In the field of geospatial analysis and remote sensing, the construction of high-quality Digital Elevation Models (DEMs) is crucial for applications such as resource monitoring, infrastructure planning, and disaster assessment. However, complex terrains like forests, glaciers, deserts, and urban areas pose significant challenges due to issues like occlusion, data gaps, and varying sensor limitations. Traditional methods relying on single data sources often fall short in these environments. To address this, we propose an integrated approach that leverages the complementary strengths of optical imagery from UAV drones and Synthetic Aperture Radar (SAR) data from satellite remote sensing. This method focuses on multi-source DEM generation and fusion, enhancing accuracy, continuity, and adaptability across diverse landscapes. By combining high-resolution UAV-based optical DEMs with SAR-derived DEMs, and incorporating advanced fusion techniques, we aim to provide a robust framework for large-scale, high-precision terrain modeling. This research not only advances the theoretical foundations of multi-source data integration but also offers practical solutions for real-world mapping challenges, pushing the boundaries of automated and intelligent remote sensing.
The theoretical underpinnings of our approach are rooted in the principles of remote sensing and data fusion. DEMs can be generated from optical imagery through stereo vision and from SAR data via interferometry. Each source has distinct characteristics: optical imagery excels in high-texture areas but is affected by weather conditions, while SAR penetrates clouds and operates in all-weather but suffers from coherence loss in vegetated or watery regions. The fusion of these sources requires a mathematical framework to ensure consistency and precision. Let us define the elevation estimation model for a remote sensing source \(i\) as:
$$D_i(x,y) = f_i(I_i, G_i, P_i, \epsilon_i)$$
where \((x,y)\) are spatial coordinates, \(I_i\) is the imagery, \(G_i\) represents ground control information, \(P_i\) denotes platform parameters, and \(\epsilon_i\) is the error term. The function \(f_i(\cdot)\) encapsulates the mapping from imagery to elevation, differing for optical and SAR data. For optical stereo, it involves disparity calculation from image pairs, while for SAR interferometry, it uses phase differences and geometric parameters.
To fuse multiple DEM sources, we adopt a weighted average model based on minimum variance unbiased estimation. The fused DEM \(D_f(x,y)\) is given by:
$$D_f(x,y) = \frac{\sum_{i=1}^{n} \omega_i(x,y) \cdot D_i(x,y)}{\sum_{i=1}^{n} \omega_i(x,y)}$$
Here, \(n\) is the number of sources, and \(\omega_i(x,y)\) is the weight function, typically inversely proportional to the elevation variance \(\sigma_i^2(x,y)\):
$$\omega_i(x,y) = \frac{1}{\sigma_i^2(x,y)}$$
This ensures that more reliable data contributes more to the final output. However, in complex terrains, simple weighting may not suffice. We introduce adaptive strategies based on regional characteristics. For instance, in low-texture or high-reflectivity areas, SAR coherence \(C\) is used as a criterion:
$$C = \frac{\left| \sum_{j=1}^{N} s_{1,j} s_{2,j}^* \right|}{\sqrt{\sum_{j=1}^{N} |s_{1,j}|^2 \cdot \sum_{j=1}^{N} |s_{2,j}|^2}}$$
where \(s_{1,j}\) and \(s_{2,j}\) are complex signals from master and slave images, and \(N\) is the window size. Higher \(C\) values indicate better coherence, favoring SAR DEMs. In urban or forested regions, optical texture richness is measured via grayscale entropy \(H\):
$$H = -\sum_{k=1}^{L} p_k \log_2 p_k$$
with \(p_k\) being the probability of gray level \(k\) and \(L\) the number of levels. When \(H\) exceeds a threshold \(H_{th}\), optical DEMs are prioritized. Additionally, to control edge effects during fusion, we impose a seamline error constraint:
$$|D_i(x,y) – D_j(x,y)| \leq \tau, \quad (x,y) \in \Omega$$
where \(\tau\) is the tolerance limit, often set to 1.5 times the reference data standard deviation \(\sigma_r\), and \(\Omega\) is the overlap region. This prevents abrupt jumps and ensures smooth transitions.
For dynamic terrain monitoring, such as glacier flow or dune migration, we incorporate temporal modeling. The elevation change \(\Delta h_t(x,y)\) between times \(t\) and \(t-1\) is:
$$\Delta h_t(x,y) = D_t(x,y) – D_{t-1}(x,y)$$
and the cumulative evolution over \(n\) phases is:
$$h(x,y,t) = h_0(x,y) + \sum_{k=1}^{t} \Delta h_k(x,y)$$
This allows for quantitative analysis of terrain dynamics, supporting applications like disaster early warning. Our theoretical framework thus integrates static and dynamic aspects, enabling robust multi-source DEM generation and fusion.
The methodology for multi-source DEM construction involves a systematic pipeline, divided into data acquisition, independent DEM generation from optical and SAR sources, and fusion with precision control. We begin with data collection using integrated platforms. UAV drones, including fixed-wing and multi-rotor systems, are equipped with high-resolution optical cameras, GNSS/IMU modules, and adaptive control units. These UAV drones fly at altitudes of 80–120 m, ensuring high overlap (≥80%) and stable geometry. For areas with cloud cover or extreme terrain, we supplement with satellite data, such as SAR imagery from missions like TanDEM-X or GaoFen-3, and optical imagery from GaoFen-7. This multi-platform approach ensures comprehensive coverage across diverse landscapes.
For optical DEM generation from UAV drones, we follow a structured process. First, stereo models are recovered using software like MapMatrix Grid, producing an initial Digital Surface Model (DSM). Then, point cloud classification separates ground and non-ground points via triangulation-based iterative algorithms. Manual verification is applied in complex areas like forests or cities to correct misclassifications. After thinning and gridding, elevation values are converted from ellipsoidal heights to orthometric heights using geoid models. Quality assessment involves stereo consistency checks and profile analysis, with manual editing in tools like MapMatrix Grid to fix discontinuities. The overall accuracy is enhanced by adaptive filtering and control point integration. This optical DEM pipeline is highly automated but retains human oversight for reliability.
SAR DEM generation follows a dedicated interferometry workflow. Key steps include interferometric pair selection, image registration, phase difference computation, flat-earth effect removal, filtering, phase unwrapping, elevation inversion, and geocoding. We use multi-temporal SAR data to improve coherence in challenging areas. For example, in glacier regions, we employ multi-baseline interferometry to reduce errors from ice movement. The phase unwrapping leverages region-growing algorithms guided by coherence maps. The elevation inversion formula is:
$$\Delta \phi = \frac{4\pi}{\lambda} B_\perp \frac{\Delta h}{R \sin \theta}$$
where \(\Delta \phi\) is the unwrapped phase difference, \(\lambda\) is the wavelength, \(B_\perp\) is the perpendicular baseline, \(\Delta h\) is the height difference, \(R\) is the range, and \(\theta\) is the incidence angle. After inversion, DEMs from multiple orbits are mosaicked and refined with reference data. This SAR DEM process is robust to weather conditions but requires careful parameter tuning.
The fusion of optical and SAR DEMs is critical for seamless output. We implement a multi-stage strategy: pre-fusion adjustment, in-fusion control, and post-fusion validation. Initially, DEMs are aligned to a common coordinate system and height datum. Then, adaptive weighting is applied based on local terrain features. For instance, in forested areas, SAR weights are reduced due to low coherence, while in urban zones, optical weights are increased. The fusion uses the weighted average model with error constraints, as described theoretically. To handle edge effects, we apply seamline smoothing with window functions. A summary of fusion parameters for different terrains is shown in Table 1.
| Terrain Type | Primary Data Source | Weighting Criteria | Error Tolerance \(\tau\) (m) |
|---|---|---|---|
| Forest | Optical (UAV drones) | High texture entropy, low SAR coherence | 1.5 |
| Glacier | SAR (multi-baseline) | High SAR coherence, ICESat control | 2.0 |
| Desert | SAR (multi-temporal) | Stability analysis, feature preservation | 1.8 |
| Urban | Optical (UAV drones) | High texture entropy, building vector data | 1.2 |
| Water Bodies | Optical (multi-temporal) | Shoreline detection, interpolation | 1.0 |
Post-fusion, we validate the DEM using check points or reference datasets. Metrics like root mean square error (RMSE) and mean absolute error (MAE) are computed. If anomalies persist, iterative correction is performed. This methodology ensures that the fused DEM maintains high accuracy and continuity across all terrain types.
Experimental validation was conducted in multiple representative terrains: forests, glaciers, deserts, urban areas, and water bodies. For each, we generated DEMs using UAV drones for optical data and satellites for SAR data, then fused them. In forest regions, UAV drones captured high-resolution imagery, but tree canopies caused elevation overestimation. We applied a tree-height reduction model and manual correction, resulting in a DEM with smooth ground representation. The RMSE relative to ground truth was 0.8 m, meeting medium-scale mapping standards. In glacier areas, SAR data from multiple passes was used to track ice movement. The fused DEM showed continuous surfaces with an elevation change detection sensitivity of 0.5 m per year. For deserts, SAR interferometry handled low texture well, and fusion with historical maps reduced noise. The DEM achieved a planimetric accuracy of 2 m. Urban areas benefited from UAV drones’ detailed optical data, with building heights accurately extracted after fusion. Water bodies like rivers and reservoirs were challenging due to fluctuating shorelines; multi-temporal optical data helped delineate boundaries, and interpolation filled gaps.
To quantify performance, we assessed relative elevation accuracy against standard grades (e.g., flat, hilly, mountainous). The results are summarized in Table 2, based on industry norms. The fused DEM consistently met or exceeded accuracy thresholds, with hilly areas showing a relative elevation median error of 0.5 m. Seamline errors between adjacent tiles were controlled within 1.5 times the allowable error, ensuring smooth transitions. Dynamic analysis in deserts and glaciers revealed trends like dune migration rates of 10 m/year, demonstrating the method’s utility for monitoring.
| Terrain Category | Fused DEM RMSE (m) | Standard Allowable Error (m) | Compliance Status |
|---|---|---|---|
| Flat | 0.9 | 1.4 | Exceeds |
| Hilly | 0.5 | 1.4 | Exceeds |
| Mountainous | 1.8 | 2.3 | Meets |
| High Mountain | 2.4 | 4.0 | Exceeds |
The use of UAV drones was pivotal in these experiments, providing flexibility and high resolution. For instance, in urban canyons, UAV drones captured oblique imagery that improved building modeling. In remote glaciers, UAV drones supplemented satellite data with local details. The integration of UAV drones with satellites created a robust multi-view data chain, enhancing overall DEM quality. However, challenges remain, such as cost and real-time processing, which future work could address through optimized algorithms and cloud computing.
In conclusion, this research presents a comprehensive framework for multi-source DEM generation and fusion, leveraging the synergies between UAV drones and satellite remote sensing. By combining optical and SAR data through theoretical models, adaptive weighting, and precision control, we achieve high-accuracy, continuous DEMs across complex terrains. The method’s effectiveness is validated in diverse landscapes, with accuracy metrics surpassing standard requirements. Key innovations include the integration of UAV drones for detailed optical mapping, multi-baseline SAR processing for robust elevation retrieval, and dynamic modeling for terrain change analysis. This work advances remote sensing by enabling automated, intelligent DEM production, with applications in topographic mapping, environmental monitoring, and disaster management. Future directions may involve incorporating lidar data or deep learning techniques to further enhance fusion quality and efficiency.