⚡ electrical engineering

Dual Scheme Adaptive Slope-Based Filtering of Photogrammetric Point Clouds for Digital Terrain Model Generation

This study proposes a dual-scheme adaptive slope-based filtering method that segments digital surface models into vegetation and building domains to improve ground point classification, achieving high vertical accuracy (RMSE of 0.148 m) in complex UAV photogrammetric surveys while identifying void extent in dense vegetation as the primary remaining limitation.

Original authors: Muhammad Ulin Nuha, Musa Damar Hadi, Rizqi Aulia

Published 2026-08-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammad Ulin Nuha, Musa Damar Hadi, Rizqi Aulia

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a cartographer trying to draw a map of the ground beneath a forest. You have a drone that flies overhead and takes thousands of photos, stitching them together to create a 3D model of the world. This model, called a Digital Surface Model (DSM), is incredibly detailed, but it has a problem: it shows everything, including the trees, the buildings, and the power lines. To get a true map of the earth itself—a Digital Terrain Model (DTM)—you need to digitally "peel away" everything that isn't the ground. It's like trying to find the shape of a cake by removing all the frosting, sprinkles, and candles without squishing the cake underneath.

The challenge is that the ground isn't always flat. Sometimes it has hills, and sometimes the "sprinkles" (like trees) are just as bumpy as the ground. Old methods tried to use a single rule for everything: "If it's steep, it's a building; if it's flat, it's the ground." But this is like trying to sort a mixed bag of marbles and rocks by only looking at their size; you might accidentally throw away a small boulder or keep a large pebble. This paper tackles the messy reality of tropical landscapes, where dense vegetation and modern buildings sit side-by-side on uneven terrain, making those old "one-size-fits-all" rules fail miserably.

The Two-Headed Filter: A Smart Way to Clean Up Drone Maps

The researchers, working at the Institut Teknologi Sumatera in Indonesia, realized that trying to use one rule to clean up a drone map of a campus was like trying to use a single pair of scissors to cut both delicate lace and thick leather. You'd either ruin the lace or fail to cut the leather. So, they invented a "Dual Scheme" filter—a smart system that acts like a bouncer at a club, checking your ID and then deciding which line you get to stand in based on what you look like.

The Setup: A Campus Full of Clues
The team flew a drone over a 121-hectare chunk of their university campus. This area was the perfect test lab: it had 70 meters of hills and valleys, tall multi-story buildings with flat roofs, and thick, dense secondary forests. The drone took photos from different angles (not just straight down, but tilted too) to build a 3D point cloud with a resolution of 4.15 cm per pixel. That's sharp enough to see individual bricks and leaves.

The Problem: The "One-Size-Fits-All" Trap
Traditional filters use a "slope-based" approach. They look at how steep a spot is. If the ground tilts sharply, the filter assumes it's a building wall or a tree trunk and deletes it. If it's gentle, it keeps it as ground. But in a place with both steep hills and flat buildings, this breaks.

If you set the rule to "delete anything steep," you accidentally delete the real hills.
If you set it to "keep anything steep," you leave the buildings standing.

The Solution: Two Different Rules for Two Different Worlds
The authors' big idea was to split the map into two zones before applying the rules. They created two "descriptors"—special maps that describe the texture of the ground.

The "Roughness" Map (Scheme A): This looks at how bumpy a small patch of ground is. Think of it like running your hand over a surface. A forest canopy is like a shaggy rug; it's very bumpy and irregular. A flat roof or a paved road is like a smooth table. The filter uses this "roughness" to identify the forest. In these bumpy areas, the filter gets "permissive." It knows the ground underneath might be steep, so it only deletes things that are really steep, saving the actual hills.
The "Edge" Map (Scheme B): This looks for sharp lines, like the outline of a building. It uses a technique called Canny edge detection, which is like a high-contrast pencil sketch that only draws the borders where things change abruptly. A flat roof is smooth inside, but it has a sharp, hard edge around the perimeter. The filter uses this to find buildings. In these smooth areas, the filter gets "strict." It knows that if there's a steep slope here, it's almost certainly a wall, so it deletes it immediately.

The Process: Sorting and Smoothing
Once the map is split into "Bumpy Forest Zone" and "Smooth Building Zone," the filter applies the correct rule to each.

In the Forest: It keeps the steep hills but removes the trees.
In the City: It removes the buildings but keeps the flat ground.

After the "non-ground" stuff (trees and buildings) is removed, there are holes in the map. It's like eating the frosting off a cake and leaving empty spots. To fix this, the team used a "hybrid interpolation" method. Imagine you have a hole in a blanket; you pull the threads from the edges to fill the gap. They used two methods: one that pulls from the nearest neighbors and another that averages the distance from all surrounding points. Finally, they applied a "Gaussian smoothing" filter, which is like gently pressing a warm iron over the fabric to make the seams invisible, ensuring the transition between the real ground and the filled-in holes looks natural.

The Results: A Map That Works (Mostly)
The team tested their new map against 12 real-world GPS points they measured on the ground.

The Good News: The new map was incredibly accurate. The error was only 0.148 meters (about 6 inches). This is almost the same accuracy as the original drone photos before any filtering happened. This proves that their smart, two-rule system didn't accidentally ruin the real ground data.
The Catch: The paper admits a limitation. In the biggest, thickest patches of forest, the drone couldn't see the ground at all. When the filter removed the trees, it left a giant hole. The "filling in" method had to guess what the ground looked like in the middle of that hole. The result was a surface that was a bit too smooth and flat, like a ramp, rather than a real, bumpy forest floor.

Why This Matters
The authors aren't claiming this is the absolute perfect method that solves every problem in the world. They aren't using complex AI or deep learning that requires massive computers and training data. Instead, they showed that by simply looking at the geometry of the world—how rough it is and where its edges are—you can create a much smarter filter than the old "one-rule" systems.

They found that for areas with mixed terrain (hills, trees, and buildings), you need a flexible approach. You can't treat a steep hill the same way you treat a building wall. By splitting the problem into two zones and using the right tool for each, they managed to create a clean, accurate map of the bare earth without needing expensive laser scanners or super-computers. The main takeaway is that while we can't see through a thick jungle canopy with a camera, we can be much smarter about how we guess what's underneath, as long as we admit where our guesses are just that—guesses.

Technical Summary: Dual Scheme Adaptive Slope-Based Filtering of Photogrammetric Point Clouds

Problem Statement
Digital Terrain Model (DTM) generation from Unmanned Aerial Vehicle (UAV) photogrammetry is frequently hindered by the limitations of conventional ground filtering algorithms, particularly Slope-Based Filtering (SBF). Standard SBF relies on a single global slope threshold to distinguish terrain from non-terrain features. This approach is insufficient in heterogeneous environments containing both dense vegetation and closely spaced buildings. In such settings, a single threshold cannot simultaneously preserve genuine steep terrain relief (e.g., 70-meter slopes) and remove steep non-terrain features like building walls or vegetation canopies. Furthermore, while learned classifiers (deep learning) offer high accuracy, they require extensive labeled training data and computational resources often unavailable in operational mapping contexts. Additionally, photogrammetric point clouds under closed tropical canopies often lack sub-canopy ground returns, meaning filtering algorithms cannot recover terrain that was never observed, a limitation distinct from LiDAR applications.

Methodology
The study proposes a "Dual Scheme Adaptive Slope-Based Filtering" workflow that operates entirely within the raster domain using a Digital Surface Model (DSM) derived from UAV photogrammetry. The method avoids the need for ancillary spectral data or multi-return information. The workflow consists of five primary stages:

Descriptor Generation: Two distinct descriptors are derived from the DSM:
- Surface Roughness (Scheme A Descriptor): Calculated as the local standard deviation of elevation within a 9-pixel window (~0.75 m). This identifies geometrically rough areas, primarily vegetation canopies and building walls.
- Edge Density (Scheme B Descriptor): Generated by applying Canny edge detection to the normalized DSM, followed by a 7-pixel mean filter. This identifies the perimeters of buildings (strong vertical discontinuities) while ignoring internal flat surfaces.
Rule-Based Zoning: A zoning raster assigns each pixel to one of two processing schemes based on the descriptors:
- Scheme A (Variable Terrain): Assigned to pixels where roughness exceeds a threshold. This domain is intended for vegetated and steep terrain.
- Scheme B (Flat Terrain): Assigned to all other pixels, including open ground and building interiors. This domain is intended for flat surfaces where steep features are likely non-terrain structures.
Scheme-Dependent Thresholding: Slope is calculated for each pixel and compared against a threshold specific to its assigned scheme:
- Scheme A Threshold ( $\theta_A$ ): Set to 25°. This permissive threshold prevents the removal of genuine steep terrain found in vegetated areas.
- Scheme B Threshold ( $\theta_B$ ): Set to 15°. This strict threshold ensures that steep features in flat areas (e.g., building walls) are correctly classified as non-terrain.
Morphological Refinement: The resulting binary mask undergoes dilation to ensure full coverage of object boundaries and hole filling to convert building outlines into solid footprints.
Void Filling and Smoothing: Pixels classified as non-terrain are set to no-data. A hybrid interpolation approach fills these voids: Inverse Distance Weighted (IDW) interpolation is used first for large gaps, followed by nearest-neighbor interpolation for residual gaps. A Gaussian filter is applied to smooth transitions between observed and interpolated surfaces.

Key Contributions
The paper's primary contribution is the demonstration that the adaptivity lacking in conventional SBF can be restored by making the processing domain adaptive rather than the threshold itself. By using a two-descriptor zoning step (roughness and edge density) derived solely from the DSM, the method applies distinct geometric rules to distinct surface types without requiring machine learning training data. The approach offers:

Transparency: All parameters (window sizes, slope thresholds) are physically interpretable and inspectable.
Computational Efficiency: The method operates on raster data using standard image-processing operators, avoiding the high computational cost of point-cloud processing or deep learning.
Geometric Logic: It successfully distinguishes between vegetation (rough) and buildings (smooth interiors with rough edges) using only geometric descriptors.

Results
The method was evaluated on a 121-hectare block of the ITERA campus in Lampung, Indonesia, characterized by dense secondary vegetation, multi-story buildings, and 70 meters of relief. The input data was an oblique UAV survey with a ground sampling distance of 4.15 cm and a point density of 145 points/m².

Vertical Accuracy: Comparison with 12 independent GNSS check points yielded a vertical Root Mean Square Error (RMSE) of 0.148 m and a Linear Error at 90% confidence (LE90) of 0.244 m.
Impact of Filtering: The filtered DTM showed an increase of only 0.004 m in RMSE compared to the input DSM. The authors note this increment is below the validation noise floor (GNSS precision and grid sampling effects), indicating that the filtering process caused no measurable degradation of elevation fidelity on observed terrain.
Visual Assessment: The resulting DTM successfully removed buildings and vegetation, preserving the broad topographic form. However, visual inspection revealed that large vegetated patches resulted in "over-smoothed" surfaces. Because the canopy was removed entirely, leaving no sub-canopy observations, the interpolation had to rely on perimeter values, resulting in planar fills that lack local terrain detail.

Significance and Limitations
The authors position this work not as a superior alternative to learned filters or cloth simulation, but as a practical, reproducible solution for operational mapping where training data is unavailable. The study highlights that the primary limitation of photogrammetric DTM generation in vegetated areas is not the classification algorithm, but the reconstruction bottleneck: if the photogrammetric cloud lacks ground returns under a closed canopy, no filter can recover the terrain, and the resulting DTM quality is dictated by the interpolation method rather than the filter.

The paper concludes that while the dual-scheme approach effectively resolves the trade-off between preserving terrain relief and removing structures in mixed environments, the accuracy of the final product in vegetated voids remains unmeasured by standard check points (which are placed on open ground). The authors emphasize that the reported accuracy metrics characterize open terrain only and should not be extrapolated to vegetated interiors where the model is weakest. Future work is recommended to focus on preserving sub-canopy evidence in the point-cloud domain and conducting stratified validation using laser-scanned references for vegetated areas.

The Two-Headed Filter: A Smart Way to Clean Up Drone Maps

More like this