TerrainGraphNet: Terrain-Constrained Graph Reasoning for Landslide Segmentation

Conference: ECCV 2026
Paper: ECCV Official
Area: Remote Sensing
Keywords: Landslide Segmentation, Terrain-Aware Learning, Graph Attention Networks, Digital Elevation Model, Remote Sensing Imagery

TL;DR

Addressing the limitation of conventional landslide segmentation where Digital Elevation Models (DEM) are treated as generic appearance channels without geomorphological constraints, TerrainGraphNet introduces terrain-modulated feature interaction and slope-constrained sparse graph reasoning to suppress boundary leakage and topological fragmentation.

Background & Motivation

Landslides represent one of the most destructive natural hazards globally, making accurate and rapid boundary delineation from high-resolution satellite imagery essential for disaster response and risk mitigation. Although deep convolutional networks and vision transformers have driven remarkable progress in Earth observation semantic segmentation, they continue to suffer from fragmented predictions and boundary leakage in complex terrain. Landslides are intrinsically geomorphological phenomena governed by gravity and terrain geometry: their spatial extent stretches along slopes, maintains topological continuity along gradient paths, and halts at natural ridges or valley discontinuities. However, landslide scars are frequently masked by heterogeneous vegetation cover, variable illumination, and deep mountain shadows, rendering purely appearance-based pixel classification vulnerable to spectral ambiguity.

To incorporate geometric guidance, standard approaches ingest aligned Digital Elevation Models (DEM) as auxiliary inputs via channel concatenation or uniform multi-branch fusion. Treating elevation as an extra visual channel exhibits a fundamental flaw: it assumes uniform cross-modal contributions across all spatial locations, overlooking the dynamic balance between steep cliffs where topography dominates and flat plains where visual textures prevail. Crucially, conventional fusion fails to impose physical topological constraints during long-range context aggregation, frequently causing feature representations to diffuse across natural ridges into non-landslide areas.

This work reframes landslide segmentation as a terrain-conditioned structured prediction problem, ensuring that feature propagation diffuses freely along terrain-consistent regions while strictly respecting geomorphological discontinuities. The core idea is to employ dual visual-terrain encoding coupled with adaptive soft-gating for terrain-modulated feature interaction, and to construct a sparse topological graph governed jointly by feature similarity and DEM slope continuity, propagating information via residual graph attention to delineate topologically consistent landslide boundaries.

Method

Overall Architecture

The TerrainGraphNet pipeline comprises four primary stages: joint visual-terrain encoding (extracting multi-scale optical semantics and DEM elevation gradient), terrain-modulated feature interaction (adaptively regulating visual and terrain features via bidirectional spatial gates), terrain-aware graph reasoning (constructing a sparse graph constrained by slope continuity and propagating information along terrain topology), and hierarchical residual decoding to produce the final segmentation mask.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input Data<br/>RGB Imagery + DEM Elevation"] --> B["Joint Visual-Terrain Encoding<br/>SAM/CNN and Slope Gradient"]
    B --> C["Terrain-Modulated Feature Interaction<br/>Adaptive Bidirectional Gating"]
    C --> D["Terrain-Aware Graph Construction & Propagation<br/>Slope-Constrained Affinity & Residual GAT"]
    D --> E["Hierarchical Residual Decoding<br/>Multi-Scale Skip Connections & Auxiliary Loss"]
    E --> F["Output Predicted Mask<br/>High-Fidelity Landslide Delineation"]

Key Designs

1. Joint Visual-Terrain Encoding: Capturing Global Context, Fine-Grained Textures, and Slope Gradients

To handle dramatic scale variations in satellite imagery and capture fine-grained boundaries, visual encoding utilizes a dual-stream architecture: a global branch employs a frozen pretrained Segment Anything Model (SAM) encoder to extract broad semantic representations \(F_g \in \mathbb{R}^{H' \times W' \times C}\), while a local branch processes medium-resolution images with a multi-scale CNN at resolutions 64, 128, and 256 to harvest fine-scale textures \(F_l\). Global and local features are concatenated and projected into a joint visual representation \(F_v\). In the terrain branch, recognizing that raw elevation values alone do not directly indicate physical slope instability, the framework extracts latent terrain features \(F_t\) while explicitly computing the elevation gradient magnitude across spatial coordinates:

\[S(x, y) = \sqrt{\left(\frac{\partial D}{\partial x}\right)^2 + \left(\frac{\partial D}{\partial y}\right)^2}\]

The resulting slope map highlights steep escarpments and terrain breaks, providing an explicit geomorphological prior for downstream cross-modal gating and topological graph construction.

2. Terrain-Modulated Feature Interaction: Spatial Adaptive Bidirectional Soft Gating

Uniform feature fusion assumes identical cross-modal utility across all pixels. In reality, terrain geometry dominates in steep slide-prone areas, whereas optical texture provides stronger discrimination in flat or homogeneous regions. To prevent naive feature dilution, the module learns two lightweight \(1 \times 1\) projections \(W_t\) and \(W_v\) to generate soft spatial gating masks \(A_t = \sigma(W_t F_t)\) and \(A_v = \sigma(W_v F_v)\), subsequently modulating representations with an additive identity bias:

\[\tilde{F}_v = F_v \odot (1 + A_t), \quad \tilde{F}_t = F_t \odot (1 + A_v)\]

The additive identity ensures that base representations are never zeroed out, but selectively boosted where the complementary modality exhibits high confidence. The modulated streams are concatenated and projected by layer \(\psi\) into a unified representation \(F\), imbuing each spatial coordinate with terrain-guided inductive bias.

3. Terrain-Aware Graph Construction and Propagation: Topology-Aligned Reasoning Under Slope Continuity

Because shadows and patchy vegetation often disrupt visual continuity within an identical landslide scar, standard convolutional kernels fail to bridge non-local fractures. To propagate structural evidence across visually disconnected regions, each spatial location is treated as a graph node. The affinity weight between node \(i\) and node \(j\) is jointly defined by multimodal feature distance and physical slope discrepancy:

\[w_{ij} = \exp\left(-\alpha \|f_i - f_j\|_2^2 - (1 - \alpha) \|s_i - s_j\|_2^2\right)\]

where \(\alpha \in [0, 1]\) balances representation similarity with slope continuity. Nodes with compatible visual-terrain features and consistent slope gradients form strong topological edges even across spatial gaps; conversely, nodes separated by steep ridges or deep valleys incur sharp slope differences and are disconnected. For each node, the top-\(K\) affinities establish a sparse graph. Message passing across \(L\) graph attention layers (GAT) updates node states anisotropically, integrated via an outer residual connection \(F' = F + h^{(L)}\). This mechanism functions as a learned terrain-weighted smoothness prior that aggregates features within coherent geomorphological units while blocking leakage across boundaries.

Loss & Training

The network is optimized end-to-end using a compound objective combining Dice loss for region overlap and Binary Cross-Entropy (BCE) for pixel-level stability:

\[\mathcal{L}_{main} = 0.7 \mathcal{L}_{Dice} + 0.3 \mathcal{L}_{BCE}\]

To ensure multi-scale feature consistency and stabilize graph gradient propagation, deep supervision is applied to intermediate decoder stages:

\[\mathcal{L} = \mathcal{L}_{main} + \lambda_{aux} \mathcal{L}_{aux}\]

with weighting coefficient set to \(\lambda_{aux} = 0.4\).

Key Experimental Results

Main Results

On the benchmark Bijie dataset (high-resolution satellite imagery with ground-truth DEM) and the Luding earthquake landslide dataset (where ground-truth DEM is absent and monocular depth estimated by Depth Anything V2 is utilized as terrain input), TerrainGraphNet demonstrates superior quantitative performance against leading semantic segmentation models and specialized landslide baselines:

Dataset	Model	Precision (%)	Recall (%)	F1 (%)	IoU (%)	mIoU (%)	Betti Error ↓	Connectivity Error ↓	HD95 ↓	Boundary F1 ↑
Bijie	U-Net	84.64	63.51	72.57	56.94	76.20	2.9500	0.8540	27.84	35.12
Bijie	SegFormer	84.71	75.87	80.05	66.74	81.32	0.3060	0.0431	19.89	43.47
Bijie	FFS-Net (Prev. SOTA)	87.64	86.61	87.12	77.19	87.15	0.0862	0.0474	15.06	52.22
Bijie	TerrainGraphNet (Ours)	88.84	87.80	88.32	79.08	88.26	0.0776	0.0172	14.42	57.31
Luding	U-Net	80.34	75.37	77.78	63.63	79.62	1.0400	0.8200	22.93	36.31
Luding	SegFormer	81.52	83.73	82.61	70.37	81.86	0.3500	0.3300	25.21	39.34
Luding	FFS-Net (Prev. SOTA)	86.00	86.85	86.42	76.09	84.73	0.9500	0.6400	25.69	50.04
Luding	TerrainGraphNet (Ours)	87.06	89.40	88.22	78.92	87.18	0.3400	0.2400	21.94	52.03

On the challenging Landslide4Sense benchmark, TerrainGraphNet similarly achieves the highest IoU of 54.20% and mIoU of 76.39%, surpassing FFS-Net (53.84% IoU).

Ablation Study

A step-by-step module ablation on the Bijie dataset (BB: SAM visual backbone, LE: local CNN encoder, TE: terrain encoder, TMFI: terrain-modulated feature interaction, HD: hierarchical decoder, TAGR: terrain-aware graph reasoning):

Config	BB	LE	TE	TMFI	HD	TAGR	Precision (%)	Recall (%)	F1 (%)	IoU (%)	mIoU (%)	Note
Baseline Backbone	✓	✗	✗	✗	✗	✗	80.93	83.65	82.30	69.92	81.34	Frozen SAM global appearance only
+ Local Textures	✓	✓	✗	✗	✗	✗	84.65	80.41	82.48	70.20	82.94	Multi-scale CNN refines boundaries
+ Terrain & Gating	✓	✓	✓	✓	✗	✗	88.14	79.75	83.75	72.04	83.72	Ingests DEM and bidirectional gating
+ Hierarchical Decoder	✓	✓	✓	✓	✓	✗	85.78	85.39	85.58	74.80	85.82	Multi-scale skip upsampling
Full Model	✓	✓	✓	✓	✓	✓	88.84	87.80	88.32	79.08	88.26	TAGR graph reasoning adds +4.28% IoU

Evaluating fusion mechanisms separately confirms that substituting TMFI with element-wise addition reduces IoU to 77.32%, and element-wise multiplication drops IoU to 76.21%. Removing TAGR entirely with standard concatenation (Concat w/o TAGR) degrades IoU to 74.55% and increases HD95 from 14.42 to 18.20, confirming the primary role of graph reasoning in boundary sharpening.

Key Findings

• TAGR drives the largest structural gain: Ablation shows TAGR alone contributes a 4.28% jump in IoU (74.80% to 79.08%) while reducing Connectivity Error (CE) from 0.1515 to 0.0172, proving that slope-guided topological propagation successfully bridges vegetation gaps.
• Sensitivity to graph density: Analysis of neighbor size indicates \(K=4, L=2\) achieves the best trade-off. Expanding to \(K=16\) degrades IoU to 77.16%, demonstrating that over-connected graphs introduce irrelevant contextual noise and blur geomorphological boundaries.
• Robustness to estimated monocular depth: On the Luding benchmark where true DEMs are missing, deriving slope maps from Depth Anything V2 yields an impressive 78.92% IoU and 52.03 Boundary F1, confirming that the framework generalizes seamlessly to pseudo-depth representations.

Highlights & Insights

• Physics-grounded topological affinity: Unlike standard GCNs that rely purely on learned feature proximity, the framework incorporates an explicit elevation gradient difference as an exponential damping penalty. This dual-constraint approach introduces minimal computational overhead (TAGR requires only 0.54 GFLOPs and 3.61 ms inference time) while preventing non-local diffusion across physical terrain boundaries.
• Backbone-agnostic adaptability: Validating with ResNet-50 demonstrates a +8.51% IoU boost (65.36% to 73.87%) and +6.66 Boundary F1 gain, confirming that the terrain-constrained graph framework operates as a plug-and-play enhancement independent of the underlying vision backbone.

Limitations & Future Work

• Omission of broader hydrological and geological factors: While slope magnitude captures primary gravity-driven boundaries, real-world landslides depend on rainfall accumulation indices, tectonic fault proximity, and lithological properties. Incorporating multidimensional geo-environmental variables into graph reasoning represents a promising direction.
• Geometric distortion in monocular depth estimation: Relying on monocular depth foundation models under extreme shadow or steep canyon angles can introduce perspective distortion, which may occasionally corrupt the derived slope affinity in emergency disaster settings lacking survey DEMs.

Related Work & Insights

• vs FFS-Net (IEEE TGRS 2023): FFS-Net relies on multi-scale cross-attention on regular Euclidean grids to fuse optical and DEM features; TerrainGraphNet operates over non-Euclidean topological graphs constrained by physical slope gradients, outperforming FFS-Net on Bijie by +1.89% IoU, +1.20% F1, and +5.09 Boundary F1.
• vs SegFormer (NeurIPS 2021): SegFormer extracts hierarchical global context via self-attention but lacks geomorphological inductive bias, causing fragmentation across shadowed zones; TerrainGraphNet integrates terrain-weighted smoothness priors to eliminate topological breaks and boundary leakage.

Rating

• Novelty: ⭐⭐⭐⭐☆ Reformulates landslide segmentation as terrain-constrained graph reasoning with a physics-informed affinity function.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Evaluated across three diverse datasets with extensive geometric/topological metrics, monocular depth transfer, and backbone generalization tests.
• Writing Quality: ⭐⭐⭐⭐⭐ Rigorous motivation, clear pipeline visualization, and well-grounded mathematical formulations.
• Value: ⭐⭐⭐⭐⭐ Highly valuable for automated geohazard mapping and remote sensing disaster response.

Related Papers

• [ECCV 2026] LineGraph2Road: Structural Graph Reasoning on Line Graphs for Road Network Extraction
• [ECCV 2026] Online Reasoning Video Object Segmentation
• [ECCV 2026] From Reasoning to Pixels: Grounded Medical Multimodal LLMs for VQA and Segmentation
• [ECCV 2026] STAC: Selective Spatiotemporal Aggregation and Compression for Video Reasoning Segmentation
• [AAAI 2026] Graph-of-Mark: Promote Spatial Reasoning in Multimodal Language Models with Graph-Based Visual Prompting