Ground Diffusion (GrounDiff): Methods & Applications

• Ground Diffusion (GrounDiff) is a paradigm that applies diffusion-based denoising processes to model ground motion, terrain features, and visual grounding with high fidelity.
• It leverages architectures such as DDPMs, U-Nets, and transformer conditioners to iteratively remove noise and extract key patterns from seismic signals, DSM data, and images.
• The approach finds broad applications in DSM-to-DTM translation, earthquake ground motion synthesis, and text-conditioned object localization, achieving state-of-the-art performance in various benchmarks.

Ground Diffusion (GrounDiff) encompasses a set of methodologies in which diffusion-based generative or denoising processes are leveraged for problems of ground motion modeling, terrestrial surface generation, and visual grounding. Applications span geophysics (earthquake seismogram synthesis, ground surface modeling), remote sensing, and zero-shot object localization. Techniques include denoising diffusion probabilistic models (DDPMs), conditional latent diffusion, wavelet packet-based diffusion frameworks, and multi-modal (text–image) diffusion pipelines, often integrated with transformers, U-Nets, and domain-specific architectural modules. The GrounDiff paradigm treats above-ground structures, unmodeled spectral noise, or localizations as “noise” to be iteratively removed, thus offering a unified view of both physical ground motion and abstract “grounding” in imagery or signals.

1. Mathematical Formulations and Model Variants

Multiple GrounDiff formulations exist, tailored to distinct signal modalities:

• Digital Surface Model to Terrain Model (DSM→DTM): Observed DSM $s\in\mathbb{R}^{H\times W}$ is mapped to the unknown ground DTM $g\in\mathbb{R}^{H\times W}$ via a forward noising chain

$$q(g_t|g_0) = \mathcal{N}(g_t;\sqrt{\bar\alpha_t}g_0,(1-\bar\alpha_t)I)$$

and reverse generative process parameterized by a U-Net predicting a residual map $\hat r$ and per-pixel confidence logit $\ell$, gated as $$\hat g_0 = \sigma(\ell)\odot s + (1-\sigma(\ell))\odot(s-\hat r)$$ (Dhaouadi et al., 13 Nov 2025).

• Earthquake Ground Motion Synthesis: Both waveform domain (time series, wavelet packets) and latent frequency-space representations are used. Diffusion in wavelet coefficient space allows direct control over spectral and temporal features per conditional vector (e.g., Arias intensity $I_A$, response spectra $S_a(T)$) (Ding et al., 7 Feb 2026, Jung et al., 2024).
• Visual Grounding/Object Grounding: Text-conditional DDPMs perturb bounding boxes, or image pixels, with noise and iteratively denoise to yield localized regions or objects based on textual queries (Chen et al., 2023, Sethi et al., 20 Apr 2026).
• Ground-to-Aerial Synthesis: Ground-level street-view images are projected into a spatial “prior” (BEV, Bird's Eye View); a diffusion model then hallucinates a consistent aerial image, via a denoising process conditioned on the BEV prior (Ye et al., 2024).

In all cases, the forward noising process follows

$$q(x_t|x_{t-1}) = \mathcal{N}(x_t; \sqrt{1-\beta_t}x_{t-1},\beta_t I)$$

with corresponding closed-form marginal and learned reverse chain

$$p_\theta(x_{t-1}|x_t, c) = \mathcal{N}(x_{t-1};\mu_\theta(x_t,c,t),\sigma_t^2 I)$$

where $g\in\mathbb{R}^{H\times W}$0 denotes the relevant conditional input(s).

2. Diffusion Model Architectures and Conditioning

Architectural Features:

• U-Net Denoisers: Central to GrounDiff is the U-Net, often enhanced with residual blocks, self-attention (at bottleneck), and time or noise-level embeddings. For visual or spectral data, cross-attention injects conditional information (e.g., geological profile, text embedding, BEV prior) at every layer (Dhaouadi et al., 13 Nov 2025, Ding et al., 7 Feb 2026, Sethi et al., 20 Apr 2026).
• Wavelet Packet & Frequency-Domain Modules: For seismic signals, wavelet packet transforms disassemble accelerograms into multiresolution bands, offering near-lossless synthesis. Diffusion operates directly on normalized subband coefficients, ensuring perfect reconstruction and interpretable frequency control (Ding et al., 7 Feb 2026).
• Transformer Conditioners: Heterogeneous (vectorial, scalar) metadata, such as seismic spectral features, scalar intensity metrics, or geometric priors, are embedded and fused via transformer encoders (Ding et al., 7 Feb 2026).
• Prior-Guided Stitching: For large images/terrains, a coarse global prior is generated via low-resolution GrounDiff inference, upsampled to guide local, high-res denoising steps, achieving seamless tiling and minimizing boundary artifacts (Dhaouadi et al., 13 Nov 2025).
• Amplitude Correction Modules: For high-fidelity waveform synthesis, an amplitude correction module (composed of TS-Conformer blocks and mask decoders) refines the final spectrograms beyond autoencoder output, restoring accurate spectral envelopes (Jung et al., 2024).
• Hybrid Pipelines: Integration with segmentation models (e.g., SAM3, RemoteSAM) refines object masks after diffusion-based localization (Sethi et al., 20 Apr 2026).

3. Training Objectives, Losses, and Inference

Training: Core objectives in GrounDiff are L2 (MSE) denoising losses between predicted and target clean signals (images, boxes, terrain maps, coefficients), optionally augmented by auxiliary regularization:

• For DTM extraction, the combined loss includes L1, L2, gradient smoothness, and binary cross-entropy terms for ground mask confidence (Dhaouadi et al., 13 Nov 2025).
• For wavelet-packet-based ground motion, explicit spectral or temporal-energy matching losses are omitted; instead, conditioning on features such as $g\in\mathbb{R}^{H\times W}$1, Husid curve $g\in\mathbb{R}^{H\times W}$2, $g\in\mathbb{R}^{H\times W}$3, and others enables implicit control (Ding et al., 7 Feb 2026).
• In visual grounding/object localization, box regression losses (SmoothL1, GIoU), cross-modal similarity, and Hungarian assignment penalize localization and alignment errors (Chen et al., 2023, Sethi et al., 20 Apr 2026).

Inference Strategies:

• Second-order Heun solvers in the EDM framework reduce inference steps (50 network evals vs. 1000 in DDPM) for efficiency without sacrificing fidelity (Ding et al., 7 Feb 2026).
• Classifier-free guidance and multi-branch cross-attention enable strong conditional sampling in pixel, box, or coefficient space.
• Dynamic model selection and fallback (as in DiffuSAM, toggling between SAM3 and RemoteSAM) maximize recall/precision in object segmentation (Sethi et al., 20 Apr 2026).
• For ground-to-aerial, a multi-to-one BEV fusion approach selects the closest-view panorama for each pixel, reducing occlusion artifacts (Ye et al., 2024).

4. Performance and Benchmarking

Terrain and DTM:

• GrounDiff sets state-of-the-art on ALS2DTM and USGS benchmarks, delivering up to 93% reduction in RMSE over learning-based baselines (e.g., DeepTerRa, HDCNN). On GeRoD road reconstruction, GrounDiff+ achieves up to 81% lower median distance error (MED) compared to FlexRoad, and smoother surfaces (road MAD $g\in\mathbb{R}^{H\times W}$4) with only DSM inputs (Dhaouadi et al., 13 Nov 2025).

Seismic Ground Motion:

• In NGA-West2 tests, GroundDiff frameworks achieve spectral R² > 0.9 in 91% of samples; explicit energy conditioning improves Husid curve R² from 25% to 88%, and significantly tightens matching bands (±1σ ≈ 0.02 in energy domain) (Ding et al., 7 Feb 2026).
• HEGGS outperforms GAN and vanilla LDM baselines in P/S phase MAE (0.48 s / 0.55 s), envelope correlation (0.82), and spectrogram MSE (0.15), generating synthetic ground motions that align with GMPEs and observed data in the 1–45 Hz band (Jung et al., 2024).

Remote Sensing/Object Grounding:

• DiffuSAM achieves 14.25% improvement in [email protected] vs. RemoteSAM (0.8049 vs. 0.6624) on NWPU-VHR-10, indicating superior localization in complex aerial scenes (Sethi et al., 20 Apr 2026).
• In language-guided visual grounding, LG-DVG attains [email protected] = 79.92% (Flickr30k), competitive with or exceeding multi-modal baselines (Chen et al., 2023).

Ground-to-Aerial Synthesis:

• SkyDiffusion produces up to 36% lower FID and 20% higher SSIM than GAN/diffusion alternatives on VIGOR-Chicago. BEV conditioning and light-style modules ensure geometric and photometric realism across urban datasets (Ye et al., 2024).

5. Key Mechanisms, Strengths, and Limitations

Key Mechanisms:

• Iterative denoising aligns with the physical intuition of filtering: above-ground artifacts or extraneous frequency content are removed akin to noise.
• Gated architectures and explicit per-pixel confidence enable region-specific correction or preservation, reducing overcorrection and retaining critical signal/terrain features (Dhaouadi et al., 13 Nov 2025).
• Cross-attention and multi-scale encoding allow GrounDiff to condition on arbitrary-length feature vectors/sequences, incorporating domain knowledge (e.g., spectra, temporal statistics, text, BEV priors) (Ding et al., 7 Feb 2026, Jung et al., 2024, Ye et al., 2024, Chen et al., 2023).

Strengths:

• Zero-shot generalization: No task-specific fine-tuning is required; the same architecture is applicable across domains, from seismic signals to visual object localization (Sethi et al., 20 Apr 2026, Ding et al., 7 Feb 2026).
• Scale-invariant modeling: The fractal nature of true ground diffusion is encoded (e.g., via the ATL law $g\in\mathbb{R}^{H\times W}$5), enabling simulation of ground drift over arbitrary spatial and temporal scales (Shiltsev, 2009).
• Fast inference: Surrogate operators (MIFNO), low-step Heun solvers, and prior-guided patching ensure practical run-times for large areas or batch generation (Perrone et al., 1 Apr 2025, Dhaouadi et al., 13 Nov 2025, Ding et al., 7 Feb 2026).

Limitations:

• Diffusion-based hallucination: Absent or small targets may be invented; axis-aligned crops can include distracting context, introducing segmentation noise (Sethi et al., 20 Apr 2026).
• High computational footprint: Training (especially for HEGGS or large GrounDiff variants) demands extensive GPU resources.
• Some models encode regional/geometric bias, and generalization to unseen domains may require re-training or transfer of priors (Jung et al., 2024, Dhaouadi et al., 13 Nov 2025).

6. Domain-Specific Applications and Impact

Seismology and Engineering: GrounDiff models facilitate rapid, spectrum- and energy-compatible ground motion scenario generation, enabling robust probabilistic seismic hazard analyses and design assessments for nuclear facilities or infrastructure (Perrone et al., 1 Apr 2025, Jung et al., 2024, Ding et al., 7 Feb 2026).

Remote Sensing and Geospatial Analysis: GrounDiff provides a framework for DSM to DTM translation, ground structure filtering, and high-fidelity road reconstruction, augmenting LiDAR or photogrammetry workflows and setting new benchmarks in accuracy and surface smoothness (Dhaouadi et al., 13 Nov 2025).

Computer Vision/Object Localization: Text-conditioned and hybrid GrounDiff variants empower robust localization and segmentation in remote sensing (DiffuSAM), natural imagery, or multi-view synthesis, outperforming single-step segmentation or GAN-based approaches (Sethi et al., 20 Apr 2026, Chen et al., 2023, Ye et al., 2024).

Physical Ground Motion Modeling: ATL/fractal random-walk models underpin the theoretical understanding of scale-dependent ground diffusion, informing engineering risk assessments and finite-element models for large-scale facilities (Shiltsev, 2009).

This suggests that GrounDiff, as both a modeling paradigm and set of architectural principles, provides a generalizable and effective approach for denoising, surface estimation, alignment, and generative modeling in earth sciences, vision, and multi-modal data domains. Its core strength lies in leveraging explicit conditioning, iterative denoising, and confidence-guided mechanisms to produce high-fidelity, interpretable, and scale-aware outputs across a wide range of high-impact geospatial and vision tasks.