2.1 True Orthoimage Production Workflow

True orthoimage generation differs from conventional orthoimage production in that it requires two additional processes: the construction of 3D precision digital elevation data and the processing of occlusion areas. First, surface data incorporating structures such as buildings must be constructed and used for image orthorectification. Subsequently, occlusion areas arising from central projection must be appropriately removed or replaced.

Fig. 1. True Orthoimage Production Workflow

Occlusion areas manifest as regions where the rear faces of structures are obscured due to the imaging geometry, and if these areas are not properly handled, double-mapping artifacts occur. Therefore, the detection and correction of occlusion areas in true orthoimage production directly affect the visual completeness and interpretability of the final product.

Fig. 2. Occlusion Area and Double Mapping

2.2 Characteristics of Digital Building Model Generation Methods

2.2.1 3D Stereo-Plotting Method

The conventional stereo-plotting method assumes 3D depiction; however, inconsistencies in results may arise due to differences in interpretation and representation among operators during the plotting process. For example, when two or more buildings with different heights are situated adjacent to one another, the conventional method—which focuses on plotting building outlines—results in discontinuous height values at the corners of individual buildings. This can produce logically inconsistent linear structures in the 3D plotted data rather than accurately representing building rooftop surfaces, and such outputs are unsuitable for true orthoimage production.

For true orthoimage production, therefore, it is necessary to apply the 3D precision stereo-plotting method, which defines objects based on building rooftop surfaces. Under this method, the number of points and lines to be depicted increases in accordance with the Level of Detail (LoD) criteria, and working time and cost increase proportionally. Furthermore, a higher degree of image overlap than that required for conventional orthoimage production must be secured. In the 3D stereo-plotting process, building rooftop surfaces and individual corners must be precisely depicted, and surfaces sharing the same height value must be represented as a single object unit. For facilities with overlapping upper and lower structures, such as overpasses, object separation based on structural hierarchy is required.

Fig. 3. Comparison of Conventional and 3D Precision Stereo-Plotting

2.3 Software Characteristics for True Orthoimage Generation

Various commercial software packages can be used for true orthoimage generation. This study compared Bentley iTwin Capture Modeler, nFrame SURE, and Trimble Inpho. Trimble Inpho provides precise orthorectification and editing capabilities well-suited to conventional photogrammetry-based workflows, while iTwin Capture Modeler and nFrame SURE offer strengths in automated matching and mesh generation from high-overlap imagery. Differences among the software packages are observed in aerial triangulation integration, mesh generation quality, boundary correction functionality, and post-processing convenience, and these differences affect the production efficiency and editing workload of the final true orthoimage.

3.1 Data Acquisition and Aerial Triangulation

In this study, aerial imagery was acquired with 80 %×80 % overlap in a single-pass configuration for true orthoimage production, securing sufficient image overlap for DIM-based automated processing. The cameras employed were the second-generation aerial cameras Falcon (378 images) and DMC2 (150 images), and the third-generation aerial cameras DMC3 (98 images) and Osprey (120 images). For aerial triangulation, 49 ground control points were distributed across the study area. The number of points used as check points varied by camera depending on point identifiability in each sensor’s imagery, ranging from 42 to 47 (see Table 2).

Fig. 4. The Study Area for Aerial Triangulation

Aerial triangulation was performed using Match-AT for imagery from each camera. Quality verification showed that all cameras achieved planimetric RMSE values below 0.03 m and vertical RMSE values below 0.04 m, confirming stable accuracy. In particular, the DMC3 achieved the lowest planimetric RMSE of 0.006 m and a vertical RMSE of 0.014 m, verifying that the ground control and orientation results were sufficiently reliable.

3.2 True Orthoimage Production for Comparison

For relative comparison, true orthoimages were produced using both the conventional 3D stereo-plotting method and the DIM mesh-structure method. In the conventional approach, building model construction and orthorectification were performed using Geo3Di and Trimble Inpho, while the automated approach utilized Bentley iTwin Capture Modeler and nFrame SURE for image matching, mesh generation, and orthoimage generation.

Processing the full workflow for 378 Falcon images using iTwin Capture Modeler version 23 required approximately 24 hours, while nFrame SURE-based processing on a single-node system required approximately 28 hours. These results demonstrate that the DIM mesh-based approach possesses an automated character well-suited to large-scale data processing.

True orthoimages generated using the DIM mesh-structure method tend to exhibit poorly defined building boundary lines due to the interpolation inherent in the image matching process. Consequently, boundary processing issues arise for tall buildings, areas with pronounced terrain undulations, or the presence of vegetation.

Post-processing-based editing and correction of the initially generated true orthoimages is feasible even without separately constructing 3D precision digital elevation data. By comprehensively incorporating the orientation elements and positional information utilized in the orthoimage production process, boundary errors can be corrected within a relatively short time. Fig. 8 presents the correction results obtained by applying image editing to the identified boundary errors.

Fig. 5. True Orthoimages Generated by 3D Stereo-Plotting: Falcon Imagery (Top Row) and DMC3 Imagery (Bottom Row)

Fig. 6. True Orthoimages Generated Using the DIM Mesh-Based Approach from DMC2 (First Row), Falcon (Second Row), DMC3 (Third Row), and Osprey (Fourth Row)

4.1 Assessment of Building Relief Displacement Positional Accuracy

To quantitatively evaluate the quality of the initial true orthoimages prior to editing, planimetric positional errors were analyzed based on 10 check points established on building rooftops. Qualitative quality was also examined through overlay with 1/1,000 digital maps.

The results showed that Falcon and DMC2 achieved favorable accuracy levels with mean errors of approximately 0.15 m, while Osprey recorded a mean error of 0.189 m and DMC3 recorded 0.359 m. The elevated mean error for DMC3 is primarily attributed to a single outlier of 1.267 m at one check point. This outlier corresponds to a tall building located in a complex rooftop area where mesh smoothing during DIM matching displaced the rooftop edge; excluding this point, the DMC3 mean error reduces to approximately 0.258 m, comparable to the Osprey result. With this caveat, the DIM mesh-based approach was assessed to have secured the positional accuracy required for practical application. It should be noted that this assessment is based on the rooftop check points used for the DIM mesh-based products and does not constitute a strict head-to-head comparison with the conventional 3D stereo-plotting approach, which would require identical check points, the same target area, and the same evaluation procedure to be applied to both methods. Such a controlled cross-method comparison is identified as a follow-up task.

To support the interpretation of the sensor-specific accuracy differences, an exploratory statistical review was performed on the n = 10 check-point measurements per camera. Using a Student’s t-distribution with nine degrees of freedom ($t_{0.025} = 2.262$), the 95 % confidence intervals for the mean planar error were approximately [0.02, 0.29] m for Falcon, [0.03, 0.27] m for DMC2, [0.02, 0.36] m for Osprey, and [0.00, 0.73] m for DMC3. The DMC2, Falcon, and Osprey intervals overlap substantially, indicating that their mean accuracies are not statistically distinguishable at this sample size. The wider DMC3 interval is driven by a single outlier (1.267 m); when this point is excluded, the DMC3 mean error reduces to approximately 0.258 m with a standard deviation of 0.252 m, placing it on the same order of magnitude as the other sensors. These results suggest that the DIM mesh-based pipeline is robust across the tested cameras, while also indicating that the present sample size of ten is sufficient for technical evaluation but limited for population-level inference; expanded validation is identified as future work in Section 5.

Fig. 7. Boundary Delineation Errors in True Orthoimages Generated Using the DIM Mesh-Based Approach

Fig. 8. Results of Boundary Editing for true Orthoimages Generated Using the DIM Mesh-Based Approach

4.2 Visual Quality Comparison by Area Type

A comparison of the 3D stereo-plotting method and the DIM mesh-structure method across three area types—industrial, apartment residential, and detached residential zones—revealed that visual differences between the two methods were minimal in the industrial zone, where building forms are relatively simple. In the apartment zone, the overall level of relief displacement correction was also satisfactory. This comparison was conducted using Falcon imagery (second-generation) and DMC3 imagery (third-generation); Trimble Inpho was used for the 3D stereo-plotting method, and Bentley iTwin Capture Modeler for the DIM mesh-structure method.

In the industrial zone, DIM mesh-based results demonstrated generally adequate correction of building rooftop relief displacement. Minor differences in boundary representation were observed in some buildings with irregular rooftop geometries or significant height differences between adjacent structures. In the apartment zone, no significant difference between the two methods was found in the level of relief displacement correction for high-rise buildings.

In contrast, in the detached residential zone—where building forms are complex and vegetation is present nearby—the DIM mesh-based products exhibited somewhat insufficient representation of fine building boundary details, associated with the smoothing of boundary discontinuities during automated matching. Quantitatively, this manifested as residual planar offsets at building edges on the order of 0.2-0.4 m, exceeding the 0.15 m mean error observed for simpler industrial-zone buildings, and is consistent with the elevated DMC3 mean error reported in Section 4.1. Nevertheless, the overall level of orthorectification and spatial interpretability was maintained at a level comparable to the conventional method. In particular, errors in 3D precision digital elevation data attributable to complex rooftop geometries and vegetation effects were found to cause ambiguous building boundary delineation, which was more pronounced in cases of greater building heights or larger elevation differences with adjacent vegetation, but correctable through software post-processing functions. Overall, the DIM mesh-structure method demonstrated variability in output completeness depending on area type, while proving to be a practical alternative in terms of workflow automation.

Fig. 9. Comparison of True Orthoimages: 3D Stereo-Plotting (Left) vs. DIM Mesh-Based Approach (Right): (a),(b) Industrial Zone; (c),(d) Apartment Zone; (e),(f) Detached Residential Zone

4.3 Comparison of Error Correction Methods

The methods for correcting errors in the initial true orthoimages generated using the DIM mesh-based approach can be broadly categorized into two types: editing the 3D precision digital elevation data, and directly editing the generated true orthoimages.

The 3D precision digital elevation data editing approach involves correcting the surface model itself using auxiliary data such as boundary line shapefiles and LiDAR, and then regenerating the true orthoimage. This approach offers the advantage of concurrently securing a corrected 3D model and ensuring high structural consistency; however, it requires additional data and expert personnel and entails a substantial re-processing burden. Positional accuracy verification can be performed based on the corrected surface model. Representative software packages supporting this approach include nFrame SURE, Pix4D Mapper, and Agisoft Metashape.

The direct true orthoimage editing approach involves correcting the initial output using image editing functions, requiring fewer additional data inputs and offering greater operational flexibility. It was found to be comparatively more efficient in large-scale operational systems and more advantageous for rapid product correction. In this study, the Retouching function of Bentley iTwin Capture Modeler was used to perform editing tasks including the removal of moving objects, the refinement of building rooftop boundary lines, and the filling of reflective or unmapped areas. In the detached residential zone, the editing process required approximately 5 to 10 minutes per building, substantially lower than the approximately 17 person-days per 1:5,000 map sheet required by the 3D stereo-plotting method. To enable a like-for-like comparison, all productivity figures were converted to a common per-sheet basis. The 24- to 28-hour automated processing time corresponds to approximately one day of machine time per coverage block, the bulk of which proceeds unattended. Assuming selective editing of approximately 100 priority structures (high-rise, landmark, and publicly significant buildings) per 1:5,000 sheet at the observed 5-10 min per building, the resulting human effort amounts to roughly 1.5 to 2 person-days per sheet. The DIM mesh-based approach therefore reduces the per-sheet human effort to approximately 3-4 person-days when accounting for both processing supervision and selective editing, compared with 17 person-days for the 3D stereo-plotting method, representing a productivity improvement of roughly four- to fivefold per sheet. It is therefore considered practically reasonable to apply selective editing focused on buildings with high public significance, landmark buildings, and high-rise structures.

Fig. 10. Comparison of Error Correction Workflows: (Top) 3D Precision Digital Elevation Data Editing Approach; (Bottom) True Orthoimage Direct Editing Approach

4.4 Redefinition of Quality Inspection Items

The DIM mesh-structure method is based on automated matching and surface reconstruction using high-overlap imagery, making it difficult to directly apply the existing quality inspection framework. Items that were critical for conventional orthoimages—such as the use of nadir imagery, image connectivity, color and brightness consistency, and digital elevation model inspection—may be of relatively lower importance in the DIM-based automated environment.

Conversely, the conformance of building boundary lines and the adequacy of occlusion area processing emerge as core inspection items governing the quality of DIM mesh-based products. The quality inspection framework should therefore be reconstructed not through simple maintenance of existing items, but through selective reduction and the addition of new items reflecting differences in the generation method. Beyond the qualitative inclusion/exclusion judgments summarized in Table 4, the redefined inspection items also call for quantitative tolerance thresholds. Drawing on the rooftop accuracy results obtained in this study (mean planar errors of 0.148–0.359 m at the building scale), preliminary tolerance candidates for two newly emphasized inspection items can be considered: (i) for boundary smoothing arising during DIM mesh matching, a planar offset tolerance on the order of one to two times the ground sampling distance for the relevant scale (e.g., approximately 0.10-0.25 m for 1:1,000 to 1:5,000 mapping) is plausible; and (ii) for occlusion area handling, the proportion of unfilled or doubly mapped pixels relative to the total building footprint area provides a measurable indicator. The exact threshold values must be calibrated against larger samples and against the specific application requirements of each product, and their formal specification is identified as a follow-up study task.