtain high-precision geometric structures on the surface of objects, laser scanning is used for the protecting and documenting of cultural heritage [1], structural deformation monitoring [2], archaeological excavation documenting [3, 4], etc. Due to the ability to obtain both high-precision geometric structures and highly realistic textures, photogrammetric reconstruction techniques are widely used for small cultural relics [5], historical buildings [6,7,8,9], archaeological sites [9,10,11,12], and archaeological excavations [13,14,15]. Some studies have combined laser scanning and photogrammetric reconstruction techniques [16,17,18,19]. The abundance of excellent work leads scholars to review the contribution of 3D technology to archaeology and cultural heritage with much praise and more expectations [19,20,21,22].

However, we found in our research that there are not many cases of 3D modeling techniques applied to dynamically documenting archaeological excavation. It may be said that the number of archaeological projects with dynamically documenting of 3D modeling techniques is not proportional to the expectations of archaeologists [13, 20, 21]. Some researchers think the main reason is the lack of a straightforward method that can be flexibly and holistically integrated into the complete workflow of archaeological excavations [20]. Some researchers also believe that the efficient integration of 3D technology into daily field work requires a deeper understanding of the technology and some excellent examples [21].

Difficulties in implementing dynamically documenting archaeological excavations and our solutions.

In practice, some reasons and difficulties prevent archaeological teams from flexibly and dynamically applying 3D modeling techniques. These include but are not limited to: (1) There is difficulty in handling and integrating data from millimeter to kilometer scales (cross-scale) and captured by different devices (multi-source). (2) The 3D modeling process is time-consuming and sometimes interferes with fieldwork progress. (3) The highly dynamic process of the excavation makes frequent requirements of 3D modeling and then amplifies the former difficulties. (4) 3D modeling is usually considered indispensable to special equipment for data collection, complex software, and expensive GPU workstations for data processing, and it is especially indispensable to specialized technicians. (5) The varying quality of 3D model data makes archaeologists uneasy.

We consider these practical difficulties in a complete archaeological excavation and solve them with a proposed workflow and several new methods. (1) We deploy a multi-level control network and propose a strategy for fusing cross-scale and multi-source data. (2) We introduce fast cloud-based modeling technology to reduce the time spent on data processing, and solve the dilemma of efficiency requirements and budget shortages. (3) We propose a 4D model fusion method that only needs to focus on the sub-area that is being excavated, thus reducing the workload due to frequent modeling. (4) Together with spatial precision, our novel geometric refinement and texture resolution metrics provide a comprehensive measure of the quality of 3D models. (5) Our self-developed software is designed to be as simple and concise as possible, making our workflow and methods easy to implement and generalize.

The 6th excavation of the Yunxian Man site—our hands-on opportunity.

The Yunxian Man site , also known as the XuetangLiangzi site, is located in the village of Mituosi in the Yunyang District (formerly “Yunxian”) of Shiyan City, Hubei Province, China (Fig. 1). It is about 10 km from Qingqu Town and about 40 km down the Han River to the urban area of Yunyang [23]. The Quyuan River flows from north to south and merges into the Han River. XuetangLiangzi is an east–west gondola, the main body of which is the fourth alluvial terrace on the north bank of the Han River, the top of which is about 50 m above the water surface of the Han River. The site preserves an area of about 1.9 million square meters, with quaternary accumulations of more than 8 m and locally more than 18 m, preserving stratigraphic deposits from different periods over more than 1 million years [24].

Fig. 1

The location of the Yunxian Man site. The site is located in the village of Mituosi in the Yunyang District (formerly “Yunxian”) of Shiyan City, Hubei Province, China. The first hominin fossil found in 1989 was named Yunxian Man, the site was named the Yunxian Man site, and these archaeological names are followed to the present day

Five seasons of excavations between 1989 and 2007 revealed two complete skulls of Homo erectus and a large number of animal fossils and stone artifacts, which attracted great attention from the academic community [23, 25,26,27,28]. Combining various geological dating methods, the layer yielding the skull of Homo erectus was dated to between $1.1 \pm 0.16$ Ma and $0.581 \pm 0.093$ Ma [29, 30].

In 2021, the site was excavated for the 6th season with three excavation areas (the Excavation Area B, C, E in Fig. 2). In 2022, the #3 hominin cranium and a large number of animal fossils and lithic artifacts were found in the calcareous cemented sand stratum. A new round of interdisciplinary research on hominin fossils, archaeological remains, site formation process, and absolute dating using multiple methods is underway [24].

Fig. 2

The location of the excavation area. In 2021, the Yunxian Man site launched its 6th season of excavation and defined three excavation areas: the Excavation Area B, the Excavation Area C, and the Excavation Area E

The Yunxian Man site serves as an excellent opportunity for practicing the concept of dynamically documenting, benefiting from two advantages. Firstly, the site is located in an open and wide location, which makes it easy to fly UAVs and provides adequate light to implement the photogrammetry. Secondly, the archaeologists meticulously planned and organized the 6th excavation of the site, invited the digital preservation team to participate in the fieldwork from the beginning, and communicated repeatedly on the implementation issues of obtaining high-quality digital documentation.

As a result, we propose a workflow and pursue several new methods to practice dynamically documenting archaeological excavation. In this paper, we present the methods, implementation details, and major results. The self-developed software used in this paper, including GET3D, DasEarth, DasMesh, and DasViewer, can be acquired uniformly from the project homepage (https://wyniu.github.io/DDAE/), or from the product pages referenced in subsequent sections. Readers could contact the authors to get a free license for academic research and cultural heritage preservation purposes. In addition, we will continuously update the project homepage to provide some more data, scripts, tutorials and demo videos to help readers reproduce and promote the methods in this paper.

Materials and methods

To achieve dynamically documenting archaeological excavations practically, we develop a workflow and pursue several new methods. As shown in Fig. 3, the workflow is based on a multi-level control network and consists of five layers:

• (1) Multi-level Control Network: Constructing a multi-level control network to place cross-scale archaeological objects into an actual and unified coordinate framework.

• (2) Data Collection: Combining multiple methods to collect original data.

• (3) Data Processing: Getting 3D models from the original data.

• (4) Data Evaluating: Quantitatively evaluating the quality of the 3D models.

• (5) Data Application: Constructing a 4D model database, generating other digital products, and supporting comprehensive analysis and dynamic exhibition.

We apply our workflow and new methods to the excavation of the #3 fossil of hominin cranium from Yunxian. Our work involves some key techniques and their necessary equipment and software, we present them in Table 1.

Table 1 The key techniques and their necessary equipment and software

Fig. 3

The workflow of dynamically documenting archaeological excavations. The workflow is based on a multi-level control network and consists of five layers. This paper focuses on the former four layers

Theoretically, the layer of data application in this workflow belongs to the phase of after excavation. As we focus on how to perform data collection and 3D modeling, and our practice archaeological site—the Yunxian Man site is currently in the phase of during excavation, we mainly describe the key methods and techniques in the former four layers in this paper.

Multi-level control network

In general photogrammetric 3D reconstruction projects, GCPs (Ground Control Points) are often used to enable 3D models and 4D products (DOM, Digital Orthophoto Map; DEM, Digital Elevation Model; DRG, Digital Raster Graphic; DLG, Digital Line Graphic) to carry geographic coordinates. Before data acquisition, GCPs are deployed in the field, and their geographic coordinates are surveyed. It is necessary to ensure that GCPs are properly sized so that they are visible in the subsequently captured images. By stabbing the surveyed coordinates of the GCPs into the images, 3D models and 4D products can be assigned geographic coordinates.

In archaeological excavation, study objects are cross-scale, as small as millimeters or as large as kilometers. Therefore, it is infeasible to deploy one set of GCPs to fulfill the requirements of being visible in different-scale images or 3D models. To place cross-scale archaeological objects into an actual and unified coordinate framework, we propose a multi-level control network.

The multi-level control network is comprised of different levels and different styles of GCPs, each level visible in the images at a corresponding scale. Tailored for archaeological excavation, we divide the multi-level control network into four levels, Level 1, Level 2, Level 3, and Level 4. The measurement objects applicable to Level 1 to Level 4 sequentially are geographical environment (kilometer level), archaeological site (hundred-meter level), excavation area (meter level), and cultural remains (millimeter level).

Deploying the multi-level control network follows the principle:

• (1) While setting up the GCPs as uniformly as possible, we should consider the topography and the importance of the local area;

• (2) The GCPs in Level 1 and Level 2 should be kept as stable as possible;

• (3) The GCPs in Level 3 and Level 4 should be partly stable and partly dynamically adjusted according to the excavation.

The main equipment used to set up a multi-level control network includes RTK (Real Time Kinematic) and ETS (Electronic Total Station). RTK is used to survey the geographic coordinates of GCPs in areas where CORS (Continuously Operating Reference System) signals are available (usually open outdoor areas). ETS is able to survey geographic coordinates without a CORS signal (usually used indoors), but it requires setting up a station at a point location with known geographic coordinates. We use Daspatial miniRTK and THINRAD TTS-112R10M ETS in our practice.

Data collection

We mainly employ four different methods for data collection (Fig. 4).

• (1) Aerial photogrammetry is mainly accomplished by a UAV carrying a multi-lens camera. The purpose of this method is to efficiently acquire oblique images for a large-scale scene, which is the necessary input data for 3D modeling. This work requires a trained UAV pilot. After setting the appropriate parameters, the camera works automatically with the UAV’s flight. In practice, we use a DJI M300 UAV carrying a self-developed multi-lens camera named Daspatial Shuangyu 7X Pro to accomplish this work.

• (2) Close-range photogrammetry is mainly accomplished by a camera. The purpose of this method is to acquire close-range images for medium-scale or small-scale scenes. Since this method requires manually taking a large number of photographs, it is usually used for areas where UAVs cannot be maneuvered. With the improvement of our algorithms, this method is not restricted to the type of camera. Whether it is a DSLR camera or a phone camera, we just need to make sure that the images are sufficient in terms of definition and overlap. In practice, we use a Canon 5D Mark IV DSLR camera to accomplish this work.

• (3) Terrestrial laser scanning is mainly accomplished by a terrestrial laser scanner. The purpose of this method is to acquire laser-scanned point clouds for medium-scale scenes, which generally provide more accurate geometric information than close-range images. After setting up the scanner in a suitable location and emptying all moving people and objects, we can remotely control the device to start scanning. In practice, we use a Trimble $\times$7 terrestrial laser scanner to accomplish this work.

• (4) Portable laser scanning is mainly accomplished by a portable laser scanner coupled with a graphic workstation. The purpose of this method is to acquire a close-range point cloud for small-scale scenes, which is promising to provide the highest accuracy point cloud from the currently available techniques. This method is relatively inefficient due to the demand for affixing many laser marking points and the demand for repeated scanning and inspection. Therefore, we use this method only on significant cultural remains that require sub-millimeter scanning. In practice, we use a KScan-Magic II portable laser scanner to accomplish this work.

Fig. 4

Data collection methods. a Data collection through aerial photogrammetry. b Data collection through close-range photogrammetry. c Data collection through terrestrial laser scanning. d Data collection through portable laser scanning

Cross-scale data fusion 3D modeling

With the constraints of a cross-scale data fusion algorithm and the multi-level control network, we can make the cross-scale data in an actual and unified coordinate framework. Our cross-scale data fusion algorithm focuses on fusing cross-scale data and constructing a unified 3D model through the following steps:

• (1) Automatically group the cross-scale data into several clusters according to the differences in camera parameters.

• (2) Separately execute dense matching to obtain a point cloud for each cluster.

• (3) Perform adjustments within the cluster to optimize the point cloud, then obtain a high-precision point cloud for each cluster.

• (4) Stabbing GCPs of multi-level control network to make the point cloud of each cluster in the actual and unified coordinate frame.

• (5) Taking the point clouds of small-scale clusters as a benchmark to gradually fine-tune the scales of point clouds of other clusters, then obtain a cross-scale fused point cloud.

• (6) Execute Delaunay triangulation on the cross-scale fused point cloud to construct the geometric model.

• (7) Map textures for the geometric model using images. Note that images from small-scale clusters should be prioritized in areas of cross-scale overlapping.

In practice, we perform the above-mentioned operations with GET3D. As a self-developed software, we embed the cross-scale data fusion algorithm into the 3D modeling workflow. Constructing 3D models from cross-scale data with GET3D takes only 4 steps:

• (a) Data importing. Use the “Import Default Photos” button to import aerial images and the “Import Ground Photos” button to import photos taken with a DSLR camera or smartphone. After clicking the “Confirm” button, the software will automatically complete the above Step 1.

• (b) ATS-free. Click the “Submit Aerial Triangulation” button and select the “Free Network” mode in the pop-up window. After clicking the Confirm button, the software will automatically complete the above Step 2 and Step 3. These steps mainly employ the photogrammetric technique of aerial triangulation surveying with free network mode, thus abbreviated as ATS-free.

• (c) ATS-control. After the previous step is finished, use the “edit GCP” button to stab the GCPs of the multi-level control network, then click the “Submit Aerial Triangulation” button and select the “Control Network” mode in the pop-up window. After clicking the Confirm button, the software will complete the above Step 4 and Step 5. These steps mainly employ the photogrammetric technique of aerial triangulation surveying with control network mode, thus abbreviated as ATS-control.

• (d) Producing a 3D model. After the previous step is finished, click the “Submit Reconstruction” button to start generating 3D models, the software will complete the above Step 6 and Step 7.

After the program finishes running, we can get a cross-scale fused 3D model. Click the “View Product” button, then GET3D (Link) will automatically call the 3D model viewer DasViewer to display the model. After the program finishes running, we can get a cross-scale fused 3D model. Click the “View Product” button, then GET3D will automatically call the 3D model viewer DasViewer to display the model. Generally, GET3D concurrently saves 3D models in OBJ and OSGB file formats. OBJ is easier to edit, and OSGB is easier to load and browse quickly.

Multi-source data fusion 3D modeling

As photogrammetry and LiDAR scanning both offer their own advantages and disadvantages, we usually use the two methods simultaneously to complement each other for excavation areas and cultural remains. To improve on the two methods, we use the a multi-source data fusion algorithm to fuse their data. The algorithm contains the following steps:

• (1) Orient and scale the laser-scanned point cloud to assign it actual scale and geographic coordinates.

• (2) Execute ATS-free using photogrammetric images to obtain dense point cloud, and then stab GCPs and execute ATS-control to assign it actual scale and geographic coordinates.

• (3) Fuse laser-scanned point cloud and dense point cloud into a composite point cloud through ICP (Iterative Closest Point) alignment algorithm [31,32,33]. The priority of the ICP alignment reference is in the following order: terrestrial laser-scanned point clouds, portable laser-scanned point clouds, and image-matched point clouds. Next, only portable laser-scanned point clouds are retained in the fused point cloud; if there are no portable laser-scanned point clouds, the fused point cloud is processed with outlier noise rejection and uniform downsampling.

• (4) Execute Delaunay triangulation on the composite point cloud to construct the geometric model.

• (5) Map textures for the geometric model using photogrammetric images.

In practice, we perform the above-mentioned operations with GET3D. As a self-developed software, we embed the multi-source data fusion algorithms into the 3D modeling workflow. Constructing a multi-source fused 3D model using GET3D is similar to a cross-scale fused 3D model, which also includes four steps: data importing, ATS-free, ATS-control, and producing a 3D model. The difference is that when importing data, it is also required to use the “Import Ground LiDAR Point Cloud” button or the “Import Mobile LiDAR Point Cloud” button to import a laser point cloud.

Fast cloud-based 3D modeling

Typically, generating 3D models using images and point clouds is computility-intensive and time-consuming. In the process of dynamically documenting archaeological excavation, it is inevitable to encounter phased surges of modeling demand. To get rid of the restriction of GPUs (which is usually limited by the funding budget) and to cope with the urgent modeling tasks, we adopt the fast cloud-based modeling technique.

The fast cloud-based modeling technique requires computers in 3 roles, including a data server, a host server, and several GPU laborers. Its workflow mainly consists of the following steps:

• (1) The user uploads the modeling data to the data server.

• (2) The host server divides the modeling task into many similar sub-tasks and dynamically organizes and manages these sub-tasks.

• (3) Dispatch the required or available number of GPU laborers.

• (4) Each GPU laborer parallelly runs the following tasks in a loop until all sub-tasks are completed:

  • fetching a sub-task from the host server,

  • fetching the data required for the sub-task from the data server,

  • executing the sub-task and getting the result data,

  • sending the result data to the host.

Theoretically, only a small part of the procedure in modeling tasks is indivisible. Therefore, most modeling tasks can be chunked, parallelized, and then executed quickly. Our team has developed the fast cloud-based modeling technology as DasEarth (Link), an online modeling platform that is open to the public. To perform 3D modeling with DasEarth, all we need to do is upload data and stab GCPs, and then the system will automatically generate a 3D model.

4D model fusion

In archaeological practice, an excavation area is usually divided into several excavation sub-areas. Each excavation sub-area is different in the excavation progress, resulting in different frequencies and dates for modeling. To facilitate unified management and application, archaeologists need to integrate the 3D models of these scattered excavation sub-areas in different time series into an overall 3D model of the excavation area. The fusion of models from different times into one overall model can be viewed as the addition of a time dimension to a 3D model, so it is also called a 4D model.

To cope with this demand, we propose a method of 4D model fusion. This method enables editing 3D model as if it was a piece of cake, as we can cut away a patch and replace it with a new one. Given that the patch $P_{old}$ that needs to be cut away is on the old 3D model $M_{old}$, and the patch $P_{new}$ that needs to be made up is on the new 3D model $M_{new}$, our method constructs an updated 3D model $M_{update}$ through the following steps (Fig. 5):

• (1) Parameter transformation. Selecting a set of corresponding points $V\{(v_{o 1}, v_{n 1}), (v_{o 2}, v_{n 2}),..., (v_{o i}, v_{n i})\}$ on $M_{old}$ and $M_{new}$ (DasViewer $\rightarrow$ Measuring $\rightarrow$ Coordinates Measuring). Using V to calculate the rotation, translation, and scaling parameters, and applying these parameters to $M_{new}$ and make sure that $M_{new}$ coincide with $M_{old}$ (DasMesh $\rightarrow$ OBJ/OSGB Coordinate Transformation $\rightarrow$ 7 Parameter Transformation).

• (2) Specifying the update range. Manually sketch an updated range R (DasViewer $\rightarrow$ Marking $\rightarrow$ Add Line $\rightarrow$ Save as KML).

• (3) Cutting the patching. Use R to cut the $M_{old}$ and $M_{new}$, the part of $M_{old}$ outside R named $P_{out}$, the part of $M_{new}$ within the R named $P_{in}$. Combine $P_{out}$ and $P_{in}$ as $M_{update}$ (DasMesh $\rightarrow$ OBJ/OSGB cropping).

• (4) Topological re-connection. Re-connect the topology of the vertices at the edge of R, so that the $P_{out}$ and $P_{in}$ are truly fused into $M_{update}$ (DasMesh $\rightarrow$ Model Fusion).

• (5) Texture re-mapping. Re-map the texture of the triangles at the edge of R, so that the texture align with geometric model (DasMesh $\rightarrow$ Model Fusion).

Although the method of 4D model fusion is relative complicated, all the steps of it are integrated into the 3D model post-processing software DasMesh (Link) and the 3D model viewer software DasViewer(Link). We noted the menu location that provides the corresponding function at the end of each step, and the right-arrow indicates access to the sub-menu or button.

Fig. 5

The process of fusing 3D models. a, b Show the alignment of $M_{new}$ and $M_{old}$ through parameter transformation. c Shows the process of specifying the update range. d Shows shows cutting the patching through update range. e Shows the process of topological re-connection. f Shows that some texture errors have occurred. g, h Show the process and result of texture re-mapping

Evaluation metrics

It is a necessary scientific step that objectively and quantitatively evaluates the 3D model quality. Practically, we combine spatial precision, geometric refinement, and texture resolution to provide a comprehensive evaluation of 3D model quality.

Spatial precision. The spatial precision is a common metric in geomatics and photogrammetry. When deploying a multi-level control network, identical methods of setting up GCPs are used to deploy and survey a few of the checkpoints. The form of checkpoints is fully the same as the GCPs of the same level, but they are not involved in the ATS-control procedure. After obtaining the 3D model, we can calculate the planimetric and elevation errors through RMSE (root mean squared error) between the surveyed coordinates of checkpoints and their corresponding model coordinates.

In the case of a 3D model covering an area with a total of $n_1$ checkpoints, given the ith checkpoint with surveyed coordinates $O_{s i}(X_{s i}, Y_{s i}, Z_{s i})$ and model coordinates $O_{m i}(X_{m i}, Y_{m i}, Z_{m i})$, then the planimetric precision $E_p$ and the elevation precision $E_e$ of the 3D model is calculated as:

$$\begin{aligned} {\left\{ \begin{array}{ll} E_p = \frac{1}{n_1} \sum _{i=1}^{n_1} \sqrt{\left( X_{s i} - X_{m i} \right) ^{2} + \left( Y_{s i} - Y_{m i} \right) ^{2} } \\ E_e = \frac{1}{n_1} \sum _{i=1}^{n_1} \sqrt{\left( Z_{s i} - Z_{m i} \right) ^{2} } \end{array}\right. } \end{aligned}$$

(1)

In practice, we load the 3D model to DasViewer and then import the surveyed coordinates of the checkpoints. Thus, we can use the precision inspection function to complete the abovementioned operations.

Geometric refinement. The geometry of the 3D model is represented by the Delaunay triangular mesh formed by connecting 3D vertices. As we know, the more pixel points are used to represent an object, the photograph is clearer and finer. In the same way, the more triangular vertices are used to represent the same object, the 3D model is finer and richer in geometric detail. Therefore, in the same spatial extent, the more vertices, the finer the geometric structure. Since an edge of any triangle necessarily connects two vertices, and the edge lengths represent the actual distances between the two vertices in real-world coordinates, the edge length is an ideal measure for considering both the spatial extent and the number of vertices. We evaluate the geometric refinement of a 3D model by counting the average edge lengths of all triangles.

Given that the 3D model contains $n_2$ edges of a triangle, the ith edge connects vertices $V_{1 i}(X_{1 i}, Y_{1 i}, Z_{1 i})$ and $V_{2 i}(X_{2 i}, Y_{2 i}, Z_{2 i})$, then the geometric refinement $R_g$ of the 3D model is calculated as:

$$\begin{aligned} R_g = \frac{1}{n_2} \sum _{i=1}^{n_2} \sqrt{\left( X_{1 i} - X_{2 i} \right) ^{2} + \left( Y_{1 i} - Y_{2 i} \right) ^{2} + \left( Z_{1 i} - Z_{2 i} \right) ^{2} } \end{aligned}$$

(2)

In practice, we programmed the geometric refinement calculation of 3D models. We will soon develop a free and handy tool in DasViewer to provide this function.

Texture resolution. The same as images, 3D models also differ in clarity. However, unlike images that can be quantitatively evaluated for clarity with resolution, there is no method to measure the clarity of a 3D model yet.

Pixel density in units of PPI (Pixels Per Inch) is a commonly used metric for evaluating image clarity. Since the 3D model is rendered as a 2D image for display on the screen, we expect to evaluate the resolution of 3D models using PPI as image. To make it easy to grasp and generalize in 3D models, we propose the PPMM by replacing inches with the more commonly used length unit of millimeters (MM). A feasible way to measure the PPMM of a 3D model is as follows:

• (1) Prepare the OBJ format file of the 3D model to be evaluated, which normally consists of one model file (.obj), one material library file (.mtl), and several texture image files (.png or .jpg).

• (2) Open one of the texture images with image processing software. Randomly select a texture block and draw a straight line $T_i$ within it, record the pixel number $N_i$ of $T_i$, and save the changed image.

• (3) Open the 3D model with a 3D viewer, find $T_i$, and measure its length, recording the length value in inches as $L_i$.

• (4) The texture resolution $R_i$ at the location of $T_i$ is calculated as the ratio of $N_i$ to $L_i$. Randomly repeated $n_3$ times, calculating the average value $R_t$ of all the $R_i$, then $R_t$ represent the texture resolution of the 3D model (Eq. 3).

$$\begin{aligned} R_t = \frac{1}{n_3} \sum _{i=1}^{n_3} \frac{N_{i}}{L_{i}} \end{aligned}$$

(3)

In practice, we use DasMesh (Step 1), Windows Paint (Step 2), DasViewer (Step 3), and Microsoft Excel (Step 4) to achieve the above-mentioned operations. We will soon develop a free and handy tool in DasViewer to provide the functions of Step 2, Step 3 and Step 4.

Normally, the geometric refinement and texture resolution are relatively uniform in a 3D model, and 3D models are often divided into multiple tiles uniformly to alleviate the memory consumption of the computer. Therefore, we can randomly select one or more tiles to calculate geometric refinement and texture resolution.

Treatment of unqualified 3D model. In case the evaluation results indicate that the spatial accuracy is not qualified, it will be necessary to re-run the modeling program from ATS-control.

In case the evaluation results indicate that the geometric refinement and/or the texture resolution is not qualified, it will be necessary to identify the local area of poor quality, and then conduct complementary data collection for this area. Combining the original data with complementary data and then re-running the modeling program will be effective in improving the quality. This method of Obtaining models—Evaluating quality—Collecting complementary data is the most effective means to ensure that there are no omissions in data collection.

Results

We practice and verify the workflow and methods by dynamically documenting the excavation of the #3 fossil of hominin cranium from Yunxian. This section shows the major results.

Multi-level control network

In the excavation of the #3 fossil of hominin cranium from Yunxian, we deploy the multi-level control network as follows:

• (1) Level 1: Painting L-shaped GCPs on the hardened road surface and surveying the geographic coordinates with RTK (Fig. 6).

• (2) Level 2: Using survey nails to set up GCPs and surveying its geographic coordinates with RTK (Fig. 7).

• (3) Level 3: Using marker board to set up GCPS and surveying its geographic coordinates with ETS from GCPs of L2 (Fig. 8).

• (4) Level 4: Using mini label marked with L to set up GCPs and surveying its geographic coordinates with ETS from GCPs of L3 (Fig. 9).

Figure 10 illustrates the distribution of the GCPs in the Level 1 and Level 2 control networks. Figure 11a illustrates the introduction of the Level 3 control network from the Level 2 control network, and Fig. 11b illustrates the distribution of the GCPs in the Level 3 control network. Figure 12 illustrates a periodic Level 4 control network.

Fig. 6

The appearance of GCPs at Level 1. a, b Show two different GCPs at Level 1 and scenarios using RTK to survey their geographic coordinates

Fig. 7

The appearance of GCPs at Level 2. a, b Show two different GCPs at Level 2

Fig. 8

The appearance of GCPs at Level 3. a, b Show two different GCPs at Level 3

Fig. 9

The appearance of GCPs at Level 4. a, b Show two different GCPs at Level 4

Fig. 10

The distribution of GCPs in Level 1 and Level 2. GCPs in Level 1 were relatively homogeneously distributed, and GCPs in Level 2 were appropriately clustered to the excavation areas

Fig. 11

Deploying the Level 3 control network based on the Level 2. a Shows that due to the absence of CORS signals, several GCPs in Level 3 were surveyed using ETS to set up stations on the GCPs in Level 2. Each of them is required once surveying and once checking to ensure the accuracy of the geographic coordinates. Then, KZ08 and KZ09 with known geographic coordinates can be used to survey and check the remaining GCPs in Level 3. b Shows that GCPs in Level 3 are relatively uniformly distributed; some of them are stable, and others are dynamically adjusted

Fig. 12

The periodic Level 4 control network. The GCPs in Level 4 are relatively uniformly distributed; some are stable, and others are dynamically adjusted. The geographic coordinates of the GCPs in Level 4 were surveyed by ETS at stations set up on the GCPs in Level 3

Under the standardized operation procedure, the error of the multi-level control network is the systematic error of RTK and ETS. If the measurements of the GCPs depend on the coordinates of the higher-level GCPs, the error will be accumulated. Table 2 shows the precision of the multi-level control network.

Table 2 The precision of the multi-level control network

Data collection

At the Yunxian Man site, we conducted 1 data collection for the geographical environment and the archaeological site because they were in a long-term stable state. During the excavation, we performed 6 data collections for the Excavation Area B and 14 data collections for the cultural remains.

Table 3 shows the Details of the data collection. The Area refers to the area of the effective coverage of the data. The Frequency refers to the times of data collection since the initiation of the current excavation in 2021. The Resolution refers to the spatial resolution, which is the most essential metric to describe the data quality.

Table 3 Details of the data collection

The spatial resolution of the point cloud equals to the geometric refinement described in “Evaluation metrics” section. The spatial resolution of an image means the actual size of the object represented in one pixel. Due to terrain relief, image tilt, and many other reasons, the spatial resolutions are not uniform and constant within the same data. Therefore, we estimated the spatial resolution and presented their orders of magnitude in Table 3.

Cross-scale data fusion 3D modeling

Before the excavation, we conducted a geological survey of the archaeological site and the geographic environment through aerial photogrammetry. Due to the difference in area and importance, we adopted different flight altitudes to improve the efficiency of data collection and optimize the cost of data processing. The flight altitude was 240 m for the geographic environment and 177 m for the archaeological site. For the Excavation Area C, we also used close-range photogrammetry to obtain more detailed data. Due to the large differences in the flight altitudes (or object distance) and camera parameters, we obtained data with different scales and resolutions. In this case, it is more difficult to obtain a unified 3D model using common methods. Thanks to the multi-level control network, we make these cross-scale data under an actual and unified coordinate framework, and then we use the method of cross-scale data fusion 3D modeling provided by GET3D to get a unified 3D model.

As a unified 3D model, it enables us to view and analyze the geographic environment, archaeological site, and excavation area just by mouse wheel zoom—without requiring us to switch to different 3D models. Figure 13 shows the 3D model of the Yunxian Man site and its geographical environment. In Fig. 14, the green lines of tile bounding indicate that it is a 3D model rather than a photograph, then sub-figure (a) shows that the 3D models of the geographic environment (outside the pink box) and the archaeological site (inside the pink box) achieve perfect fusion, sub-figure (b) shows that the 3D models of the archaeological site (outside the pink box) and the Excavation Area C (inside the pink box) achieve perfect fusion.

Fig. 13

The 3D model of the Yunxian Man site and its geographical environment. a Shows the high-precision geometric model. b Shows the high-realistic textured model

Fig. 14

Cross-scale models integrated into a unified framework. a That the 3D models of the geographic environment (outside the pink box) and the archaeological site (inside the pink box) achieve perfect fusion. b shows that the 3D models of the archaeological site (outside the pink box) and the Excavation Area C (inside the pink box) achieve perfect fusion. Through the wireframe mode, c, d show that the 3D model has a higher geometric refinement in the area of the archaeological site

Multi-source data fusion 3D modeling

In the excavation of the #3 fossil of hominin cranium from Yunxian, we documented each transient but significant state of the Excavation Area B and the cultural remains using both photogrammetry and LiDAR scanning. After each data collection, we use GET3D to fuse photogrammetric images and laser point clouds to obtain high-precision and high-realistic 3D models.

Figure 15 shows a point cloud acquired by the terrestrial laser scanner at one station. We can see that occlusion and other reasons result in significant data deficiencies (i.e., voids), and the laser-scanned data are naturally absent of texture images. In practice, we can try to fill in these voids using multi-station scanning, but it is difficult to eliminate them completely as shade is ubiquitous. Figures 16 and 17 shows that using multi-source data fusion, a geometrically complete and highly realistic textured 3D model was obtained.

Fig. 15

A point cloud acquired by the terrestrial laser scanner at one station. It shows that occlusion and other reasons result in significant data deficiencies (i.e., voids), and the laser-scanned data are naturally absent of texture images. a Shows a circle void caused by blind areas at the scanner stand location. b Shows missing data in the radial distribution caused by occlusion. c Shows resolution degradation due to laser beam divergence

Fig. 16

The 3D model of the Excavation Area B at the Yunxian Man site. a Shows the high-precision geometric model. b Shows the wireframe model. c Shows the high-realistic textured model

Fig. 17

Two partial zoom-in views of the 3D model of the Excavation Area B at the Yunxian Man site. a, b Show that the 3D model has a high geometric quality. c, d Show that the 3D model has different geometric refinement in different regions

Fast cloud-based 3D modeling

In the excavation of the #3 fossil of hominin cranium from Yunxian, we conducted 6 data collections using close-range photogrammetry for the Excavation Area B (Table 3). Since it is not possible to fly UAVs in the cramped indoor area, these close-range photogrammetry can only be taken by manual photography. The Excavation Area B is over 200 $\textrm{m}^2$; it is too large to prevent omissions for manual photography. To engage the method of Obtaining models—Evaluating quality—Collecting complementary data to prevent omissions, we used DasEarth to process all these close-range images and generated 3D models (Fig. 16).

In our experiment, the single GPU workstation is configured with an Intel Core CPU and a Nvidia RTX4090ti GPU. DasEarth called 1 Synology NAS and 3 workstations, the workstations were equally configured with an Intel Core CPU and a Nvidia RTX4090ti GPU. Thanks to the efficiency of cloud modeling, we don’t have to go through the dreaded long wait in the close-range photogrammetry for the Excavation Area B. We enumerate the runtime taken for each modeling with DasEarth in Table 4. For a single GPU workstation, we tested the runtime using the 6th modeling data (bold font in Table 4) and estimated the runtime for the other five modeling data (regular font in Table 4) based on an approximate linear relationship between the number of photos and the runtime.

Table 4 The number of images and the modeling times used for each update

Obviously, DasEarth is faster than a single GPU workstation. Moreover, DasEarth doesn’t need a powerful and expensive GPU workstation; it just needs a mediocre laptop with Internet access or even a smartphone to accomplish the task.

4D model fusion

Archaeologists identified the integrity and significance of the #3 fossil of hominin cranium from Yunxian when it was just revealed. Then, a fossil excavation sub-area was defined by extending the location of the fossil outwards to four trenches (the pink rectangle in Fig. 18a). The works for the fossil excavation sub-area were meticulous and resulted in a more frequent requirement for digital documentation. As a result, a total of 14 data collections were made from the time the fossil was exposed until it was fully excavated (Cultural Remains in Table 3).

We used the 4D model fusion method to fuse the 3D models of the fossil excavation sub-area with the entire excavation area. The fusion result is a set of 4D models (Fig. 18). Each temporal phase in the set is a complete model of the entire excavation area, which supports archaeologists performing stratigraphic research and spatial analyses for the cultural remains. In sub-figure (a), the pink rectangles mark the update area. In sub-figure (b), multiple 3D models of the update area are shown by time; the first row shows the geometric model, and the second row shows the 3D model with texture.

Fig. 18

A set of 4D models for the Excavation Area B. a Marks the fossil excavation sub-area in the Excavation Area B. b Shows the 4D models of the fossil excavation sub-area. More details will be updated on the project homepage (https://wyniu.github.io/DDAE/) in the future

Data quality

According to the evaluation metrics in “Evaluation metrics” section, we evaluated the 3D model quality of the excavation of the #3 fossil of hominin cranium from Yunxian. Table 5 presents all the evaluation results.

Table 5 The evaluation of data quality

Discussion

Multi-level control network

• (1) In our practice, we constructed 4D models for the excavation of the #3 fossil of hominin cranium from Yunxian. Thanks to the multi-level control network, the documented 3D models all carry actual geographic coordinates. It proves that the multi-level control network is capable of placing cross-scale archaeological objects in an actual and unified coordinate framework.

• (2) In our proposed multi-level control network, the GCPs in Level 1 and Level 2 are relatively fixed, while the GCPs in Level 3 and Level 4 could change dynamically. The stabilized GCPs ensure the accuracy of the overall multi-level control network. The flexible GCPs enable the multi-level control network to continuously support the dynamically documenting archaeological excavations.

Cross-scale data fusion 3D modeling

• (1) From the practice of fusing the data of the geographic environment, archaeological site, and excavation area together to construct a unified 3D model in “Cross-scale data fusion 3D modeling” section, we can conclude that (a) aerial photogrammetry enables fast data collection for large areas; (b) aerial photogrammetry with reduced altitude and close-range photogrammetry enables meticulous data collection to obtain adequate details; (c) together with multi-level control network, the method of cross-scale data fusion 3D modeling enables fuse cross-scale data to construct a unified 3D model.

• (2) From a practical perspective, the method of cross-scale data fusion 3D modeling allows us to select data scales according to how many details are required—so as to optimize labor, time, and cost of data collection and processing.

Multi-source data fusion 3D modeling

It can be seen from “Multi-source data fusion 3D modeling” section that our technique and software achieve the fusion modeling of laser-scanned point clouds and photogrammetric images. We can pick some details to analyze the advantages of multi-source data fusion 3D modeling.

• (1) Observing Fossil A in Fig. 19, the point cloud captured by the portable laser scanner is denser than the dense matched point cloud from photogrammetry. Naturally, the former also generates a more refined mesh than the latter. We fused the point clouds and then mapped the textures to obtain a 3D model with both high-precision geometry and highly realistic texture.

• (2) Observing Fossil B in Fig. 19, the point cloud obtained by the terrestrial laser scanner produces some missing points due to viewpoint occlusion. Obviously, it produces meshes with voids. We fused the point clouds, and then the voids were repaired.

Fig. 19

Data fusion of photogrammetry and laser scanning. a Shows that the laser-scanned data of fossil A is not textured, while the photogrammetric data lacks sufficient geometrical refinement, and the fusion method yields a geometrically refined and textured 3D model. b Shows the laser-scanned data has a void caused by shading, and the fusion method fills it up

Eventually, the multi-source data fusion 3D modeling method enables us to obtain 3D models that preserve the dual advantages of LiDAR scanning and photogrammetry.

Fast cloud-based 3D modeling

• (1) In the practice of “Fast cloud-based 3D modeling” section, when we use DasEarth to generate 3D models, we only require a mediocre laptop to upload data and stab GCPs, and the rest of the modeling work can be done entirely using the computility of the cloud. Therefore, the fast cloud-based 3D modeling technology helps us get rid of the dependence on local GPU workstations. As a result, archaeological teams do not need to finance purchasing GPU workstations, nor do they need to worry about the idle and wasteful use of expensive GPU workstations during times when they are not conducting field excavations.

• (2) Analyzing the technical theory of fast cloud-based 3D modeling in “Fast cloud-based 3D modeling” section, multiple GPU laborers working in parallel will certainly be faster than a single GPU workstation locally. Table 4 also confirms this assumption. Therefore, the fast cloud-based 3D modeling technology could save a lot of time and also nicely solve the problem of the phased surge of 3D modeling demand during archaeological excavations.

Overall, the fast cloud-based modeling technology offers multiple benefits, including financial savings, decreasing 3D modeling time, and making 3D modeling and sharing easier. This technology will be a true utility to facilitate dynamically documenting archaeological excavation, and a true platform to promote collaborative research through data sharing.

4D model fusion

As the practice of fusing the 3D models of the frequently changing excavation sub-area and the entire excavation area in “4D model fusion” section, the method of 4D model fusion enables us to document the dynamic process of change in the excavation area so as to support the sequential analysis along time.

To better analyze the value of the methods in this paper, we provide an additional case study of the Zhengjiahu Cemetery [34] in the “Appendix”. To fulfill the requirement of comparatively analyzing the three tombs in the additional case study, we need to simultaneously present them in their just-opened state. However, the just-opened states of the three tombs were separated by several months. Thanks to the 4D model fusion method, we fused 3D models of the three tombs in their just-opened state into a unified 3D model. Therefore, the 4D model fusion method enables us to synthesize transient states of several excavation sub-areas into virtual scenes so as to support comparative analysis across time.

From a practical perspective, the method of 4D model fusion solves much tedious and time-consuming 3D modeling work by allowing us to simply focus on the sub-area that is being excavated—Otherwise, to get a complete model, we would have to conduct a data collection and modeling of the excavation area B of 200 $\textrm{m}^2$ once whenever the fossil excavation sub-area of 4 $\textrm{m}^2$ changes.

Data quality

The planimetric precision $E_p$ and the elevation precision $E_e$ in Table 5 are commonly better than 0.03 m, even may be better than 0.01 m, and reach the millimeter level. Such evaluation results indicate that the 3D models carry geographical coordinates with ideal spatial accuracy. Moreover, since the algorithms of 3D measurement are based on geographic coordinates, the ideal spatial accuracy means we can perform reliable measurements for position, distance, area, and volume on these 3D models.

The geometric refinement $R_g$ and texture resolution $R_t$ in Table 5 are gradually enhanced from geographical environment to cultural remains. Indeed, our data collection is progressively finer from large to small scale. It indicates that our proposed geometric refinement and texture resolution are effective measures for the data quality.

Moreover, the data in Table 5 show that our methods enable high-precision modeling with geometric refinement reaching the sub-millimeter level and texture resolution of more than 10 PPMM.

Conclusions

Technological advancements promoted greater demands on using 3D modeling techniques to document archaeological excavations than ever before. The concept of documenting various momentary states for an ongoing archaeological excavation using 4D models can be formulated as dynamically documenting. Due to the practical difficulties of implementing dynamically documenting, we proposed a complete workflow and pursued several new methods, and practiced them in the excavation of the #3 fossil of hominin cranium from Yunxian. Our practical results show that:

• (1) The multi-level control network enables all documented 3D models to be in an actual and unified coordinate framework.

• (2) The proposed methods are capable of fusing data collected from different scales, equipment, and periods.

• (3) The fast cloud-based modeling technology serves dual benefits. On the one hand, it aids archaeological teams in economically underdeveloped regions by alleviating their financial burdens, allowing them to engage in dynamically documenting without the concern of purchasing a sufficient number of GPU workstations. On the other hand, it offers all archaeologists and cultural heritage protectors a more efficient and user-friendly modeling service than several local GPU workstations.

• (4) Through the comprehensive evaluation metrics, the documented 3D models can be precisely evaluated.

Our workflow and methods are capable to perform dynamically documenting archaeological excavation and obtain cross-scaled 4D models. Nevertheless, improvements still need to be incorporated into our technology. For example, images obtained from close-range photogrammetry inevitably contain some out-of-focus areas, which occasionally cause the texture of 3D models to be unclear. We will improve our technique by studying the optimal texture selection strategy.

In general, the proposed workflow and methods are proven effective and efficient since they eliminate tedious and time-consuming procedures. In addition, the entire process is easy to implement and generalize since the equipment and software we adopted are user-friendly. As a result, the workflow and methods enable us to obtain unified and 4D models with high-precision geometry and highly realistic texture. The 3D views enhance archaeologists’ spatial perception of sites and relics, and the high-quality digital documentations support collaborative and long-term research based on data sharing. Therefore, the widespread adoption of the workflow and methods is promising to be conducive to archaeological fieldwork and in-depth research.

Appendix: Additional Case Study of the Zhengjiahu Cemetery

A brief introduction to the site. The Zhengjiahu Cemetery is located in Yunmeng County, Hubei Province, China, which is an important Warring States through Qin and Han cemetery. The Excavation Area C of the Zhengjiahu Cemetery was excavated in 2021. There, two Qin culture tombs, Nos. 274 and 234, yielded wooden gu slip with long texts and painted burial accouterments, respectively. The wooden gu slip, rarely in form and rich in content, is of significant value for the study of social history and ideology during the late Warring States. In addition, discovered painted burial accouterments fill a gap in current understandings of Qin and Han painting materials and types, providing crucial data for the study of Qin descendants’ burial customs and rites, as well as religious ideology.

3D technology to document the excavation. We documented many transient but significant states of the Excavation Area C using the methods described in this paper. The significance of these digital documenting works lies in the fastest speed and highest resolution to capture and document the real-life spatial information of the fragile organic archaeological remains, thus avoiding the omission of information that may be caused by the rapid oxidation and disappearance of cultural relics and the slow speed of manual documenting, and providing the possibility of future in-depth research, protection, and display.

The application of 4D model fusion. There are three adjacent tombs, M274, M275, and M276, in the Zhengjiahu Cemetery (Fig. 20). As the cultural relics in the tombs are extremely susceptible to oxidation and discoloration, the three tombs could only be excavated one after the other—so as to ensure that there were sufficient technicians available to protect the unearthed relics in time. Therefore, the coffins of the three tombs were opened at different times. After each coffin was opened, we used a UAV to conduct close-range photogrammetry as fast as possible, and then the archaeologists immediately transferred the relics. All of these tasks are usually accomplished within a few hours, and the appearance of the coffin just being opened is no longer in existence. However, archaeologists want to research the three tombs and analyze the relative orientation of their cultural relics as they were just being opened. Therefore, we employed the 4D model fusion method and fused the 3D models of three tombs constructed at different times into a unified 3D model. The fused 3D model is shown in Fig. 21, and it satisfied the needs of archaeologists in later studies.

Fig. 20

The tombs ichnography in the Excavation Area C of the Zhengjiahu Cemetery. The red polygons label the objects studied in this paper, M274, M275, and M276

Fig. 21

The result of 4D model fusion. This figure shows the M274, M275, and M276 appearing to be realistically excavated together. However, the scene never really existed. Each tomb of them only existed for a few hours as they appear on the figure, and they were excavated several months apart from each other. Such a synthesized virtual scene is useful for archaeological research and spatial analysis

This case demonstrates that the method of 4D model fusion enables us to fuse the spatial 3D models obtained at different excavation seasons and in different excavation areas into a complete, virtual 3D model, to obtain the complete information that can’t be obtained from the excavations in different areas, and to discover the spatial features that can’t be found from normal observations.