PaperlayerView exact revision on arXiv

Source-linked AI summary

A systematic review and meta-analysis of Digital Elevation Model (DEM) fusion: pre-processing, methods and applications

Chukwuma Okolie, Julian Smit

arXiv:2203.15026v2cs.CVeess.IV

TL;DR

Differences among DEM products make fusion useful, yet the subject has lacked a comprehensive review. This paper systematically reviews DEM fusion and its applications; it reports that variational models can improve accuracy beyond classic weighted averaging, with additional computational expense and slower processing.

  • Problem

    DEM products differ in coverage, resolution, and accuracy, while DEM fusion has lacked a comprehensive review despite developed methods and heterogeneous studies.

    Source: §1 · pp. 3–4

  • Method

    The paper systematically reviews DEM-fusion preprocessing, methods, applications, and challenges, including comparative analysis of methods and their performance.

    Source: §1 · p. 4

  • Results

    Variational models can improve fusion accuracy beyond classic weighted averaging, at the cost of additional computational expense and slower processing.

    Source: §4.4 · p. 40

  • Takeaways & Limitations

    Weighted averaging remains preferred for low computational expense and fast processing despite drawbacks, while variational models trade those costs for potential accuracy gains.

    Source: §4.4 · p. 40

    Up-sampling from an input DEM can introduce artefacts into the fused DEM.

    Source: §4.3.2 · p. 26

Abstract

from arXiv · show

The remote sensing community has identified data fusion as one of the key challenging topics of the 21st century. The subject of image fusion in two-dimensional (2D) space has been covered in several published reviews. However, the special case of 2.5D/3D Digital Elevation Model (DEM) fusion has not been addressed till date. DEM fusion is a key application of data fusion in remote sensing. It takes advantage of the complementary characteristics of multi-source DEMs to deliver a more complete, accurate and reliable elevation dataset. Although several methods for fusing DEMs have been developed, the absence of a well-rounded review has limited their proliferation among researchers and end-users. It is often required to combine knowledge from multiple studies to inform a holistic perspective and guide further research. In response, this paper provides a systematic review of DEM fusion: the pre-processing workflow, methods and applications, enhanced with a meta-analysis. Through the discussion and comparative analysis, unresolved challenges and open issues were identified, and future directions for research were proposed. This review is a timely solution and an invaluable source of information for researchers within the fields of remote sensing and spatial information science, and the data fusion community at large.

1. Introduction

DEMs support topographic analysis and many geospatial applications, while differences among available products motivate fusion. This review addresses the limited coverage of DEM fusion in prior reviews by synthesizing its workflow, methods, and applications.

1. Introduction

  • DEMs underpin spatial data infrastructure and support mapping, orthophoto generation, and 3D analysis across fields including hydrology, environmental modelling, urban planning, and geology.Their use also spans topographic analysis and digital relief maps.Source: §1 · p. 2
  • Differences in DEM coverage, resolution, accuracy, surface model, error distribution, datum, and trends motivate fusion to combine complementary source characteristics.The paper notes that products are available from different remote-sensing platforms, sensors, and reconstruction algorithms.Source: §1 · p. 3
  • Prior reviews have focused substantially on 2D image fusion, while DEM fusion has received little attention and lacks a comprehensive review despite heterogeneous studies.The paper defines DEM fusion as algorithmic integration of at least two source DEMs to generate a more detailed DEM; quality depends on application.Source: §1 · pp. 3–4
  • The review examines fusion methods and their strengths and weaknesses, preprocessing, DEM properties, terrain parameters, accuracy assessment, applications, and future research directions.It also describes the general DEM-fusion approach and workflow.Source: §1 · pp. 4–5

2. General Workflow of DEM Fusion

The review organizes DEM fusion as a workflow from source selection through preprocessing and fusion to post-processing and quality assessment. Choices at each stage depend on data characteristics and the intended use.

2. General Workflow of DEM Fusion

  • DEM fusion is an ill-posed problem, and its goal is a continuous, contiguous, complete terrain representation across applications such as merging technologies or updating DEMs.Fusion may also combine DEMs with different acquisition geometries or use auxiliary elevation data to improve accuracy.Source: §2 · p. 5

2.1 DEM selection

  • Source selection depends on fusion purpose, data availability and quality, sensor characteristics, and suitable algorithms, with complementary sources potentially compensating for sensor limitations.Finance, access restrictions, software, computing and storage resources, data quality, and latency can constrain selection.Source: §2.1 · pp. 6–7

2.2 DEM pre-processing

  • Preprocessing addresses source-DEM issues including resolution disparities, registration errors, invalid pixels, vertical discrepancies, horizontal displacement, and inconsistent datums, coordinate systems, or formats.Common operations include accuracy improvement, void filling, co-registration, resampling, and generating height-error or weight maps.Source: §2.2 · p. 7

2.3 DEM fusion

  • Many pixel-level fusion methods use height-error or weight maps to determine each source DEM’s influence on the fused elevation.The review places the actual fusion after preprocessing in its generalized procedure.Source: §2 · p. 6, §2.3 · pp. 7–8

2.4 DEM post-processing

  • Post-processing can smooth the fused surface or blend transition areas using trend surfaces, denoising algorithms, spatial filters, or point-reduction filters.The cited examples include adaptive mean, Gaussian, and high-pass filters.Source: §2.4 · p. 8

2.5 Fusion quality assessment

  • Fused-DEM quality is assessed through visual inspection, statistical comparisons with source or reference DEMs, or performance in an application.Application-based assessment can examine outcomes such as surface-flow modelling, land-cover mapping, or watershed delineation rather than per-point accuracy statistics.Source: §2.5 · p. 8

3. Review Methodology

The review follows PRISMA 2020, screens literature against defined eligibility criteria, and synthesizes extracted study characteristics. Searches across scholarly databases and supplementary sources yielded 45 studies for review.

3. Review Methodology

  • PRISMA 2020 structures the review into identification, screening, and inclusion stages.The flow diagram reports 48,331 identified studies: 46,153 from databases and 2,178 from other methods; 45 studies were selected.Source: §3 · p. 8, Figure 2 · p. 9

3.1.1 Inclusions

  • Eligible studies used theoretically grounded methods to fuse raster, multi-source or multi-sensor DEMs, including real-world experiments and quality assessments.The criteria also included fusion along blend or transition zones and advanced geostatistical methods developed for fusion.Source: §3.1.1 · p. 10

3.1.2 Exclusions

  • Exclusions covered insufficiently rigorous or abstract studies, non-ground or non-DEM fusion topics, and studies outside the review’s stated DEM-fusion scope.The excluded topics included void filling, traditional tile mosaicking, extra-terrestrial data, bathymetric fusion, and DEM representation combinations without actual fusion.Source: §3.1.2 · pp. 10–11

3.2 Identification of studies

  • The search combined title, abstract, and keyword queries in Scopus, Web of Science, and IEEE Xplore with citation searching and a Connected Papers website search.Searches had no publication-period limit and used “Digital Elevation Model” and “Fusion” as main keywords.Source: §3.2 · p. 11

3.3 Article screening

  • Double screening was used to help minimize screening bias and resolve inclusion or exclusion conflicts through discussion and consensus.After duplicate removal, titles and abstracts of 32,163 records were screened manually.Source: §3.3 · pp. 12–13

3.4 Data extraction and analysis of included studies

  • Screening yielded 54 eligible reports, representing 45 studies because some studies appeared in two reports.The included studies comprised 28 journal papers and 17 conference papers.Source: §3.4 · p. 13
  • Extracted study characteristics included fusion methods, preprocessing, source and reference DEM properties, production technologies, and study-area land cover and terrain class.The extracted data were grouped into categories and synthesized for presentation.Source: §3.4 · p. 13

4.1 Characteristics of reviewed studies

The reviewed studies prominently feature TanDEM-X and weight-map and variational fusion methods. Publication activity was irregular, and study sites were concentrated in the global north.

4.1 Characteristics of reviewed studies

  • TanDEM-X was the most frequently occurring DEM, while InSAR was the most common production technology and weight-map and variational methods were prominent.SRTM, ASTER Global DEM, and Cartosat-1 were also prominent DEMs.Source: §4.1 · p. 13
  • 53% of studies were published between 2014 and 2018, although the publication trend showed no steady growth.The publication record was described as irregular.Source: §4.1 · p. 14
  • Study sites were concentrated in the global north, with hardly any studies from South America and Africa.The United States had 11 sites, China 7, Germany and Switzerland 5 each, and Spain 4.Source: §4.1 · p. 15
  • Sites were distributed across Asia (27%), Europe (36%), and North America (25%).Africa and Australia each accounted for 4% of sites, and South America for 2%.Source: §4.1 · p. 16

4.2 DEM pre-processing workflow

DEM pre-processing is tailored to source quality and characteristics, addressing discrepancies, voids, surface offsets, and alignment before fusion. The review also identifies regionalisation modelling and commonly used operations.

4.2 DEM pre-processing workflow

  • Pre-processing steps depend on source DEM quality and characteristics, and the workflow includes correction, alignment, and harmonisation before fusion.Alignment includes resampling and co-registration to achieve coincident pixels.Source: §4.2 · pp. 16–17

4.2.1 Initial comparison

  • Before fusion, discrepancies between source DEMs can be assessed through statistical analysis and DEM differencing.The DEMs provide separate perspectives of the same environment and can deviate from the true surface.Source: §4.2.1 · p. 17

4.2.2.1 Void filling

  • Voids are systematic errors that reduce DEM quality, completeness, and utility; they can be filled using auxiliary data and methods such as interpolation or Delta surface fill.Before filling, differences between the DEM and auxiliary data—such as datum, resolution, or error patterns—need to be resolved.Source: §4.2.2.1 · pp. 17–18

4.2.2.2 Filtering of non-ground points

  • Removing tree heights and other above-ground objects can help produce bare-earth elevations, a requirement for applications such as hydrological modelling.Approaches include tree-height estimates, ancillary data, stochastic methods, and mathematical morphology; the step is optional and was used by few reviewed studies.Source: §4.2.2.2 · p. 18

4.2.2.3 Vertical accuracy improvement

  • Source DEM vertical accuracy can be improved through difference surfaces, outlier filtering, and modelling errors against land and topographic parameters.The cited parameters include land use/land cover, slope, and aspect.Source: §4.2.2.3 · p. 19

4.2.2.4 Spatial filtering/smoothing

  • Spatial filtering can remove high-frequency artefacts with low-pass filters or low-frequency artefacts with high-pass filters.Reviewed approaches include two-step mean filtering and adaptive multi-scale smoothing.Source: §4.2.2.4 · pp. 19–20

4.2.3 DEM alignment; 4.2.3.3 Co-registration

  • Alignment includes resampling DEMs to a common resolution and harmonising coordinate and height systems; co-registration estimates and corrects vertical and horizontal offsets.Resizing a coarser grid to a finer one may introduce additional errors; co-registration aligns source data pixel by pixel to the same ground locations.Source: §4.2.3.1 · p. 20, §4.2.3.2 · p. 20, §4.2.3.3 · pp. 20–21

4.2.4 Regionalisation modelling

  • Regionalisation links environmental variables to height-error patterns or residuals, supporting height-error maps and predicted fusion weight maps.The most applied pre-processing operations were resampling, co-registration, void filling, smoothing/filtering, and harmonising spatial reference systems.Source: §4.2.4 · p. 21

4.3 Review of DEM fusion methods

This section reviews DEM-fusion methods, covering weighted averaging and variational methods alongside alternative strategies such as clustering, filtering, rules, weighting, and outlier handling. It surveys their mechanisms, applications, and limitations, including reported improvements in accuracy, completeness, and terrain detail that vary by method and evaluation measure.

4.3 Review of DEM fusion methods

  • Weighted averaging is the most frequently used DEM fusion method, followed by variational methods.The review presents the methods’ occurrence in the reviewed studies and introduces detailed discussions of individual approaches.Source: §4.3 · p. 22, Figure 8 · p. 24

4.3.1 Simple averaging

  • Simple averaging combines coincident input elevations cell by cell, with other locations filled from whichever DEM has data; one application reported an RMSE of 9.7 m.The cited glacier application averaged elevations only where both DEMs were available and their absolute difference was ≤ 10 m.Source: §4.3.1 · p. 25

4.3.2 Weighted averaging

  • Weighted averaging is simple and computationally inexpensive, but its performance depends on accurate weights and it can blur urban detail or introduce upsampling artefacts.In cited examples, fusion reduced RMSE by 2–4 m with 84% fewer voids; ANN-predicted weights improved relative accuracy by up to 50% in urban areas and 22% in non-urban areas.Source: §4.3.2 · pp. 25–27

4.3.3 Variational models

  • Variational fusion can outperform weighted averaging in noise reduction, while TV-L1 preserves edges but may create staircasing artefacts.A weighted TV-L1 version further enhanced fused DEM quality compared with unweighted TV-L1.Source: §4.3.3 · pp. 30–31

4.3.4 Super-resolution (SR) methods

  • Regularised super-resolution combines resolution enhancement, void filling, and noise suppression, yielding more comprehensive terrain detail and seamless fusion in reported experiments.The cited study used low-resolution data as foundational information and part of the high-resolution data as a constraint.Source: §4.3.4 · p. 31

4.3.5 Kalman filter

  • A multi-scale Kalman filter fusion reduced global MSE by 88–90% for bare-surface heights and 66–87% for vegetation heights.Later applications reported lower height uncertainty than individual InSAR DEMs, but an extended Kalman filter result remained deficient on steep front slopes because of layover.Source: §4.3.5 · p. 32

4.3.6 Multiple-point geostatistical simulation

  • A feature-based FILTERSIM variant fused coarse GMTED2010 with sparse fine-resolution SRTM data, improving accuracy and preservation of spatial structures.It derived terrain features from the DEM residual surface, vector ruggedness measure, and ridge-valley class.Source: §4.3.6 · p. 33

4.3.7 Multi-scale decomposition

  • Among four tested decomposition methods, N-AMD performed best, detecting DEM noise while maintaining trend consistency; the fused surfaces showed clear detail and no mutation regions.The study applied slope position-based linear regression to enhance the fused DEMs.Source: §4.3.7 · p. 34

4.3.8 Other methods

  • Bayesian fusion influenced the resulting DEM more than maximum-likelihood fusion and smoothed generated structures, but its discrepancies against GNSS checkpoints were higher.Sadeq et al. (2016) used prior knowledge to fuse 1-m Worldview-1 and Pleaides DEMs over parts of Glasgow; Jiang et al. (2014) used phase-noise variance and height of ambiguity to determine fusion weights.Source: §4.3.8 · p. 36
  • Frequency-domain fusion reduced RMSE to 16.4m from 18.3m for SRTM and 28.3m for ASTER, a 42% reduction versus ASTER and 10% improvement over SRTM.Karkee et al. (2008) co-registered and void-filled the DEMs, then filtered and summed their spectra before returning to the spatial domain; the fused DEM also improved accuracy and completeness.Source: §4.3.8 · p. 36, Figure 11 · p. 37
  • Self-consistency fusion identifies and excludes unreliable elevation outliers before combining reliable points, significantly reducing residual errors without apparent anomalies.Schultz et al. (1999) introduced this DEM-fusion application, which was also used in Schultz et al. (2002) and Papasaika et al. (2011a,b).Source: §4.3.8 · pp. 37–38
  • K-means fusion reduced SD, MAE, and RMSE from 5.9m, 6.6m, and 8.0m with simple averaging to 5.2m, 5.3m, and 6.4m, respectively.Fuss et al. (2016) combined slope and elevation thresholding, pixel-level clustering, filtering, and smoothing.Source: §4.3.8 · p. 38
  • Combining maximum-height and maximum-correlation rules yielded ~10% lower area error, >1m lower mean height error, and improved vertical accuracy; preferential weighted fusion produced a gap-free, more accurate DEM with greater topographic detail.Gamba et al. (2003) fused Star-3i InSAR and EagleScan LiDAR DEMs; Deng et al. (2011) combined prior-based selection with correlation-weighted arithmetic fusion.Source: §4.3.8 · pp. 38–39
  • An edge-preserving guided filter reduced TanDEM-X RMSE by 6–16% across three fusion experiments, while a weighted spectral-component method outperformed filter-based spectral fusion.Dong et al. (2018) fused ascending/descending multi-orbit, multi-baseline observations; Karakasis et al. (2014) tested five spectral expansion methods on SRTM and ASTER DEMs.Source: §4.3.8 · p. 39
  • Other weighting approaches used reference-site error variance and correlation or slope-based K-nearest-neighbour weight maps; adaptive-threshold weighted fusion produced an artefact-free, highest-resolution, temporally up-to-date DEM.Pham et al. (2018) transferred slope–weight relationships to regions without reference data; Papasaika et al. (2009) fused a 4m IKONOS image-matching DEM with a 2m airborne LiDAR DEM.Source: §4.3.8 · pp. 39–40
  • Sparse-representation fusion with adaptive regularisation produced more realistic terrain texture, preserved topographic features, and yielded stable error quality.Guan et al. (2020) combined 30m SRTM1, 90m TanDEM-X, and 30m AW3D-30; sparse representation extracted terrain information from the lower-resolution, higher-precision TanDEM-X DEM, while adaptive regularisation addressed discontinuities such as noise.Source: §4.3.8 · pp. 36, 37, 40

4.4 Pros and cons

DEM fusion methods trade simplicity and speed against artifact handling, accuracy, and terrain-detail preservation. Their benefits and limitations vary with landscape complexity and the assumptions used to model elevation errors.

Method trade-offs

  • Variational models can improve fusion accuracy beyond weighted averaging, especially in morphologically complex or high-frequency landscapes, but add computational expense and slow processing.Simple averaging is fast and inexpensive but may leave artifacts and edge effects, creating discontinuities or artificial landforms and altering correct source heights.Source: §4.4 · p. 40, Table 4 · p. 42

Super-resolution and self-consistency

  • Super-resolution methods suppress noise and integrate high- and low-resolution information, while self-consistency distributions do not map exactly to geospatial error distributions.A more comprehensive formulation is needed to predict geospatial-error statistics precisely from self-consistency distributions.Source: §4.4 · pp. 40, 43

Clustering

  • K-means clustering requires no prior input-DEM error information, but its assumption that more clustered elevations are more accurate can introduce uncertainty.Only a few user-defined data-distribution parameters are needed to set thresholds.Source: §4.4 · pp. 41, 44

Kalman methods

  • The multiscale Kalman smoother can substantially improve fused DEMs, particularly in steep channels where InSAR data are affected by layover and foreshortening.Reported benefits include smaller height uncertainty than in source DEMs and improved channel cross-sections.Source: §4.4 · pp. 41, 44

Multi-scale decomposition and other methods

  • Multi-scale decomposition with slope-position-based regression produced clear texture and detail, avoided mutation regions, and supported noise detection and trend-signal consistency.Other reported advantages across methods include improved terrain detail, smoother solutions, and reduced elevation-error dispersion.Source: §4.4 · pp. 43–44

Filter-based methods

  • Guided filters use local spatial context to reduce noise, fill voids, and preserve terrain details through gradient consistency, while remaining computationally efficient.Filter-based methods are also associated with improved vertical accuracy.Source: §4.4 · p. 44

4.5 Analysis of DEM sources and characteristics

The reviewed studies most often used TanDEM-X, SRTM, and ASTER as source DEMs, while LiDAR was the most frequent reference dataset. InSAR dominated source-DEM production, and reference DEMs were generally finer in resolution than the fused sources.

Source DEMs

  • TanDEM-X was the most common source DEM, appearing 12 times, with SRTM and ASTER also frequently used, partly reflecting their global coverage and free availability.No spaceborne LiDAR DEMs, including ICESat and ICESat-2, appeared in the reviewed studies.Source: §4.5 · p. 46, Figure 12 · p. 47

Reference datasets

  • LiDAR DEMs, especially airborne LiDAR, were the most frequently used references because their high resolution, accuracy, and vegetation penetration support validation.Ground control points and topographic maps or data were also used; GCPs are often unavailable in difficult terrain, and classical surveys are laborious and time-consuming.Source: §4.5 · p. 47, Figure 13 · p. 48

Production technologies

  • Most source DEMs were generated through InSAR, followed by stereo-optical photogrammetry, while LiDAR was used mainly for reference data.Other source-production approaches included Structure-from-Motion photogrammetry, image matching, and stereo-radargrammetry.Source: §4.5 · p. 48, Figure 14 · p. 49, Figure 15 · p. 49

Spatial resolution

  • Most reference DEMs had finer spatial resolution than the source DEMs being fused, while 30 m SRTM and ASTER were commonly adopted as sources, likely due to availability rather than accuracy.A few DEMs with resolution greater than 30 m, such as ICESat LiDAR, were used for validation.Source: §4.5 · pp. 49–50

4.6 Parameters influencing weight maps and height error maps

Fusion weights and height error maps commonly rely on source-DEM uncertainty information, but global or modeled error measures may miss important errors. The reviewed studies also use terrain classes and learned weights to account for spatially varying conditions.

Weight maps and height error maps

  • Weight maps and height error maps model source-DEM height-error distributions, but global uncertainty can miss accuracy variation caused by scene, processing, acquisition, and sensor differences.HEMs may also omit major errors such as InSAR elevation offsets and phase-unwrapping errors.Source: §4.6 · p. 51, Table 5 · p. 52

Terrain-based parameters

  • Terrain-based weighting can assign one weight to grids in the same slope and visibility class, reducing fusion cost; weighted averaging using slope, slope variability, and landform type improved accuracy and removed most artifacts.The latter approach assessed SRTM and ASTER GDEM across flat areas, valleys, and mountain slopes.Source: §4.6 · p. 52

Learned weights

  • ANN-predicted weights improved DEM relative accuracy by up to 50% in urban areas and 22% in non-urban areas, versus up to 20% and 10% respectively for HEM-based fusion.The ANN used features including slope, aspect, surface roughness, anisotropic coefficient of variation, and height residuals against LiDAR ground truth.Source: §4.6 · p. 53

Study-site terrain

  • Of the reviewed sites, 54% were hilly or mountainous, where SAR geometric distortions can negatively affect InSAR DEM quality and suitability for fusion.Another 25% of sites were flat lands or floodplains, and 8% were low-relief areas or lowlands.Source: §4.6 · p. 53, Figure 18 · p. 54, Figure 19 · p. 54

4.7 Fusion of multi-geometry InSAR DEMs

Multi-geometry InSAR fusion uses complementary acquisitions to address geometric distortions, data voids, and height errors. The reviewed studies report improved DEM quality from combining ascending and descending passes in difficult terrain.

Complementary acquisition geometries

  • Fusing ascending and descending InSAR DEMs can compensate for geometric distortions and data voids, especially in mountainous areas, while multi-baseline or multi-frequency fusion can reduce some errors.Different acquisition properties can be useful in different conditions, including larger heights of ambiguity for forested areas and dual-baseline phase unwrapping for phase-unwrapping stability.Source: §4.7 · p. 55

Reliability parameters

  • InSAR DEM reliability depends on phase-unwrapping method, coherence-based height error, and height of ambiguity, which in turn depend on processing method and orbit geometry.Geometric thresholds such as slope, aspect, baseline, incidence angle, and satellite heading can introduce additional artifacts.Source: §4.7 · p. 55

Reported outcomes

  • Combining ascending and descending TanDEM-X passes was reported to improve fused DEM quality in difficult terrain and areas affected by layover and shadow.The investigated fusion methods were weighted averaging and the TV-L1 and Huber variational models.Source: §4.7 · p. 56

4.8 Accuracy assessment

Accuracy assessment relies mainly on reference elevations and RMSE, but terrain, land cover, data availability, and inconsistent study settings complicate comparisons. Across reviewed studies, reported RMSEs range from below 2 m for some methods to above 10 m for others.

4.8 Accuracy assessment

  • RMSE is the most widely adopted vertical-accuracy metric, but can be unreliable when data are not normally distributed.Reference heights such as GCPs or higher-accuracy DEMs are commonly used; where they are unavailable or inadequate, non-classical measures include neighborhood comparisons, slope- and aspect-based anomaly detection, and elevation or aspect histograms.Source: §4.8 · p. 56
  • Comprehensive evaluation requires dense ground-truth coverage, but the resources for extensive ground-truthing are often infeasible.A single whole-product accuracy value can also miss regional differences associated with terrain steepness and land cover.Source: §4.8 · p. 56
  • Some Bayesian inference, linear-combination weight estimation, and multi-scale decomposition methods achieved RMSEs below 2 m, while weighted averaging, maximum likelihood, guided filters, and variational models generally achieved 2–5 m.Clustering and simple averaging reached 5–10 m, with reported RMSEs as high as 6.4 m and 9.7 m; the highest accuracies reported for Kalman filtering, weighted sparse representations, FFT, and MPS were above 10 m.Source: §4.8 · p. 57
  • Cross-study accuracy comparisons are difficult because experiments vary in terrain, land cover, DEM combinations, resolution, preprocessing, and reported metrics.The review therefore summarized average RMSEs, extracting each study’s lowest reported RMSE and one primary method where methods were compared.Source: §4.8 · p. 57

4.9 Applications

The review describes DEM fusion applications ranging from wide-area elevation products and terrain analysis to hazard assessment, hydrology, water-body mapping, and 3D building models. Across these uses, fused DEMs support improved height estimates or analyses that depend on terrain representation.

4.9 Applications

  • Fused DEMs have been used for wide-area products, hazard assessment, 3D building modelling, military planning, flood and water-flow modelling, surface-water detection, and other hydrological applications.Weighted averaging has been extensively used to combine small-area and wide-area DEMs into global products, including TanDEM-X, GTOPO30, GMTED2010, EarthEnv-DEM90, and EU-DEM.Source: §4.9 · pp. 57–58
  • For Nevado del Ruiz, a fused 10 m DEM produced simulated lahar inundation zones with longer run-out, more channelized flows, and thicker deposits over a larger extent than the 30 m SRTM DEM.The authors report that the fused DEM enabled more detailed volcano-hazard assessment.Source: §4.9 · p. 59
  • Two water-flow modelling experiments showed the importance of robust, seamless DEM fusion for realistic simulation of flow patterns.The cases fused LiDAR with UAV DEMs and a scanned physical sand model with a LiDAR watershed DEM.Source: §4.9 · p. 59
  • Multi-geometry TanDEM-X fusion combined water-body detection and DEM generation, with the refined DEM separating water from linear objects such as bridges and calculating water-area elevations.The weighting scheme used interferometric phase error and perpendicular baseline alongside layover and shadow maps.Source: §4.9 · pp. 59–60
  • In Munich, LOD1 prismatic building models made from OSM footprints and fused elevations had lower RMSEs than the source DEM models, though all models remained systematically biased against high-resolution LiDAR.Reported RMSEs were 10.01 m for Cartosat-1 and 10.16 m for TanDEM-X, versus 9.97 m for ANN fusion, 9.5 m for weighted-average TanDEM-X fusion, 8.95 m for TV-L1 fusion, and 9 m for Huber fusion.Source: §4.9 · p. 60, Figure 22 · p. 61
  • Coherence-weighted averaging, maximum-coherence selection, and coherence-mixture decisions yielded DEMs more accurate than their source DEMs in a topographic-profile application.The fusion used coherence information to represent elevation reliability.Source: §4.9 · p. 62
  • For Finke River Gorge, fusing space-borne and airborne InSAR DEMs produced the greatest improvement in steep channels affected by shadowing and foreshortening.The fused DEM also enabled spatio-temporal evaluation of fluvial structure and monitoring of changes in meandering channel portions.Source: §4.9 · p. 62

5. Challenges, Open Issues and Future Directions

The review identifies practical, methodological, and evaluation gaps that constrain DEM fusion research and use. It recommends targeted parameter analysis, accessible workflows and collaboration, stronger benchmarking, and further study across underexplored settings and methods.

5. Challenges, Open Issues and Future Directions

  • Fusion workflows should balance user expectations with method robustness, computational efficiency, and global convergence, while providing accessible step-by-step procedures using available DEMs and software.Computationally intensive methods may be unsuitable for users with limited computing resources, and advanced algorithms may still fail to deliver the desired outcome.Source: §5 · p. 62
  • A pre-analysis of terrain parameters against height error is recommended to address multicollinearity and remove less influential predictors, since adding parameters does not necessarily improve fusion quality.The review also notes that screening source DEMs for non-ground components such as buildings and forests has been uncommon.Source: §5 · pp. 62–63
  • The review identifies machine learning and deep learning as future avenues for automating DEM fusion, including weight-map preparation, while noting no recorded use case for direct ML/DL DEM fusion in the reviewed literature.Existing neural-network support work predicted weight maps but did not integrate the fusion process into the network; computation burden, technical complexity, and processing time also limit implementation.Source: §5 · pp. 63–64
  • Decision-level fusion remains underexplored and could combine outcomes from multiple algorithms to resolve source-DEM differences and reinforce common terrain interpretations.Source: §5 · p. 64
  • International collaboration could use optical, InSAR, and LiDAR global DEM coverage to produce demand-driven, standardized analysis-ready fused DEM tiles for national and regional use.LiDAR’s preprocessing, format, capacity, and storage requirements can put it out of reach for non-technical user groups in global-south research institutes.Source: §5 · p. 65
  • Public wide-area DEMs are heterogeneous and often coarse, while limited high-quality reference data, ground truth, and training samples complicate fair evaluation; the review calls for benchmark datasets.It also identifies technical expertise as a barrier to wide-area fusion in data-sparse regions, including low-income countries and the global south.Source: §5 · pp. 65–66
  • Research has addressed vertical accuracy and some terrain representation, but topological consistency, continuity, and completeness remain future targets, and fusion in volcanic and arctic terrain is underexplored.The review also points to extra-terrestrial DEM fusion as a likely future research area as planetary DEMs and exploration expand.Source: §5 · p. 66

FUNDING

The authors acknowledge funding support for doctoral studies.

FUNDING

  • CJO received the University of Cape Town Postgraduate Funding Office International Student Award for doctoral studies.Source: FUNDING · p. 67
Loading 2203.15026v2…