⚡ electrical engineering

ASND‑DEM: An Adaptive Spatial Nyquist Decomposition Framework for Optimal DEM Resolution Mapping from Photogrammetric Point Clouds in Open‑Pit Mining

This study introduces ASND-DEM, an adaptive framework that optimizes DEM resolution for photogrammetric point clouds in open-pit mining by dynamically adjusting grid density based on local point spacing and terrain roughness, achieving significant data compression while preserving critical geological features.

Original authors: Amir Sarvi, Abbas Abedini, Farzaneh Dadrass Javan

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amir Sarvi, Abbas Abedini, Farzaneh Dadrass Javan

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 tasked with drawing a map of a massive, rugged open-pit mine. You have a drone that took millions of photos, creating a 3D "point cloud" of the terrain. This cloud has 96.5 million dots, each representing a spot on the ground.

Your goal is to turn these dots into a Digital Elevation Model (DEM)—a smooth, grid-based map that engineers can use to check if the mine walls are safe or to calculate how much rock has been moved.

The Problem:
If you try to draw the map using a single grid size for the whole mine, you run into a dilemma:

Too Coarse (Big Grid): If your grid squares are too big, you miss the sharp, dangerous edges of the mine steps (benches) and the steep slopes. It's like looking at a jagged mountain range through a foggy window; the details blur away.
Too Fine (Tiny Grid): If your grid squares are tiny enough to catch every single rock and crack, the file becomes huge. It's like trying to carry a library of encyclopedias in your pocket. The computer slows down, storage fills up, and you end up recording "noise" (sensor errors) instead of real terrain.

The paper argues that there is no "one size fits all" grid size because a mine has both smooth, flat haul roads (which don't need a tiny grid) and jagged, broken rock faces (which do).

The Solution: ASND-DEM
The authors created a smart framework called ASND-DEM. Think of it as an intelligent, shape-shifting camera that automatically adjusts its focus and zoom level depending on where it is looking.

Here is how it works, broken down into four simple steps:

1. The "Nyquist" Rule (Setting the Limits)

First, the system looks at how close the drone's dots are to each other. It uses a rule from physics called the Nyquist-Shannon theorem.

The Analogy: Imagine you are trying to hear a musical note. To hear it correctly, you need to sample the sound at least twice as fast as the note's pitch. If you sample too slowly, you hear the wrong note (aliasing).
In the Mine: The system calculates the "median distance" between dots. It decides that the map grid cannot be larger than twice that distance, or it will lose the shape of the mine. This sets the "coarsest possible" grid size allowed.

2. Measuring "Roughness" (The Texture Check)

Next, the system walks through the point cloud and asks: "How rough is this specific spot?"

It looks at three things: how much the height changes, how steep the slope is, and how much the ground curves.
The Analogy: Think of a piece of fabric. Some parts are smooth silk (the flat haul roads), and some are crumpled burlap (the blasted rock walls).
The system uses a smart math trick called Entropy to decide which of these three measurements matters most. It found that for mines, "slope" and "curvature" are the best indicators of roughness, while simple height changes are often redundant.

3. The "Smart Map" (Matching Grid to Terrain)

Now, the system creates a custom map that tells it exactly what grid size to use for every single square inch of the mine.

The Analogy: Imagine a tailor making a suit. Instead of using the same fabric thickness for the whole body, they use thick, durable fabric for the knees (rough areas) and thin, light fabric for the chest (smooth areas).
The Result: On the flat haul roads, the system says, "Use a big, coarse grid." On the jagged bench edges, it says, "Switch to a tiny, fine grid."
It found that the best "average" grid size for this specific mine was about 0.4 meters, but it allowed the grid to shrink to 0.13 meters in rough spots and grow to 0.53 meters in smooth spots.

4. The "Octree" (The Magic Filing Cabinet)

Finally, the system organizes the 96.5 million dots into a structure called an Octree (a 3D tree-like filing system).

The Analogy: Imagine a library. Instead of giving every single book its own shelf, the librarian groups books by topic. If a section has 10,000 books, it gets a big shelf. If a section has only 5 books, it gets a tiny shelf.
The Result: The system compressed the massive 96.5 million points down into just 542,443 "tiles."
The Magic: It reduced the data volume to just 0.56% of the original size (a 178-fold reduction!) without losing the sharp edges of the mine walls. It kept the "rough" details where they mattered and threw away the unnecessary "smooth" data where it didn't.

The Bottom Line

The paper proves that you don't need to choose between a blurry map and a massive file. By using this ASND-DEM framework:

Safety is preserved: The sharp edges of the mine walls (critical for stability) are kept in high detail.
Efficiency is maximized: The smooth, flat areas are simplified, saving massive amounts of computer storage and time.
It's automatic: The system figures out all the settings itself; you don't need to guess the right grid size.

In short, it's a smart way to turn a mountain of data into a manageable, high-quality map that knows exactly where to zoom in and where to zoom out.

Technical Summary: ASND‑DEM Framework for Optimal DEM Resolution Mapping

1. Problem Statement

Selecting an appropriate resolution for Digital Elevation Models (DEMs) generated from photogrammetric point clouds in open-pit mining remains a critical unresolved challenge. A resolution that is too coarse fails to capture essential engineered features such as bench edges, slope breaks, and erosion patterns, thereby compromising slope stability analysis and volumetric calculations. Conversely, a resolution that is too fine amplifies sensor noise, inflates data volume, and increases computational costs quadratically.

The core difficulty lies in the fact that no universal fixed resolution exists. Optimal resolution depends on both the local point density and the surface roughness. Mining terrains are particularly heterogeneous, containing smooth haul roads (requiring coarse resolution) alongside rugged bench faces (requiring fine resolution). Existing methods often rely on fixed-grid approaches or simple filtering that cannot simultaneously preserve critical discontinuities and compress data efficiently across such variable landscapes.

2. Methodology: The ASND‑DEM Framework

The study proposes ASND‑DEM (Adaptive Spatial Nyquist Decomposition for Digital Elevation Models), a four-stage framework designed to answer two questions: (1) What fixed resolution is appropriate for a given dataset? and (2) How can resolution vary spatially to match local complexity?

Stage 1: Spatial Nyquist Analysis

This stage derives theoretically permissible resolution scales based on the Nyquist–Shannon sampling theorem adapted to the spatial domain.

Process: The framework calculates the $k$ -nearest neighbor distances ( $k=3$ ) for the input point cloud to estimate local sampling density. A Gaussian Mixture Model (GMM) identifies the dominant sampling mode, yielding the median nearest-neighbor distance ( $P_{50}$ ).
Output: A discrete scale table of permissible resolutions is generated, ranging from $0.5 \times P_{50}$ to $2.0 \times P_{50}$ . The spatial Nyquist limit is defined as $R_{Nyquist} = 2 \times P_{50}$ , representing the coarsest resolution that avoids spatial aliasing.

Stage 2: Multivariate Roughness Characterization

This stage quantifies local surface complexity using three geometric descriptors: elevation standard deviation ( $r_{std}$ ), slope gradient ( $r_{slope}$ ), and curvature residual ( $r_{curv}$ ).

Adaptive Weighting: Four weighting schemes (Equal, PCA, Entropy, Spectral) are evaluated to fuse these descriptors. An automatic decision score (balancing information entropy, detail preservation, and stability) selects the optimal weighting method.
Result: In the case study, the Entropy-based method was selected, assigning near-equal weights to slope and curvature while discarding elevation standard deviation due to redundancy.

Stage 3: Roughness-to-Resolution Mapping

This stage maps the normalized roughness values to the optimal local resolution ( $R_{best}$ ) within the permissible scale set.

Mapping Functions: Three families of functions (Polynomial, Exponential, Logarithmic) are tested.
Selection: An information-theoretic composite score (incorporating entropy, fine-detail preservation, and stability) selects the optimal function. The study found that an Exponential mapping ( $\alpha=5$ ) best captured the relationship, where even moderate roughness triggers a rapid shift to finer resolutions.
Fixed Resolution: For applications requiring uniform grids, an optimal fixed resolution ( $R_{opt}$ ) is derived as the median of the spatially varying $R_{best}$ map, determined via a cost-quality trade-off analysis.

Stage 4: Adaptive Octree Tiling

The final stage constructs a hierarchical spatial partition (referred to as an adaptive octree, functioning as a quadtree for 2.5D height fields) driven by the $R_{best}$ map.

Splitting Criteria: Nodes are split based on two conditions derived from local resolution statistics: (1) the tile size exceeds a threshold relative to the local mean resolution, and (2) the local resolution heterogeneity exceeds a threshold.
Compression: This process compresses the point cloud into tiles where the resolution is proportional to local surface complexity. Smooth areas are represented by coarse tiles, while rough areas retain fine tiles.

3. Key Contributions

Theoretical Adaptation: Adapts the Nyquist–Shannon theorem to the spatial domain for point cloud-derived DEMs, establishing a physically constrained scale table ( $0.5\times$ to $2.0\times P_{50}$ ).
Multivariate Roughness Fusion: Introduces a multivariate roughness characterization with automatic weighting selection (via Entropy, PCA, etc.), eliminating manual parameter tuning.
Information-Theoretic Mapping: Develops a roughness-to-resolution mapping strategy that selects the optimal functional form (polynomial, exponential, or logarithmic) based on information-theoretic metrics.
Adaptive Octree Compression: Implements a roughness-driven adaptive octree that achieves data compression proportional to local surface smoothness, rather than using fixed voxel sizes.
Large-Scale Validation: Validates the framework on a real-world, large-scale dataset (96.5 million points) from an open-pit mine, demonstrating near-linear scaling and significant compression.

4. Results

The framework was applied to a photogrammetric point cloud of 96,555,133 points from an open-pit mine in Iran.

Resolution Analysis: The dominant point spacing ( $P_{50}$ ) was 0.264 m, establishing a Nyquist limit of 0.528 m. The adaptive resolution varied spatially from 0.132 m (on rough bench edges) to 0.528 m (on smooth haul roads), a 4× range.
Optimal Fixed Resolution: The analysis identified 0.396 m (1.5× $P_{50}$ ) as the optimal fixed resolution for uniform-grid DEMs, representing the "knee point" where further coarsening yields diminishing returns in data reduction relative to error increase.
Compression Performance: The adaptive octree compressed the 96.5 million points into 542,443 tiles (averaging ~178 points/tile). This achieved a storage compression ratio of 0.56% (approximately 178× reduction) compared to the original point cloud volume.
Feature Preservation: Critical topographic features (bench edges, slope breaks) were preserved in fine-resolution tiles, while smooth areas were aggressively compressed.
Generalization: The framework was successfully applied to a second independent mine site (20.2 million points), achieving similar compression ratios (0.62%) and automatically adapting the optimal fixed resolution (0.240 m) to the site's specific topography.
Comparison: Compared to fixed-space subsampling and Cloth Simulation Filter (CSF) baselines, ASND-DEM offered superior storage efficiency while retaining all structural points (unlike CSF, which discards off-ground objects) and preserving fine features better than uniform subsampling.

5. Significance and Claims

The paper claims that ASND-DEM provides a physically constrained, data-driven solution for variable-resolution DEM generation in heterogeneous landscapes. Its significance lies in:

Eliminating Trial-and-Error: It replaces ad-hoc resolution selection with a theoretically grounded criterion based on point density and surface roughness.
Efficiency vs. Fidelity: It demonstrates that variable-resolution mapping can reduce storage requirements by two orders of magnitude while preserving centimeter-scale features essential for mining safety and volume calculations.
Scalability: The near-linear computational scaling ( $O(N \log N)$ ) makes the framework suitable for large-scale mining operations where point clouds routinely exceed 100 million points.
Generalizability: The framework is designed to be applicable to any photogrammetric or LiDAR point cloud, not just mining environments, though it was specifically validated on engineered mining terrain.

The authors note that the framework requires a complete point cloud as input and does not currently support real-time incremental processing, but it offers a robust offline solution for optimizing DEM generation in complex, engineered terrains.