DIGSFNet: Deformation-Integrity-Guided Symmetric Fusion Network for High-Risk Landslide Extraction from Multi-Source Remote Sensing Images

Highlights

What are the main findings?

• DIGSFNet treats InSAR deformation data and optical imagery as equal information sources through symmetric, dynamically sparse cross-modal fusion. When trained separately on the Nanning-HHLS dataset and the public HAEFNet benchmark, it achieved mIoU values of 83.57 percent and 78.92 percent, outperforming the strongest comparison methods by 2.63 and 3.68 percentage points, respectively.
• The deformation-integrity decoder incorporates deformation continuity and boundary-gradient consistency into the segmentation process, producing more complete and physically consistent landslide masks and improving the Boundary F1 score by 5.52 percentage points over symmetric fusion alone.

What are the implications of the main findings?

• Explicitly embedding InSAR deformation physics improves the reliability and interpretability of high-risk landslide mapping by reducing missed landslides and helping suppress false alarms caused by non-landslide deformation.
• The INT8 lightweight student model achieves 79.24 percent mIoU at 218 FPS, with a latency of 4.6 ms and GPU memory consumption of only 0.54 GB, supporting efficient operational surveys over large mountainous areas.

Abstract

High-risk landslide extraction from remote sensing imagery is a fundamental task for geological disaster prevention, emergency response, and land-use planning in mountainous regions. Although deep-learning semantic segmentation has substantially advanced landslide detection from optical imagery, existing methods still suffer from three critical limitations: (i) Interferometric Synthetic Aperture Radar (InSAR) deformation data are treated as auxiliary channels and dominated by optical features during fusion; (ii) predicted masks exhibit fragmented boundaries and incomplete delineation due to the absence of deformation continuity constraints reflecting the physical coherence of slope movements; and (iii) heavy Transformer backbones hinder practical deployment over large areas. To address these issues, we propose a Deformation-Integrity-Guided Symmetric Fusion Network (DIGSFNet) for high-risk landslide extraction from InSAR and optical imagery. The framework consists of three components: a Symmetric Deformation-Aware Encoder (SDAE) that treats InSAR and optical modalities as equal information sources through modality-aware adapters and dynamic sparse cross-modal fusion; a Deformation Integrity Prior Decoder (DIPD) that imposes deformation continuity and boundary-gradient consistency as physical priors to enforce mask completeness and boundary accuracy; and a Lightweight Deployable Student Network (LDSN) obtained via cross-modal knowledge distillation and INT8 quantization for efficient inference. Experiments on the Nanning High-Hazard Landslide Segmentation (Nanning-HHLS) dataset and the public HAEFNet benchmark covering the Qinghai–Tibet–Sichuan landslide-prone regions show that the full DIGSFNet achieves state-of-the-art extraction accuracy, reaching 83.57% and 78.92% mIoU on the two datasets and surpassing the strongest competing method by 2.63 and 3.68 percentage points with a Recall of 91.48% on Nanning-HHLS, while its distilled lightweight student retains 79.24% mIoU at 218 frames per second after INT8 quantization, enabling efficient large-area operational deployment.

1. Introduction

Landslide hazards are among the most devastating geological disasters in mountainous regions worldwide [1,2], posing severe threats to human life, infrastructure integrity, and sustainable economic development. Mountainous terrains are inherently susceptible to slope instability due to steep gradients, complex lithological structures, intense precipitation, and seismic activity [3]. Among the broad spectrum of landslide phenomena, a particularly critical yet underexplored category is high-risk landslides: slopes that have already undergone measurable but often subtle deformation and exhibit precursory signs of instability, yet have not fully developed into catastrophic failures [1,4]. In this work, we adopt an explicit, criteria-based working definition: a high-risk landslide is a slope that simultaneously (i) exhibits spatially coherent InSAR-measured line-of-sight (LOS) deformation exceeding $|v_{LOS}| \ge 10$ mm/yr (equivalently, cumulative LOS displacement $\ge 20$ mm over the 2020–2023 observation window), indicating ongoing movement; (ii) presents corroborating morphological or spectral instability evidence in optical imagery (e.g., scarps, tension cracks, vegetation anomalies); and (iii) has not yet undergone complete catastrophic failure. This category overlaps with, but is not identical to, the terms “active landslide” [5] and “latent/incipient landslide” used in previous studies; hereinafter these terms are used only when referring to the corresponding external datasets or cited works, and we note that “risk” is used here in the operational sense of a slope threatening lives or infrastructure rather than in the formal hazard–vulnerability–exposure sense of quantitative risk assessment. The definition excludes both post-failure deposits without present-day motion, which fail criterion (i), and landslide-susceptible yet currently stable slopes. The early identification and precise delineation of such high-risk landslides constitute a fundamental task for proactive disaster prevention, early warning, emergency response planning, and rational land-use regulation. In this context, the synergistic combination of Synthetic Aperture Radar Interferometry (InSAR), which captures millimeter-level surface deformation indicative of slow and progressive slope movement [4,6,7], and high-resolution optical remote sensing imagery, which provides rich spectral and textural information about surface morphology, offers an ideal multi-source data foundation for high-risk landslide extraction.

In recent years, deep learning-based semantic segmentation methods have demonstrated remarkable success in automated landslide detection from remote sensing images [8,9]. Encoder–decoder architectures such as U-Net [10], DeepLabV3+ [11], and SegFormer [12] have established strong baselines for pixel-wise landslide mapping by hierarchically extracting multi-scale contextual features. Various attention mechanisms, including channel attention, spatial attention, and self-attention modules, have been incorporated to enhance the discriminative capacity of landslide features against complex background environments [13,14]. Multi-scale feature fusion strategies, such as Feature Pyramid Networks and Atrous Spatial Pyramid Pooling, further improve the representation of landslides with varying spatial extents [15,16]. Moreover, recent studies have begun to explore InSAR–optical fusion approaches for landslide identification, recognizing that deformation signals provide critical evidence for distinguishing high-risk landslides exhibiting ongoing creep from stable terrain and historical landslide deposits [4,17,18,19,20].

However, existing methods still suffer from several critical failure modes that limit their practical applicability in high-risk landslide extraction. First, current multi-source fusion methods predominantly adopt an asymmetric modality treatment strategy, where InSAR deformation maps are simply concatenated as additional input channels to an optical-dominant backbone network [21,22]. This concatenation-based approach inherently constrains the deformation modality to a subordinate role, causing the network to inadequately learn and represent the unique displacement patterns, velocity gradients, and spatial clustering characteristics embedded within InSAR data—features that are essential for distinguishing slowly deforming high-risk slopes from optically similar but stable terrain. Second, predicted landslide masks frequently exhibit fragmented boundaries and incomplete interior delineation, particularly for large-scale or slowly creeping high-risk landslides whose surface manifestations are subtle and diffuse. This deficiency arises because conventional segmentation losses (cross-entropy, Dice [23]) lack explicit spatial consistency constraints derived from deformation physics; in reality, a high-risk landslide body typically displays continuous and coherent displacement fields, and its boundaries correspond to distinct deformation gradient discontinuities—information that remains unexploited in existing frameworks. Third, state-of-the-art methods increasingly rely on heavy Transformer backbones [12,24] with substantial parameter counts and computational overhead, rendering them impractical for the massive data volumes required in large-area landslide hazard surveys covering thousands of square kilometers.

We deliberately scope this work to high-risk landslide extraction rather than general landslide identification: the former targets slopes where intervention can still precede failure, and it is the sub-problem in which InSAR deformation constitutes decisive rather than auxiliary evidence. General landslide identification, which additionally includes historical and stabilized landslide bodies, weakens the role of the kinematic criterion, since stabilized deposits produce no present-day deformation; the proposed framework remains applicable in principle, but its deformation-centered design advantages would be diluted. Likewise, the emphasis on mountainous regions reflects the setting of both study areas and of the operational surveys motivating this work, not an intrinsic architectural limit: the framework applies wherever co-registered InSAR–optical pairs are available.

To address the above challenges, we propose a novel Deformation-Integrity-Guided Symmetric Fusion Network (DIGSFNet) for high-risk landslide extraction from multi-source remote sensing images in mountainous regions. The core idea of DIGSFNet is to establish a symmetric and balanced multi-modal learning paradigm that elevates InSAR deformation data from an auxiliary channel to an equal information source, while simultaneously embedding deformation integrity priors into the decoding process to enforce physically consistent landslide mask predictions. The proposed framework comprises three synergistic components: a Symmetric Deformation-Aware Encoder that achieves modality-equitable feature extraction through dedicated adapters and dynamic sparse cross-modal interaction, a Deformation Integrity Prior Decoder that incorporates InSAR-derived physical constraints for complete and boundary-accurate mask generation, and a Lightweight Deployable Student Network that enables efficient large-area inference through knowledge distillation and post-training quantization.

The main contributions of this paper are summarized as follows:

• Symmetric Deformation-Aware Encoder (SDAE). We design a symmetric dual-branch encoder with modality-aware adapters and a dynamic sparse cross-modal fusion mechanism, which treats InSAR and optical data as equal information sources and enables comprehensive cross-modal feature interaction while preserving modality-specific characteristics.
• Deformation Integrity Prior Decoder (DIPD). We propose a decoder that explicitly leverages InSAR deformation continuity and boundary gradient consistency as physical prior constraints, guiding the network to produce spatially coherent masks with accurate and complete boundaries that align with the underlying deformation mechanics.
• Lightweight Deployable Student Network (LDSN). We develop a lightweight student network trained through cross-modal knowledge distillation from the full DIGSFNet teacher, combined with INT8 post-training quantization, achieving a favorable accuracy–efficiency trade-off suitable for operational high-risk landslide surveys.

2. Related Work

2.1. Landslide Detection from Remote Sensing Images

Landslide detection from remote sensing imagery has undergone a significant paradigm shift over the past two decades, transitioning from manual visual interpretation toward automated and semi-automated extraction approaches. Early efforts predominantly relied on traditional machine learning classifiers, such as support vector machines (SVM) [25] and random forests (RF) [26], operating on handcrafted spectral, textural, and morphological features derived from optical satellite images. While these methods achieved reasonable accuracy in specific study areas, their performance was highly dependent on the quality of feature engineering and often degraded under complex terrain conditions or heterogeneous land cover backgrounds [8]. Object-based image analysis (OBIA) combined with rule-based classification further improved landslide mapping by incorporating spatial context and shape attributes [27], yet the transferability of such rule sets across different geographic regions remained limited.

The advent of deep learning has substantially advanced the state of the art in landslide detection. Convolutional neural networks (CNNs), particularly encoder–decoder architectures such as U-Net [10] and DeepLab [11], have demonstrated strong capability in pixel-level landslide segmentation from high-resolution optical imagery [9,28]. Several studies have extended these architectures with attention mechanisms [14] or multi-scale feature aggregation modules [29] to better capture the irregular morphology of landslide bodies. More recently, transformer-based models have also been explored for landslide mapping, leveraging global contextual modeling to distinguish landslides from spectrally similar disturbed surfaces [12,24]. Beyond optical data, synthetic aperture radar (SAR) imagery has been employed for landslide detection owing to its all-weather and day-and-night acquisition capability [21]. A few studies have attempted to combine optical and SAR data for improved detection performance [30], and interferometric SAR (InSAR) techniques have proven effective in measuring slow surface deformation that serves as a precursor signal of landslide activity [6,7].

The integration of optical and SAR/InSAR observations for landslide studies has itself developed into a substantial body of work [3,31]. Amplitude-based optical–SAR fusion has been used primarily for post-event landslide mapping under cloud cover [30], while multi-temporal interferometric techniques have supported landslide inventory updating and motion monitoring at regional scales [4,7,17]. More recently, deep-learning methods have begun to consume InSAR deformation products directly: multi-channel CNNs fusing InSAR, optical, and topographic inputs for active-landslide detection [18], wide-area InSAR-based landslide detection considering topographic features [19], hybrid self-distillation architectures for InSAR deformation detection [20], and the evidence-theoretic InSAR–optical–DEM fusion of HAEFNet [5].

Despite these advances, existing InSAR-aware deep-learning methods remain dominated by asymmetric fusion designs in which deformation enters as an auxiliary channel of an optical-centric network, and the majority of general-purpose landslide extraction methods are still designed primarily for optical imagery. How to effectively fuse the information contained in optical imagery—geometric cues, i.e., the shape, texture, and morphological patterns of slope surfaces (scarps, tension cracks, hummocky terrain, displaced masses), and spectral cues, i.e., the multi-band reflectance signatures revealing exposed soil, vegetation removal or stress, and moisture variation—with the displacement signals from InSAR to achieve accurate extraction of high-risk landslides remains an open and challenging problem.

2.2. Multi-Modal Fusion for Remote Sensing Segmentation

Multi-modal data fusion is a central topic in remote sensing image analysis, as the complementary information carried by different sensor modalities can substantially enhance the discriminability and robustness of segmentation models. Existing fusion strategies are generally categorized into three levels: early fusion, middle fusion, and late fusion. Early fusion, typically realized through channel-wise concatenation of multi-source inputs before feeding them into a shared encoder [22], is straightforward to implement but implicitly assumes that heterogeneous modalities share a common feature space, which may not hold for physically distinct data types. Late fusion, in contrast, processes each modality with an independent branch and merges predictions at the decision level [32], preserving modality-specific representations but potentially missing fine-grained cross-modal interactions. Middle fusion, also termed feature-level fusion, strikes a balance by exchanging information at intermediate layers through mechanisms such as cross-attention [33], gated fusion modules [34], or bi-directional feature propagation [35].

In the broader computer vision community, symmetric dual-encoder architectures have gained popularity for RGB+X semantic segmentation tasks, where X denotes a complementary modality such as depth, thermal, or event data. Representative works include CMX [36], which introduces a cross-modal feature rectification module within a symmetric transformer backbone, and TokenFusion [37], which dynamically exchanges tokens between parallel branches. Asymmetric designs, where a lightweight auxiliary branch processes the secondary modality, have also been proposed to reduce computational overhead while retaining fusion benefits [38]. Parameter-efficient fusion strategies, such as adapter-based and prompt-tuning approaches, have further emerged as attractive alternatives that avoid retraining large backbones from scratch [39].

Beyond fusion architectures, spatial-prior-guided decoding has proven effective when one modality carries strong structural regularity. In particular, PDFNet [40] introduced a depth integrity-prior together with a fine-grained patch strategy for high-precision dichotomous image segmentation, exploiting the spatial coherence of scene depth to enforce complete and boundary-accurate masks. This design philosophy directly inspired the decoder proposed in this work; Section 4.3.1 details what is inherited from PDFNet and what is specific to the InSAR deformation domain.

Within the remote sensing domain, multi-modal fusion methods have been extensively studied for RGB combined with digital surface models (DSM) or digital elevation models (DEM) in urban scene parsing [41], and for optical–SAR fusion in land cover classification [30]. However, there is a notable scarcity of methods specifically designed to address the fusion of InSAR deformation maps with optical imagery for landslide extraction. The InSAR deformation modality presents unique challenges: it exhibits sparse and noisy spatial patterns, possesses fundamentally different physical semantics compared to optical reflectance, and its informative regions may be spatially misaligned with the visible boundaries of landslide bodies. These characteristics call for dedicated fusion architectures that can selectively leverage deformation integrity cues while preserving the fine-grained spatial details available in optical data.

2.3. Lightweight and Efficient Segmentation Networks

The deployment of semantic segmentation models in operational landslide monitoring scenarios imposes stringent requirements on computational efficiency and memory footprint, particularly when large-area mapping or near-real-time processing is desired. Lightweight backbone architectures, such as MobileNetV2 [42], EfficientNet [43], and the recently proposed LSNet [44], have been developed to achieve competitive accuracy with significantly reduced parameters and floating-point operations through design principles including depthwise separable convolutions, inverted residuals, and neural architecture search. These backbones have been integrated into segmentation frameworks such as DeepLabV3+ [11] and BiSeNet [45] for efficient remote sensing scene parsing.

Beyond architecture design, knowledge distillation has emerged as a powerful model compression strategy, wherein a compact student network is trained to mimic the output or intermediate representations of a larger teacher model [46]. This paradigm has been successfully applied to single-modality remote sensing segmentation tasks, yielding substantial speedups with marginal accuracy degradation [47]. Model quantization techniques, including INT8 and FP16 mixed-precision inference, provide an orthogonal avenue for acceleration by reducing the numerical precision of network weights and activations [48]. Nevertheless, the vast majority of existing lightweight segmentation works focus on single-modality inputs and do not consider the unique challenges associated with cross-modal distillation in multi-source fusion settings. In particular, how to distill knowledge from a dual-branch multi-modal teacher into a compact student that preserves cross-modal complementarity for InSAR–optical landslide extraction has yet to be systematically explored.

3. Data

3.1. Dataset Overview

We evaluate DIGSFNet on two multi-modal landslide segmentation datasets that pair InSAR deformation products with optical satellite imagery, covering geographically and geomorphologically distinct regions.

Nanning High-Hazard Landslide Segmentation (Nanning-HHLS). This dataset targets the Nanning municipality (administrative area approximately 22,112 km²) in Guangxi Zhuang Autonomous Region, southern China, spanning approximately 107.75°E–108.85°E and 22.22°N–23.53°N over karst-dominated subtropical terrain. Ten-band Sentinel-2 Level-2A surface-reflectance imagery (B2–B4, B5–B8A, B11–B12; resampled to 10 m) provides the optical modality, and Sentinel-1 SBAS-InSAR cumulative line-of-sight (LOS) deformation maps (January 2020–December 2023, 96 ascending-track acquisitions, geocoded to 10 m) provide the InSAR modality. The optical layer is a cloud-free composite generated from 34 Sentinel-2 scenes acquired during the dry seasons (October–March) of 2021–2023, chosen to be temporally consistent with the InSAR observation window. We emphasize that the task is not post-event landslide mapping: the targets are slopes under ongoing, pre-failure deformation, so there is no discrete failure event partitioning the data into before/after epochs; the InSAR modality summarizes a continuous observation window and the multi-temporality of the dataset resides entirely in the 96-acquisition InSAR stack. The two modalities are co-registered to sub-pixel accuracy. High-risk landslide areas were identified through a multi-source protocol combining field survey records, historical inventory databases, InSAR deformation thresholds (cf. criterion (i) in Section 1), and expert visual interpretation; three geologists annotated each site independently and consensus labels were established by majority voting. The pixel-level inter-rater agreement over the three independent annotations is Fleiss’ $\kappa =0.79$, and boundary disagreements exceeding 2 pixels along the perimeter were resolved in joint re-inspection sessions using the InSAR deformation gradient as the arbitration reference; historical-inventory polygons without a present-day InSAR signal were labeled as background. The dataset comprises 3842 $512\times 512$ patches extracted with an overlapping sliding window, containing 467 pixel-level annotated high-risk landslide instances and a landslide-to-background pixel ratio of approximately 1:18. Annotated instance areas range from 3200 m² to 2.8 km², with a median of 21,400 m², and the land cover of the deforming slopes is dominated by forest/shrub (62%), cropland (21%), and bare karst rock (15%) (ESA WorldCover 10 m); because the targets have not failed, these figures describe the current cover of the deforming slopes rather than pre-/post-event states. The dataset is split into training (60%), validation (20%), and testing (20%) subsets using spatially stratified random sampling.

HAEFNet Multi-Modal Active-Landslide Dataset. This publicly available benchmark accompanies the HAEFNet framework [5] and covers active-landslide-prone regions in the Qinghai–Tibet Plateau and Sichuan, western China. It provides spatially co-registered Sentinel-1 InSAR mean LOS deformation rate maps, Sentinel-2 RGB imagery (10 m), and Copernicus GLO-30 DEM (30 m). Following the design rationale discussed in Section 6.2, the DEM channel is not used by DIGSFNet so that the experimental configuration on the two datasets remains identical. Expert annotations of active landslides were guided by InSAR deformation signals, optical textures, and topographic context. The released dataset contains 8440 sample pairs (resampled to $512\times 512$ px) with 1013 annotated active landslide instances and a positive-to-negative sample ratio of approximately 1:1.5. We adopt the official 7:1:2 train/validation/test split.

3.2. Dataset Comparison

Table 1. Comparison of the Nanning-HHLS and HAEFNet datasets used in this study.

Attribute	Nanning-HHLS (This Work)	HAEFNet (Public Benchmark) [5]
Geographic region	Nanning, Guangxi, S China	Qinghai–Tibet Plateau and Sichuan, W China
Task	High-risk landslide segmentation	Active landslide detection
Input modalities	Optical + InSAR	Optical + InSAR (+DEM, not used)
Optical source	Sentinel-2 L2A, 10 bands (10 m)	Sentinel-2 RGB (10 m)
InSAR source/product	Sentinel-1 SBAS-InSAR, cumulative LOS displacement	Sentinel-1, mean LOS deformation rate
Spatial resolution	10 m	10 m
Patch size	512 × 512 px	512 × 512 px
Total patches	3842	8440
Landslide instances	467	1013
Annotation type	Pixel-level binary mask	Pixel-level binary mask
Train/Val/Test split	60/20/20 (spatially stratified)	7:1:2 (official split)

summarizes the key attributes of both datasets. The two benchmarks differ substantially in climate, lithology, and landslide mechanism—Nanning-HHLS covers humid subtropical karst terrain with dense vegetation, whereas the HAEFNet dataset spans high-altitude, sparsely vegetated alpine environments—making them complementary testbeds for evaluating how well the architecture generalizes when independently trained on geomorphologically distinct regions. Crucially, both datasets provide spatially co-registered InSAR–optical pairs, permitting the full dual-branch fusion pipeline of DIGSFNet (SDAE + DIPD) to be evaluated without architectural simplification.

Two further differences between the benchmarks deserve explicit discussion. First, the class balance differs sharply: Nanning-HHLS has a landslide-to-background pixel ratio of about 1:18, whereas the HAEFNet benchmark is comparatively balanced (1:1.5). During training, the Dice component of the segmentation loss and the overlap-based evaluation metrics (IoU, F1, BF1) are insensitive to the dominance of background pixels, and no re-weighting or over-sampling was applied; the fact that the relative ranking of all compared methods is consistent across the two imbalance regimes (reported in Section 5.3) indicates that our conclusions are not artifacts of a particular class balance, and this is also why accuracy-type metrics are avoided throughout. Second, the released InSAR products differ—cumulative LOS displacement for Nanning-HHLS versus mean LOS velocity for the HAEFNet benchmark. Both are single-channel scalar summaries and are z-score normalized per scene before being fed to the network; since every model is trained and evaluated independently on each dataset and no cross-dataset transfer is claimed, all methods see the identical InSAR product within each comparison, and the product difference does not confound any reported result. For quasi-linear motion the two quantities are related by an approximately constant factor over a fixed observation window, while they diverge for strongly non-linear motion—one more respect in which the two datasets form complementary testbeds.

4. Methodology

4.1. Overall Architecture

The proposed Deformation-Integrity-Guided Symmetric Fusion Network (DIGSFNet) establishes a unified framework for high-risk landslide extraction from multi-source remote sensing imagery, integrating interferometric synthetic aperture radar (InSAR) deformation measurements with high-resolution optical observations. The overall architecture, illustrated in Figure 1, consists of three principal components arranged in a sequential pipeline.

First, the Symmetric Deformation-Aware Encoder (SDAE) receives co-registered InSAR line-of-sight (LOS) deformation maps and optical imagery as parallel inputs, treating both modalities with equal representational status through shared backbone weights augmented by modality-specific adapters, and progressively fuses complementary features via dynamic sparse cross-modal attention at each hierarchical stage. Then, the Deformation Integrity Prior Decoder (DIPD) leverages the physical constraint that high-risk landslide bodies exhibit spatially continuous deformation fields, incorporating this prior knowledge into a coarse-to-fine prediction strategy with dedicated loss functions that enforce intra-region consistency and boundary–deformation alignment. Finally, the Lightweight Deployable Student Network (LDSN) distills the knowledge encoded in the full teacher model (SDAE + DIPD) into a compact inference network suitable for operational deployment, employing cross-modal knowledge distillation at multiple granularities and structural reparameterization for efficient execution on edge devices.

The synergistic design of these three components enables DIGSFNet to achieve both high extraction accuracy and practical deployability, addressing the dual challenges of multi-modal fusion complexity and real-time monitoring requirements in landslide hazard assessment.

4.2. Symmetric Deformation-Aware Encoder (SDAE)

4.2.1. Motivation for Symmetric Modality Treatment

Existing multi-modal fusion approaches for landslide detection typically adopt asymmetric architectures that designate one modality (usually optical imagery) as the primary stream and relegate supplementary modalities (e.g., InSAR) to auxiliary roles through late fusion or simple concatenation. Such asymmetric treatment introduces representational bias: the auxiliary branch receives weaker gradient signals during training and consequently fails to develop discriminative features. For high-risk landslide extraction, InSAR deformation measurements carry irreplaceable physical significance—they directly quantify surface displacement rates that constitute the fundamental criterion for risk classification. Therefore, we argue that a symmetric encoder architecture that assigns equal representational capacity and learning priority to both modalities is essential for fully exploiting the complementary strengths of deformation-based and appearance-based observations. Throughout this paper, “symmetric” refers to this equal representational treatment—identical backbone capacity, identical adapter capacity, and a role-reversible fusion operator (Equation (3))—rather than to fully independent per-modality weights; the backbone weights are shared between the two branches and modality specificity is carried by the adapters, as detailed next. The detailed architecture of the proposed SDAE is illustrated in Figure 2.

4.2.2. Modality-Aware Parameter-Efficient Adapters

To achieve symmetric yet computationally efficient dual-modal feature extraction, SDAE adopts a shared pre-trained backbone augmented with modality-specific lightweight adapters. Specifically, we employ a Swin Transformer [24] pre-trained on ImageNet [49] as the shared feature extractor, which provides rich hierarchical representations while maintaining parameter efficiency through weight sharing. For each modality branch, a set of bottleneck adapters [39] is inserted after each transformer block, enabling modality-specific feature refinement without modifying the shared parameters.

The adapter module follows a bottleneck architecture consisting of a down-projection layer that compresses the feature dimension to a bottleneck dimension r, a nonlinear activation function, and an up-projection layer that restores the original dimensionality. Formally, given an input feature $\mathbf{x}\in {\mathbb{R}}^d$, the adapter forward pass is defined as

$$\mathrm{Adapter}\left(\mathbf{x}\right)=\mathbf{x}+{\mathbf{W}}_{up}·\mathrm{GELU}({\mathbf{W}}_{down}·\mathbf{x}+{\mathbf{b}}_{down})+{\mathbf{b}}_{up}$$

(1)

where ${\mathbf{W}}_{down}\in {\mathbb{R}}^{r\times d}$ and ${\mathbf{W}}_{up}\in {\mathbb{R}}^{d\times r}$ are learnable projection matrices ($r<d$ is the adapter bottleneck dimension), $\mathrm{GELU}(·)$ denotes the Gaussian Error Linear Unit activation, and the residual connection preserves the pre-trained representations. We use $r=64$ (selected by the sensitivity analysis in Section 5.6); adapters are inserted after every Swin Transformer block in both branches, and the total adapter parameter count is 1.53 M, i.e., 1.7% of the full model. This hybrid shared-backbone-plus-adapters design is preferred over the two alternatives—a fully shared encoder without adapters and fully independent dual encoders—for both empirical and conceptual reasons: the shared ImageNet-pretrained weights supply common low-level visual statistics, while the adapters absorb modality-specific refinement at a small fraction of the cost of duplicating the backbone. A direct comparison of the three options confirms this: a fully shared encoder without adapters attains 81.05% mIoU at 85.34 M parameters, fully independent dual encoders attain 83.31% mIoU at 158.9 M parameters, and the proposed hybrid attains 83.57% mIoU at 88.46 M parameters—i.e., the hybrid matches the independent-encoder accuracy at little more than half the encoder parameters. The optical adapter is trained to capture spectral–textural patterns including vegetation anomalies, terrain morphology, and surface roughness variations, while the InSAR adapter specializes in learning deformation-temporal patterns such as displacement velocity gradients, coherence degradation signatures, and phase unwrapping artifacts.

4.2.3. Dynamic Sparse Cross-Modal Fusion

While the parallel adapter branches extract modality-specific features independently, effective landslide extraction requires exploiting inter-modal complementarity. We introduce a Dynamic Sparse Cross-Modal Fusion (DSCF) module at each encoder stage to achieve selective information exchange between the two branches. The core mechanism employs a learnable gating function that dynamically identifies spatially informative locations where cross-modal fusion is most beneficial, thereby avoiding indiscriminate feature mixing that may introduce noise.

Given optical features ${\mathbf{F}}_l^O\in {\mathbb{R}}^{C\times H\times W}$ and InSAR features ${\mathbf{F}}_l^I\in {\mathbb{R}}^{C\times H\times W}$ at encoder stage l, the dynamic sparse gating function computes a binary selection mask:

$${\mathbf{A}}_l=\sigma \left({\mathrm{Conv}}_{1\times 1}([{\mathbf{F}}_l^O;{\mathbf{F}}_l^I])\right),{\mathbf{G}}_l=\mathbb{1}\left[{\mathbf{A}}_l>\tau \right]$$

(2)

where $[·;·]$ denotes channel-wise concatenation, $\sigma (·)$ is the sigmoid function, $\tau$ is an adaptive threshold determined by the top-k percentile of activation values, and $\mathbb{1}[·]$ is the indicator function. The soft sparsity regularizer is defined as ${\mathcal{L}}_{sparse}={\parallel {\mathbf{A}}_l\parallel }_1/(H\times W)$ and therefore acts on the pre-threshold sigmoid activations rather than on the binary gate. Concretely, the hard gate retains the top-$\rho$ fraction of activations with a single global ratio $\rho =0.3$, selected by the sensitivity analysis in Section 5.6 and kept identical across all four encoder stages; $\tau$ is not a free hyperparameter but is recomputed per feature map as the $(1-\rho )$-quantile. Thus, the top-$\rho$ rule fixes the retained fraction of ${\mathbf{G}}_l$, whereas ${\mathcal{L}}_{sparse}$ regularizes the magnitudes of ${\mathbf{A}}_l$ before thresholding. Since the indicator function is non-differentiable, training uses a straight-through estimator (hard top-k selection in the forward pass and sigmoid gradients in the backward pass).

The cross-modal fusion operation is then formulated as

$${\widehat{\mathbf{F}}}_l^O={\mathbf{F}}_l^O+{\mathbf{G}}_l\odot \mathrm{MHA}({\mathbf{F}}_l^O,{\mathbf{F}}_l^I,{\mathbf{F}}_l^I)$$

(3)

where $\mathrm{MHA}(·)$ denotes multi-head cross-attention [33] with the first argument as query and the remaining as key-value pairs, and ⊙ represents element-wise multiplication with the sparse gate. The symmetric counterpart ${\widehat{\mathbf{F}}}_l^I$ is computed analogously with reversed roles.

4.2.4. Dual-Modal Self-Teaching

To further strengthen the discriminative capacity of each individual modality branch, we introduce a dual-modal self-teaching mechanism. The fused dual-modal prediction, generated from the combined features $[{\widehat{\mathbf{F}}}_l^O;{\widehat{\mathbf{F}}}_l^I]$, serves as a soft supervision signal for each single-modal branch. This self-teaching loss is formulated as

$${\mathcal{L}}_{st}={\displaystyle \frac{1}{2}}[\mathrm{KL}\left(\mathrm{sg}\left(p_{fused}\right)\parallel p_O\right)+\mathrm{KL}\left(\mathrm{sg}\left(p_{fused}\right)\parallel p_I\right)]$$

(4)

where $p_{fused}$, $p_O$, and $p_I$ denote the softmax probability distributions from the fused, optical-only, and InSAR-only predictions, respectively, $\mathrm{sg}(·)$ denotes the stop-gradient operation, and $\mathrm{KL}(·\parallel ·)$ is the Kullback–Leibler divergence. The stop-gradient breaks the potential circular dependency between the fused teacher signal and the single-modal branches from which it is derived: gradients of ${\mathcal{L}}_{st}$ flow only into the single-modal predictions, while the fused branch is trained solely by the segmentation supervision, i.e., $p_{fused}$ is detached before being used as the self-teaching target. This mechanism ensures that each single-modal branch progressively learns to approximate the richer fused representation, thereby improving robustness when one modality is degraded or unavailable during inference.

4.3. Deformation Integrity Prior Decoder (DIPD)

4.3.1. Motivation for the Deformation Integrity Prior

The design of DIPD is inspired by the depth integrity-prior and fine-grained patch strategy of PDFNet [40], which exploits the spatial coherence of depth maps for dichotomous image segmentation. DIPD transfers this design philosophy from scene depth to InSAR surface deformation, which entails three substantive differences. First, the prior is grounded in landslide kinematics rather than scene geometry: deformation coherence reflects the mechanical coupling of a sliding mass, and deformation-gradient discontinuities correspond to physical shear boundaries. Second, the loss formulation differs: DIPD combines a deformation-affinity-weighted intra-region consistency term (Equation (7)) with a boundary-gradient cosine alignment term (Equation (8)), whereas PDFNet enforces its depth integrity-prior through a structural consistency objective on depth-derived cues together with a patch-level refinement loss. Third, DIPD operates inside a symmetric cross-modal fusion framework in which the prior-carrying modality (InSAR) is itself a learned input branch rather than an auxiliary map, and the patch-selection criterion (Equation (6)) jointly uses prediction entropy and deformation-gradient magnitude. The head-to-head comparison with PDFNet adapted to our task (reported in Section 5.3) isolates the benefit of these deformation-specific designs.

The concept of “deformation integrity prior” is rooted in the geophysical observation that high-risk landslide bodies exhibit spatially continuous and coherent deformation patterns in InSAR time-series measurements [4,6]. Unlike vegetation changes or shadow artifacts that may create discontinuous optical signatures, the mechanical coupling within a landslide mass ensures that adjacent pixels within the unstable region share similar displacement velocities and deformation directions. This physical property provides a powerful constraint for segmentation: predicted landslide masks should respect the spatial continuity of the underlying deformation field. Conventional semantic segmentation decoders lack mechanisms to enforce such domain-specific priors, often producing fragmented predictions that violate the physical integrity of deformation zones. The DIPD addresses this limitation by explicitly encoding the deformation integrity constraint into both the decoding architecture and the loss function design. The detailed structure of the proposed DIPD is illustrated in Figure 3.

4.3.2. Coarse-to-Fine Prediction with Fine-Grained Patch Strategy

The DIPD adopts a coarse-to-fine prediction paradigm that progressively refines landslide boundaries while maintaining global deformation consistency. In the coarse prediction stage, the decoder generates an initial probability map ${\mathbf{P}}_{coarse}\in {[0,1]}^{H\times W}$ from the deepest encoder features:

$${\mathbf{P}}_{coarse}=\sigma \left({\mathrm{Conv}}_{1\times 1}\left(\mathrm{Up}\left({\mathbf{F}}_4\right)\right)\right)$$

(5)

where ${\mathbf{F}}_4$ denotes the stage-4 encoder output and $\mathrm{Up}(·)$ represents bilinear upsampling to the full spatial resolution.

Subsequently, a fine-grained patch resampling strategy is applied to refine predictions in uncertain boundary regions. Candidate patches are selected based on a composite criterion that jointly considers prediction uncertainty and deformation gradient magnitude:

$${\mathcal{S}}_{patch}=\{(i,j)\mid H({\mathbf{P}}_{coarse}^{(i,j)})>{\theta }_H\mathrm{and}\parallel \nabla {\mathbf{D}}^{(i,j)}\parallel > {\theta }_D\}$$

(6)

where $H(·)$ denotes the binary entropy function measuring prediction uncertainty, $\nabla \mathbf{D}$ is the spatial gradient of the InSAR LOS deformation field, and ${\theta }_H$, ${\theta }_D$ are adaptive thresholds: ${\theta }_H$ is the 80th percentile of the per-image entropy map and ${\theta }_D$ is the 70th percentile of the per-image deformation-gradient magnitude; i.e., both are computed adaptively per image rather than fixed globally, which makes the selection robust to scene-dependent deformation dynamic ranges. Selected patches are processed at higher resolution with full multi-scale feature aggregation, enabling precise boundary delineation at deformation transition zones.

4.3.3. Deformation Integrity Loss

The deformation integrity loss enforces two complementary physical constraints on the predicted landslide mask. The first component, the intra-region consistency term, penalizes prediction discontinuities within regions of coherent deformation:

$${\mathcal{L}}_{deform}={\displaystyle \frac{1}{\left|\mathrm{\Omega }\right|}}\sum _{(i,j)\in \mathrm{\Omega }}\exp \left(-{\displaystyle \frac{\parallel {\mathbf{D}}^{(i,j)}-{\mathbf{D}}^{(m,n)}{\parallel }^2}{2{\sigma }_D^2}}\right)·|p^{(i,j)}-p^{(m,n)}|$$

(7)

where $\mathrm{\Omega }$ denotes the set of spatially adjacent pixel pairs $\left((i,j),(m,n)\right)$, $(m,n)$ being a spatial neighbor of $(i,j)$ under 8-connectivity (all first-order neighbors), ${\mathbf{D}}^{(i,j)}$ is the LOS deformation value at location $(i,j)$, and the exponential weighting ensures that prediction consistency is enforced most strongly between pixels sharing similar deformation characteristics. The bandwidth ${\sigma }_D$ is estimated per batch as the standard deviation of neighboring-pixel LOS deformation differences. Formally, Equation (7) is a deformation-affinity-weighted pairwise smoothness term, analogous to the intensity-weighted pairwise potentials of conditional random fields but with the affinity defined in the deformation domain and optimized end-to-end; it is a physically motivated regularizer encoding the quasi-coherent motion of a mechanically coupled sliding mass, rather than a constraint derived from first principles.

The second component, the boundary gradient alignment term, encourages predicted mask boundaries to coincide with sharp transitions in the deformation field:

$${\mathcal{L}}_{boundary}=1-{\displaystyle \frac{\langle \nabla {\mathbf{P}}_{pred},\nabla \mathbf{D}\rangle }{\parallel \nabla {\mathbf{P}}_{pred}\parallel ·\parallel \nabla \mathbf{D}\parallel +ϵ}}$$

(8)

where $\nabla {\mathbf{P}}_{pred}$ is the spatial gradient of the predicted probability map and $ϵ$ is a small constant for numerical stability. This term maximizes the cosine similarity between prediction boundaries and deformation boundaries. We note that Equation (8) by itself cannot identify the origin of a deformation discontinuity; the discrimination between landslide boundaries and other deformation edges (e.g., mining subsidence rims) is performed jointly with the learned cross-modal features and the segmentation supervision, which determine where landslide predictions appear in the first place. Because the term is only active where $\nabla {\mathbf{P}}_{pred}$ is non-negligible, it refines the boundaries of regions the network already considers landslide candidates instead of attracting predictions toward every deformation edge in the scene (see also Section 6.2).

The total training loss integrates segmentation supervision with the physical constraints:

$${\mathcal{L}}_{total}={\mathcal{L}}_{seg}+{\lambda }_1{\mathcal{L}}_{boundary}+{\lambda }_2{\mathcal{L}}_{deform}+{\lambda }_3{\mathcal{L}}_{ssim}+{\lambda }_4{\mathcal{L}}_{st}+{\lambda }_5{\mathcal{L}}_{sparse}$$

(9)

where ${\mathcal{L}}_{seg}$ combines binary cross-entropy and Dice loss [23], ${\mathcal{L}}_{ssim}$ is the structural similarity loss that preserves local structural patterns, computed as the standard single-scale SSIM with an $11\times 11$ Gaussian window ($\sigma =1.5$) between the single-channel predicted probability map and the ground-truth mask at full output resolution, and ${\mathcal{L}}_{st}$ and ${\mathcal{L}}_{sparse}$ are the dual-modal self-teaching loss and the cross-modal sparsity regularization defined earlier in this section. The coefficients ${\lambda }_1$, ${\lambda }_2$, ${\lambda }_3$ are empirically set to 0.5, 1.0, and 0.2, respectively, based on the sensitivity analysis reported in Section 5.6, while ${\lambda }_4$ and ${\lambda }_5$, which weight the auxiliary self-teaching and sparsity-regularization terms, are fixed to 0.5 and 0.01, respectively. To make the balance among the terms transparent,

Table 2. Typical magnitude of each weighted loss component at three training stages on Nanning-HHLS (running-average weighted terms).

Weighted Term	Epoch 1	Epoch 60	Epoch 120
${\mathcal{L}}_{seg}$	0.812	0.214	0.128
${\lambda }_1{\mathcal{L}}_{boundary}$	0.341	0.112	0.068
${\lambda }_2{\mathcal{L}}_{deform}$	0.287	0.093	0.051
${\lambda }_3{\mathcal{L}}_{ssim}$	0.124	0.047	0.029
${\lambda }_4{\mathcal{L}}_{st}$	0.196	0.061	0.033
${\lambda }_5{\mathcal{L}}_{sparse}$	0.021	0.009	0.006

reports the typical magnitude of each weighted component at the beginning, middle, and end of training, confirming that the physics-based terms act as regularizers of comparable but sub-dominant magnitude relative to the segmentation loss.

4.3.4. Integration with Multi-Scale Features

The DIPD integrates hierarchical features $\{{\mathbf{F}}_1,{\mathbf{F}}_2,{\mathbf{F}}_3,{\mathbf{F}}_4\}$ from the encoder through skip connections equipped with attention-weighted fusion. At each decoder level, the skip-connected encoder feature is re-weighted by a channel attention module that evaluates its relevance to the current decoding context. This selective integration mechanism prevents irrelevant low-level details from contaminating the physically constrained predictions while preserving fine-grained spatial information necessary for accurate boundary recovery. The attention weights are jointly learned with the decoder parameters, enabling adaptive feature selection across different landslide scales and deformation patterns.

4.4. Lightweight Deployable Student Network (LDSN)

4.4.1. Teacher–Student Framework

While the full DIGSFNet model (SDAE + DIPD) achieves superior extraction accuracy through its dual-modal symmetric encoder and physically constrained decoder, its computational complexity presents challenges for operational deployment scenarios requiring real-time processing or execution on resource-constrained edge platforms. To bridge this gap, we design a Lightweight Deployable Student Network that inherits the representational knowledge of the full model through systematic knowledge distillation. The overall teacher–student distillation pipeline is illustrated in Figure 4.

The teacher network comprises the complete DIGSFNet with a Swin-B backbone, dual-branch SDAE with cross-modal fusion modules, and the full DIPD with coarse-to-fine refinement. The student network adopts a significantly more compact architecture that, by default, uses LSNet-T [44]—the tiny variant of the LSNet lightweight backbone family—as the backbone encoder (MobileNetV4 [50] can serve as an alternative), paired with a lightweight single-stream decoder with depth-wise separable convolutions. The student receives both modalities as a channel-concatenated input and relies primarily on distilled knowledge rather than heavy dual-branch fusion modules to achieve effective multi-modal understanding. To enable explicit alignment of deformation-aware reasoning between the teacher and the student, a single lightweight deformation-aware attention head is retained at the deepest decoder stage of the student. This head consists of a depth-wise $3\times 3$ convolution followed by a $1\times 1$ convolution and a sigmoid activation operating on the deepest decoder features, producing a spatial response map that serves as the alignment target for the deformation-prior distillation described below. The total parameter count of this attention head accounts for less than 0.4% of the student model and does not affect the overall lightweight property. Specifically, the head maps the $C=256$-channel deepest decoder feature to a single-channel response map (≈2.6k parameters in total); it is used solely as the alignment target of Equation (12), does not gate the segmentation path, and is retained at inference (though it can be removed without affecting the segmentation output).

4.4.2. Cross-Modal Knowledge Distillation

The knowledge distillation framework operates at three complementary levels to transfer the teacher’s multi-modal reasoning capability to the compact student. The mask-level distillation aligns the student’s output probability distribution with the teacher’s soft predictions:

$${\mathcal{L}}_{mask-KD}=\mathrm{KL}\left(\mathrm{softmax}(z_T/T)\parallel \mathrm{softmax}(z_S/T)\right)$$

(10)

where $z_T$ and $z_S$ are the teacher and student logits, T is the temperature parameter, and $\mathrm{softmax}(·)$ denotes the softmax function.

The feature-level distillation transfers intermediate representational knowledge through projection-aligned feature matching:

$${\mathcal{L}}_{feat-KD}={\displaystyle \sum _{l=1}^L}{\parallel {\varphi }_l\left({\mathbf{F}}_l^S\right)-\mathrm{sg}\left({\mathbf{F}}_l^T\right)\parallel }_2^2$$

(11)

where ${\varphi }_l(·)$ is a learnable projection head that aligns the student feature dimensions with the teacher, and $\mathrm{sg}(·)$ denotes the stop-gradient operation.

The deformation-prior distillation specifically transfers the teacher’s learned understanding of deformation integrity constraints:

$${\mathcal{L}}_{prior-KD}={\parallel {\mathbf{A}}_{deform}^S-\mathrm{sg}({\mathbf{A}}_{deform}^T)\parallel }_F^2$$

(12)

where ${\mathbf{A}}_{deform}^T$ denotes the deformation-aware attention map from the teacher’s DIPD (i.e., the attention-weighted skip-connection response at the deepest decoder stage), ${\mathbf{A}}_{deform}^S$ is the response map produced by the lightweight deformation-aware attention head retained in the student decoder as described in Section 4.4.1, and ${\parallel ·\parallel }_F$ denotes the Frobenius norm. Bilinear interpolation is applied to align the spatial resolution between the teacher and student attention maps before computing the distance.

The total student training loss combines task supervision with all distillation objectives:

$${\mathcal{L}}_{student}={\mathcal{L}}_{seg}+{\alpha }_1{\mathcal{L}}_{mask-KD}+{\alpha }_2{\mathcal{L}}_{feat-KD}+{\alpha }_3{\mathcal{L}}_{prior-KD}$$

(13)

where ${\alpha }_1=1.0$, ${\alpha }_2=0.5$, and ${\alpha }_3=0.8$ are determined through grid search on the validation set, over the ranges ${\alpha }_1\in \{0.5,1.0,2.0\}$, ${\alpha }_2\in \{0.1,0.5,1.0\}$, and ${\alpha }_3\in \{0.2,0.5,0.8,1.0\}$.

4.4.3. Deployment Optimization

To maximize inference efficiency for operational deployment, the trained student network undergoes several post-training optimizations. First, structural reparameterization [51] converts multi-branch training-time architectures (e.g., residual connections in adapter blocks) into equivalent single-path inference architectures, eliminating memory access overhead. Subsequently, the model is exported to the ONNX intermediate representation format for cross-platform compatibility. For GPU-based deployment, TensorRT optimization is applied with layer fusion, kernel auto-tuning, and precision calibration. Both FP16 mixed-precision and INT8 post-training quantization are supported, with INT8 calibration performed on a representative subset of 500 samples from the training set to minimize quantization error, using TensorRT entropy (KL-divergence) calibration over the 500 samples in 63 batches of 8, with per-channel quantization for weights and per-tensor quantization for activations, and the first and last layers kept in FP16. These optimizations collectively reduce inference latency by approximately 11.8× for the INT8 student compared to the full teacher model, while the FP32 student preserves about 95.6% of the teacher’s mIoU and the INT8-quantized student still retains over 94.8% (see Section 5.5 for full numerical details), enabling real-time landslide monitoring at regional scales.

5. Experiments and Results

5.1. Experimental Settings

The proposed DIGSFNet was implemented using the PyTorch 2.1 deep learning framework and all experiments were conducted on a workstation equipped with four NVIDIA RTX 4090 GPUs (24 GB VRAM each; NVIDIA Corporation, Santa Clara, CA, USA). The AdamW optimizer [52] was employed for model parameter optimization with an initial learning rate of $6\times {10}^{-5}$ and a weight decay coefficient of $5\times {10}^{-4}$. The learning rate followed a cosine annealing schedule with linear warm-up over the first five epochs to stabilize the early training dynamics. The teacher model (Swin-B backbone with SDAE and DIPD) was trained for 120 epochs with a batch size of 8, while the student model (LSNet-T backbone with the simplified decoder) was trained for 80 epochs under the knowledge distillation framework with a batch size of 16. Input images were cropped into non-overlapping patches of $512\times 512$ pixels during training. Extensive online data augmentation strategies were adopted to mitigate overfitting, including random horizontal flipping, random vertical flipping, random rotation (0^∘, 90^∘, 180^∘, 270^∘), and random color jitter (brightness $\pm 0.2$, contrast $\pm 0.2$, saturation $\pm 0.2$, hue $\pm 0.05$) for the optical modality. The corresponding spatial transforms (flipping and rotation) were applied identically to the InSAR modality to maintain strict spatial alignment, while color jitter was not applied to InSAR features since deformation values carry physical meaning. Regarding the interaction between rotation and radar geometry: the stored InSAR modality is a scalar per-pixel LOS displacement (or velocity) map rather than a vector field, so spatial rotation permutes pixel locations without altering the scalar values, and no re-projection of the LOS vector is required for the stored quantity to remain well-defined. What rotation does alter is the statistical relationship between the orientation of the deformation pattern and the (implicit) radar viewing geometry and slope aspect. We treat this as a deliberate trade-off: restricting rotations to multiples of 90^∘ preserves the raster structure, and the augmentation encourages the network to rely on rotation-covariant deformation coherence and gradient cues rather than absolute orientation, which is desirable for cross-track generalization. A control experiment without rotation augmentation yields 83.39% mIoU (a 0.18-point drop), indicating the augmentation is mildly beneficial. The Swin-B backbone was initialized with ImageNet-22K pre-trained weights [49], while the modality-specific adapters, dynamic sparse fusion modules, and decoder components were initialized with Kaiming uniform initialization [53]. Because the residual formulation of Equation (1) keeps each adapter close to an identity mapping when its output branch is small, the randomly initialized adapters perturb the pretrained backbone representations only mildly at the start of training; combined with the five-epoch linear learning-rate warm-up, we observed no divergence or feature collapse in the early phase. Mixed-precision training (FP16) was enabled via PyTorch’s automatic mixed-precision module to accelerate training throughput.

5.2. Evaluation Metrics

To comprehensively evaluate segmentation accuracy and practical utility, we adopt the following evaluation metrics: mean Intersection over Union (mIoU), F1-score (harmonic mean of Precision and Recall), Precision, Recall, Area Under the ROC Curve (AUC), and Boundary F1-score (BF1, computed at a two-pixel tolerance threshold) for assessing segmentation quality. Additionally, Missing Rate (defined as $1-\mathrm{Recall}$) and False Alarm Rate (defined as $1-\mathrm{Precision}$) are reported to quantify the practical reliability of landslide detection. For computational efficiency, we report model parameters (Params, M), floating-point operations (FLOPs, G), inference speed (FPS), ONNX inference latency (ms), and INT8 inference latency (ms), all measured on a single RTX 4090 GPU with $512\times 512$ input resolution. FPS and latency denote pure model inference (batch size 1), excluding data loading and preprocessing, measured with CUDA event timing after 200 warm-up iterations and averaged over 1000 timed iterations. The two-pixel tolerance of BF1 (20 m at the 10 m ground sampling distance) is chosen to match the boundary uncertainty of the reference annotations themselves—the mean inter-annotator boundary deviation on Nanning-HHLS is 1.7 pixels—so that stricter tolerances would penalize deviations smaller than the label noise; BF1 at one- and three-pixel tolerances is additionally reported in Supplementary Table S2, with unchanged method ranking. For high-risk landslide extraction, Recall and Missing Rate are particularly critical: in practical disaster prevention, it is preferable to tolerate a small number of false alarms rather than miss high-risk landslides that may pose direct threats to human lives and infrastructure.

5.3. Comparison with State-of-the-Art Methods

To validate the effectiveness of the proposed DIGSFNet, we conduct comprehensive comparisons with nine state-of-the-art semantic segmentation, multi-modal fusion, and dedicated active-landslide detection methods. The comparison methods include four single-modal architectures—U-Net [10], DeepLabV3+ [11], SegFormer [12], and Swin-UNet [54]—that were originally designed for single-modality input and were adapted to accept concatenated optical–InSAR input along the channel dimension. Additionally, four advanced fusion-capable architectures are compared: UNetFormer [55], SegMAN [56] (adapted for dual-modal input by replacing its RGB-only encoder with symmetric dual-modal branches), KTB [57] (adapted for InSAR–optical fusion following its balanced multi-modal fusion paradigm), and PDFNet [40] (adapted for the landslide extraction task by modifying its input interface). For each adapted baseline, Supplementary Table S1 specifies the exact modification performed: which encoder consumes which modality, how the input stems were reshaped, and which auxiliary branches were removed; the parameter counts reported in

Table 3. Quantitative comparison on the Nanning-HHLS dataset. The best results are marked in bold and the second-best results underlined. Bold and underline apply to the accuracy metrics only; the last column (model parameters) is an efficiency measure for which a smaller value is preferable and is therefore left unmarked.

Method	mIoU (%)	F1 (%)	Precision (%)	Recall (%)	AUC (%)	BF1 (%)	Miss. (%)	FA (%)	Params (M)
U-Net	71.24	82.45	83.12	81.79	91.56	68.32	18.21	16.88	31.04
DeepLabV3+	72.86	83.67	84.22	83.13	92.31	69.78	16.87	15.78	41.26
SegFormer	74.53	85.01	85.67	84.36	93.12	71.45	15.64	14.33	47.28
Swin-UNet	75.12	85.42	86.04	84.81	93.47	72.11	15.19	13.96	59.84
UNetFormer	76.38	86.32	86.91	85.74	93.92	73.28	14.26	13.10	11.69
SegMAN	77.92	87.42	87.83	87.01	94.56	74.86	12.99	12.18	52.37
KTB	79.14	88.29	88.52	88.06	95.03	76.04	11.94	11.48	64.22
PDFNet	80.26	89.06	89.31	88.81	95.47	77.12	11.19	10.70	71.85
HAEFNet	80.94	89.52	89.62	89.43	95.78	77.86	10.57	10.38	76.32
DIGSFNet (Ours)	83.57	91.16	90.84	91.48	96.82	82.34	8.52	9.16	88.46

Miss. = Missing Rate; FA = False Alarm Rate.

and

Table 4. Quantitative comparison on the public HAEFNet benchmark [5]. Same format and methods as Table 3; the DEM channel is disabled for all methods. The best results are marked in bold and the second-best results underlined. Bold and underline apply to the accuracy metrics only; the last column (model parameters, an efficiency measure) is not marked.

Method	mIoU (%)	F1 (%)	Precision (%)	Recall (%)	AUC (%)	BF1 (%)	Miss. (%)	FA (%)	Params (M)
U-Net	67.12	79.84	80.42	79.27	89.84	64.18	20.73	19.58	31.04
DeepLabV3+	68.83	81.05	81.62	80.49	90.51	65.74	19.51	18.38	41.26
SegFormer	70.62	82.38	82.89	81.88	91.27	67.43	18.12	17.11	47.28
Swin-UNet	71.45	82.92	83.51	82.34	91.68	68.21	17.66	16.49	59.84
UNetFormer	72.34	83.59	84.08	83.11	92.06	69.27	16.89	15.92	11.69
SegMAN	73.58	84.42	84.81	84.04	92.71	70.62	15.96	15.19	52.37
KTB	74.62	85.13	85.42	84.85	93.18	71.85	15.15	14.58	64.22
PDFNet	75.18	85.51	85.74	85.28	93.46	72.43	14.72	14.26	71.85
HAEFNet	75.24	85.84	85.92	85.76	93.62	73.18	14.24	14.08	76.32
DIGSFNet (Ours)	78.92	88.14	87.95	88.34	95.04	78.45	11.66	12.05	88.46

Miss. = Missing Rate; FA = False Alarm Rate. The HAEFNet results listed here are re-produced under the InSAR + optical setting with the DEM branch removed. For reference, the original HAEFNet paper reports 75.2% IoU/85.8% F1 (single landslide class) using its full InSAR + optical + DEM input; although that single-class IoU is not strictly comparable to the mIoU reported here, our DEM-free re-implementation attains a similar accuracy level, suggesting that the DEM branch contributes only a marginal gain on this benchmark—consistent with our decision to exclude DEM (Section 6.2). For a like-for-like check under the original paper’s metric, our DEM-free re-trained HAEFNet attains 74.6% single-class landslide IoU on this benchmark, i.e., 0.6 points below the original DEM-inclusive result.

were re-measured on the adapted variants actually trained, using the same profiler as for our model. To establish a strong landslide-specific baseline on the public benchmark, we additionally include HAEFNet [5], the reliability-aware evidence-theoretic InSAR–optical–DEM fusion network that originally accompanies the HAEFNet dataset. For a fair head-to-head comparison on Nanning-HHLS, the DEM-dependent branch of HAEFNet is disabled and the network is re-trained with InSAR + optical inputs only. All comparison methods were retrained on the same training data using their officially released code with recommended hyperparameters, and the best-performing checkpoints on the validation set were selected for fair comparison.

5.3.1. Results on the Nanning-HHLS Dataset

The quantitative evaluation results on the Nanning-HHLS dataset are presented in Table 3. Consistent with the operational priority argued in Section 5.2, we discuss Recall first: DIGSFNet achieves the highest Recall of 91.48% with a Missing Rate of only 8.52%—2.05 percentage points higher Recall than the strongest baseline HAEFNet—which is the metric of primary consequence for disaster prevention, where a missed high-risk landslide can be catastrophic while a false alarm merely costs verification effort (at the instance level this corresponds to six fewer missed high-risk landslides on the test set). Overall, the proposed DIGSFNet achieves the best performance across all accuracy-related metrics, attaining an mIoU of 83.57%, outperforming the second-best method HAEFNet (re-trained with InSAR + optical inputs) by 2.63% in terms of mIoU and 1.64% in terms of F1-score. The superior Boundary F1-score of 82.34% further demonstrates the effectiveness of the deformation integrity prior in the DIPD module for producing boundary-accurate landslide masks that faithfully delineate the extent of the sliding body. All headline results are reported as the mean over three independent runs with different random seeds; standard deviations and paired significance tests against HAEFNet are given in Section 5.7.

Among the single-modal architectures adapted for dual-modal input, U-Net and DeepLabV3+ achieve moderate performance, yet they fail to detect landslides on vegetated slopes where spectral changes are subtle but deformation signals are prominent. SegFormer and Swin-UNet improve upon the CNN-based methods by capturing longer-range spatial dependencies through self-attention, yet they still suffer from fragmented predictions and elevated false alarm rates due to the lack of deformation-aware constraints. The multi-modal fusion methods—SegMAN, KTB, PDFNet, and HAEFNet—demonstrate improved performance over single-modal baselines by incorporating InSAR information, confirming the value of deformation data for landslide identification. Among them, HAEFNet, which is specifically tailored for active-landslide detection through evidence-theoretic InSAR–optical fusion, ranks second on the Nanning-HHLS dataset, validating its strong design but also revealing that its reliability-aware fusion paradigm is still vulnerable to the karstic, densely-vegetated subtropical terrain that characterizes Guangxi. The asymmetric or insufficiently exploitative fusion strategies of the remaining baselines lead to sub-optimal deformation feature utilization, resulting in lower Recall and higher Missing Rate compared with the proposed symmetric fusion approach. DIGSFNet outperforms HAEFNet by 2.05% in Recall and 4.48% in BF1, indicating that the symmetric modality treatment and deformation integrity constraints provide substantial advantages for identifying complete and accurately bounded landslide regions.

5.3.2. Results on the Public Dataset

To further validate that the architectural design of the proposed method is not specific to a single geomorphological setting, we conduct additional experiments on the publicly available HAEFNet benchmark [5], which covers active-landslide-prone regions in the Qinghai–Tibet Plateau and Sichuan—geological environments that differ substantially from the karstic, subtropical Nanning study area in terms of lithology, vegetation cover, slope failure mechanisms, and climatic forcing. The HAEFNet dataset provides spatially co-registered Sentinel-1 InSAR mean LOS rate, Sentinel-2 RGB imagery, and Copernicus GLO-30 DEM. To keep the experimental protocol consistent with the Nanning-HHLS experiments and with the rationale in Section 6.2, the DEM channel is ignored, and DIGSFNet is independently trained and evaluated on this benchmark under its full dual-modal (InSAR + optical) configuration; the only architectural change is that the first input convolution of the optical stem is reshaped from 10 to 3 channels to match the RGB input of the HAEFNet dataset, while all other modules remain unchanged. All baselines that originally consume DEM (i.e., HAEFNet) are likewise re-trained without the DEM branch, while all other baselines use InSAR + optical inputs identical to the Nanning-HHLS setup, ensuring a strictly head-to-head comparison. We emphasize that this protocol evaluates how well the architecture generalizes when retrained on each region in turn, rather than a single-model zero-shot transfer between regions; a true cross-region transfer experiment is left as future work.

The corresponding results are presented in Table 4. DIGSFNet maintains its superior performance on this geomorphologically distinct benchmark, achieving 78.92% mIoU and 88.14% F1-score, which surpass the second-best method HAEFNet by 3.68% mIoU and 2.30% F1-score. The consistent performance advantage observed when the same architecture is independently trained on two datasets with substantially different geological and climatic conditions indicates that the architectural inductive biases of DIGSFNet—symmetric modality treatment and deformation-integrity priors—are not tied to any single region. In particular, the gap between DIGSFNet and HAEFNet on the HAEFNet benchmark is even larger than on Nanning-HHLS, suggesting that the symmetric deformation-aware encoder and the deformation integrity prior remain effective when applied to high-altitude, sparsely vegetated terrain where the InSAR signal-to-noise characteristics, optical reflectance statistics, and landslide morphology all differ from the Nanning study area. Importantly, the absolute mIoU lead of DIGSFNet over the runner-up method actually widens on the more challenging benchmark—from 2.63% on Nanning-HHLS (against HAEFNet) to 3.68% on the HAEFNet benchmark (also against HAEFNet)—and the lead in BF1 likewise grows from 4.48% to 5.27%. This indicates that the deformation-integrity prior and the symmetric encoder both contribute to robustness when the network is retrained on regions of varying difficulty, while all methods reach a lower absolute mIoU on the HAEFNet benchmark, which is consistent with its larger spectral and geomorphological diversity.

5.3.3. Qualitative Comparison

The qualitative comparison results are visualized in Figure 5, where we present the segmentation outputs of all compared methods on several representative challenging scenes drawn from both datasets. On densely vegetated subtropical slopes of Nanning-HHLS where the optical appearance is highly ambiguous, single-modal methods (U-Net, DeepLabV3+) produce significant false-negative errors, missing large portions of the high-risk landslide area. In contrast, DIGSFNet accurately delineates these regions by effectively leveraging InSAR deformation information through the symmetric fusion mechanism. On the high-altitude, sparsely-vegetated slopes of the HAEFNet Qinghai–Tibet scenes, the dominant failure mode shifts to false alarms triggered by glacial moraine textures, periglacial scree, and non-landslide creep; DIGSFNet correctly suppresses these false detections owing to the deformation-integrity constraint that decouples landslide-style continuous deformation from broader-scale periglacial signals. Small isolated target landslides that occupy only a few dozen pixels are completely missed by most comparison methods on both datasets, whereas DIGSFNet successfully detects them. Existing methods also produce pronounced false alarms in regions with non-landslide deformation sources such as construction activities and mining subsidence, while DIGSFNet correctly suppresses these false detections. Furthermore, existing fusion methods (SegMAN, KTB) generate fragmented masks with broken boundaries and even HAEFNet, despite its evidence-theoretic reliability weighting, occasionally produces over-confident yet boundary-inaccurate masks at the periphery of the sliding body; in contrast, DIGSFNet produces spatially complete and boundary-accurate predictions on both datasets, demonstrating the effectiveness of the boundary gradient alignment constraint in the DIPD.

5.4. Ablation Studies

To systematically validate the effectiveness and necessity of each proposed component, we conduct comprehensive ablation studies on the Nanning-HHLS dataset. Both “progressive addition” and “component removal” strategies are employed to provide a thorough analysis. The eleven ablation configurations and their corresponding results are presented in

Table 5. Ablation study results on the Nanning-HHLS dataset. Within the teacher variants (Configs 1–9) and the student variants (Configs 10–11), the best value in each accuracy column is shown in bold; the parameters and FLOPs columns (efficiency measures) are not marked.

Configuration	mIoU (%)	F1 (%)	Recall (%)	BF1 (%)	Params (M)	FLOPs (G)
(1) Optical only	74.12	85.18	82.15	66.48	85.34	197.6
(2) InSAR only	71.85	83.56	79.42	62.78	85.34	197.6
(3) Optical + InSAR direct concatenation	76.42	86.64	83.68	68.92	85.42	198.9
(4) Asymmetric fusion (InSAR as auxiliary)	78.56	87.92	85.41	71.34	86.78	208.3
(5) Symmetric fusion (SDAE, w/o DIPD)	81.03	89.53	88.56	76.82	87.62	211.2
(6) Full model w/o deformation integrity loss	81.86	89.98	89.87	78.12	88.46	215.8
(7) Full model w/o fine-grained patch decoder	82.64	90.52	90.68	79.45	87.85	211.2
(8) Full model w/o self-teaching	82.15	90.26	90.23	81.08	88.46	215.8
(9) Full DIGSFNet (teacher)	83.57	91.16	91.48	82.34	88.46	215.8
(10) LDSN (student, w/o distillation)	76.48	86.68	84.12	72.58	11.32	27.6
(11) LDSN (student, with full distillation)	79.92	88.79	87.96	77.84	11.32	27.6

, with representative qualitative improvements across configurations visualized in Figure 6.

Single-modal baselines (Configs 1–2). The optical-only baseline (Config 1) achieves 74.12% mIoU, providing a strong baseline due to the rich spectral-textural information available in the optical modality. The InSAR-only baseline (Config 2) achieves 71.85% mIoU, which is lower than the optical baseline but still demonstrates meaningful landslide identification capability, confirming the intrinsic value of deformation information for this task. The performance gap between the two single-modal baselines motivates the design of a fusion framework that can fully exploit the complementary strengths of both modalities.

Fusion strategy comparison (Configs 3–5). Direct concatenation of optical and InSAR inputs along the channel dimension (Config 3) improves upon the optical-only baseline by 2.30% in mIoU, demonstrating the basic benefit of incorporating InSAR data. Asymmetric fusion (Config 4), where InSAR features are fed as auxiliary inputs to the optical backbone through a cross-attention mechanism, achieves a further improvement of 2.14% over direct concatenation. The proposed symmetric fusion with SDAE (Config 5) yields the most substantial improvement, surpassing direct concatenation by 4.61% in mIoU and surpassing asymmetric fusion by 2.47%. Notably, the improvement in Recall is particularly pronounced (3.15% over asymmetric fusion), confirming that symmetric modality treatment enables more effective exploitation of deformation signals for detecting high-risk landslides that might otherwise be missed when InSAR features are suppressed by the dominant optical branch.

DIPD decoder component analysis (Configs 5–7, 9). Comparing Config 5 (SDAE only, without DIPD) with Config 9 (full DIGSFNet), the addition of the DIPD module brings an improvement of 2.54% in mIoU and a particularly significant 5.52% improvement in BF1, validating the effectiveness of deformation integrity priors for boundary refinement. Removing the deformation integrity loss (i.e., setting ${\lambda }_1={\lambda }_2=0$ while keeping ${\mathcal{L}}_{seg}$ and ${\mathcal{L}}_{ssim}$, Config 6) causes a 1.71% drop in mIoU and a 4.22% drop in BF1 compared with the full model, demonstrating its critical role in ensuring spatial completeness and boundary accuracy. Removing the fine-grained patch decoder (Config 7) results in a 2.89% decrease in BF1, confirming its targeted effectiveness for local boundary refinement in regions with high prediction uncertainty and strong deformation gradients. We further note that the cumulative impact of removing individual DIPD sub-components (1.71% + 0.93% in mIoU) is larger than the single-step DIPD-overall improvement (2.54%), indicating positive interaction effects among modules.

SDAE auxiliary mechanism analysis (Config 8). Removing the dual-modal self-teaching mechanism (which is part of the SDAE encoder rather than the DIPD decoder) leads to a 1.42% decrease in mIoU compared with the full model. This degradation is qualitatively most apparent in scenes containing degraded InSAR coherence or partial cloud cover in the optical modality, validating its contribution to cross-modal robustness when one modality is partially corrupted.

Knowledge distillation evaluation (Configs 9–11). The LDSN student model trained without knowledge distillation (Config 10) achieves 76.48% mIoU with only 12.8% of the teacher’s parameters and 12.8% of the FLOPs, representing a baseline for the lightweight architecture. With full cross-modal knowledge distillation (Config 11), the student model’s mIoU improves to 79.92%, recovering approximately 95.63% of the teacher’s performance.

Contribution of individual distillation levels. To attribute the distillation gain to its three components,

Table 6. Ablation of the three distillation levels on the Nanning-HHLS dataset (student = LSNet-T). The best value in each column is shown in bold.

Distillation Configuration	mIoU (%)	F1 (%)	Recall (%)	BF1 (%)
None (Config 10)	76.48	86.68	84.12	72.58
Mask-level only	77.81	87.41	85.63	74.06
Feature-level only	77.95	87.53	85.81	74.39
Deformation-prior only	78.16	87.66	85.44	75.90
Mask + feature	78.84	88.21	86.99	75.60
Full three-level (Config 11)	79.92	88.79	87.96	77.84

ablates the mask-level, feature-level, and deformation-prior terms individually and in combination. Each level contributes positively; deformation-prior distillation yields the largest single-term BF1 gain ($+3.32$ points over the no-distillation baseline), and the three levels are complementary, the full combination recovering 95.63% of the teacher mIoU.

Architectural merit versus model capacity. To distinguish architectural effects from capacity effects, we consider three observations. First, the capacity gap to the strongest baselines is moderate (HAEFNet: 76.32 M; PDFNet: 71.85 M), yet the accuracy gap widens rather than narrows on the benchmark for which HAEFNet was designed (Table 4). Second, Configs 3–9 in Table 5 constitute a capacity-controlled comparison: all share the same Swin-B backbone with nearly identical parameter counts (85.42–88.46 M), yet span 76.42–83.57% mIoU; a 7.15-point spread produced at essentially constant capacity, with the entire increment from Config 3 to Config 9 costing only +3.04 M parameters (+3.6%), is attributable to the architectural mechanisms rather than to size. Third, the distilled student (11.32 M, smaller than every compared baseline) still reaches 79.92% mIoU, within 1.02 points of the strongest 76 M-scale baseline (HAEFNet, 80.94%), which would not be possible if the improvements were primarily a capacity effect.

5.5. Lightweight Deployment Evaluation

To demonstrate the practical deployability of the proposed method for large-area landslide hazard surveys, we conduct a systematic efficiency comparison between the teacher model and the student model under different optimization configurations. The results are presented in

Table 7. Efficiency comparison between teacher and student models. All inference metrics measured on a single RTX 4090 GPU with $512\times 512$ input.

Model	Params (M)	FLOPs (G)	FPS	ONNX (ms)	INT8 (ms)	GPU Mem. (GB)	mIoU (%)	F1 (%)
DIGSFNet Teacher (FP32)	88.46	215.8	18	54.2	—	4.85	83.57	91.16
LDSN Student (FP32)	11.32	27.6	76	13.2	—	1.24	79.92	88.79
LDSN Student (FP16)	11.32	27.6	142	7.1	—	0.82	79.86	88.74
LDSN Student (INT8)	11.32	27.6	218	—	4.6	0.54	79.24	88.31

.

The efficiency comparison results demonstrate that the LDSN student model achieves a highly favorable accuracy–efficiency trade-off suitable for practical deployment. The Student (FP32) model reduces the parameter count by 7.81× and FLOPs by 7.82× compared with the teacher model, with an mIoU degradation of only 3.65%. After FP16 optimization, the inference speed is further improved to 142 FPS with negligible accuracy loss (about 0.06% mIoU drop). The fully optimized Student (INT8) model achieves the most aggressive compression, reducing inference latency by 11.78× (comparing the FP32 teacher’s ONNX latency with the INT8 student’s TensorRT latency) compared with the teacher while maintaining 79.24% mIoU. We note that this 11.78× end-to-end latency reduction is not expected to be proportional to the 7.82× FLOPs reduction: FLOPs is a precision-independent operation count comparing architectures, whereas the latency figure compounds three multiplicative factors—the architectural FLOPs reduction, the higher throughput of INT8 integer arithmetic, and TensorRT graph optimizations (layer fusion and reduced memory traffic); latency on modern GPUs is, moreover, partly memory-bandwidth-bound rather than purely compute-bound. The like-for-like architectural comparison at equal precision is FP32 teacher (54.2 ms, ONNX) versus FP32 student (13.2 ms, ONNX), i.e., 4.1×. At 218 FPS with 4.6 ms per-frame latency and only 0.54 GB GPU memory consumption, the INT8 student model enables rapid inference across large areas, meeting the practical requirements for operational landslide hazard surveys where thousands of square kilometers must be processed within constrained time budgets. These results confirm that the proposed cross-modal knowledge distillation and quantization pipeline successfully preserves the teacher’s deformation-aware segmentation capability in a deployment-ready lightweight model.

5.6. Sensitivity Analysis

We conduct a sensitivity analysis on six key hyperparameters to provide practical guidance for model configuration. The hyperparameters examined include the loss weights (${\lambda }_1$, ${\lambda }_2$, ${\lambda }_3$) controlling the contribution of the boundary gradient alignment loss, intra-region consistency loss, and structural similarity loss, respectively; the adapter bottleneck dimension (r); the sparsity ratio ($\rho$) in the dynamic sparse cross-modal fusion module; and the fine-grained patch size in the DIPD. Each hyperparameter is varied independently while keeping all others fixed at their default values, and both mIoU and F1-score are recorded.

The sensitivity analysis results presented in Figure 7 indicate that DIGSFNet exhibits robust and stable performance across a reasonable range of hyperparameter values for all six parameters. The scans cover ${\lambda }_1\in \{0.01,0.05,0.1,0.2,0.5,1.0,2.0\}$ and ${\lambda }_2\in \{0.01,0.05,0.1,0.2,0.5,1.0,2.0,5.0\}$. As shown in Figure 7a,b, mIoU peaks at ${\lambda }_1=0.5$ (83.57%) and declines to 83.31% and 82.87% at ${\lambda }_1=1.0$ and $2.0$, and peaks at ${\lambda }_2=1.0$ (83.57%), decreasing to 83.29% and 82.61% at ${\lambda }_2=2.0$ and $5.0$; the adopted defaults are therefore interior optima. Within the explored ranges, mIoU varies by less than 0.85% for ${\lambda }_1,{\lambda }_2\in [0.05,0.5]$. For ${\lambda }_3$, the curve is comparatively flat over $[0.1,0.5]$—the mIoU variation across this interval is 0.12 points (83.45 at ${\lambda }_3=0.1$, 83.57 at the peak, 83.50 at ${\lambda }_3=0.5$; Figure 7c)—and reaches its peak at ${\lambda }_3=0.2$, which is therefore adopted as the default; this confirms that the structural similarity constraint provides a stable regularization effect without requiring careful tuning. The adapter bottleneck dimension $r=64$ provides the best accuracy–efficiency trade-off: smaller values ($r\le 32$) lead to insufficient modality-specific adaptation capacity, while larger values ($r\ge 128$) yield marginal accuracy gains at the cost of increased parameters. The sparsity ratio $\rho =0.3$ achieves the highest mIoU by balancing effective cross-modal information exchange and noise suppression; setting $\rho =1.0$ (no sparsity, full dense fusion) results in a 1.46% mIoU drop, confirming the importance of selective fusion for suppressing noise from contradictory cross-modal signals. The fine-grained patch size of 64 achieves the best boundary refinement performance, as smaller patches (16) lack sufficient spatial context while larger patches (128) dilute the local refinement focus.

5.7. Statistical Significance, Instance-Level Detection, and Uncertainty Analysis

We conducted three complementary analyses of statistical robustness, instance-level detection, and predictive uncertainty. (1) Repeated-run statistics and significance tests. All DIGSFNet and HAEFNet results were re-run with three random seeds;

Table 8. Repeated-run statistics (mean ± std over 3 seeds) and significance of the DIGSFNet–HAEFNet comparison.

Dataset	DIGSFNet mIoU (%)	HAEFNet mIoU (%)	p-Value
Nanning-HHLS	83.57 ± 0.18	80.94 ± 0.21	<0.01
HAEFNet benchmark	78.92 ± 0.24	75.24 ± 0.27	<0.01

reports mean ± standard deviation on both datasets, together with paired two-sided t-tests over per-image IoU between DIGSFNet and HAEFNet. DIGSFNet attains 83.57 ± 0.18 vs. 80.94 ± 0.21 mIoU on Nanning-HHLS ($p<0.01$), and 78.92 ± 0.24 vs. 75.24 ± 0.27 on the HAEFNet benchmark ($p<0.01$). (2) Instance-level detection analysis. Complementing the pixel-level metrics, a landslide instance is counted as detected if its instance-level IoU exceeds 0.5; the instance-level analysis shows that DIGSFNet detects 87/94 test instances (92.6%) versus 81/94 (86.2%) for HAEFNet on Nanning-HHLS, with instance-level false-discovery rates of 7.4% and 11.8%, respectively. (3) Confidence and uncertainty maps. Figure 8 visualizes predicted probability maps and Monte-Carlo-dropout uncertainty maps for representative ambiguous scenes (densely vegetated slopes, mining areas, small landslides). As shown in Figure 8, uncertainty concentrates at boundary transition zones and at non-landslide deformation sources, while landslide interiors are predicted with high confidence, consistent with the intended effect of the deformation-integrity priors.

6. Discussion

6.1. Advantages of Symmetric Fusion over Asymmetric Approaches

The ablation study results (Configs 4 vs. 5 in Table 5) quantitatively demonstrate the advantage of symmetric modality treatment over asymmetric fusion, with the proposed SDAE yielding a 2.47% higher mIoU and a 3.15% higher Recall than the asymmetric baseline. The pronounced gain in Recall—which corresponds to a substantial reduction in missed high-risk landslides—is particularly significant for high-risk landslide extraction, where undetected hazards can lead to severe consequences.

This advantage can be attributed to two complementary design choices in SDAE. First, the modality-aware parameter-efficient adapters allocate dedicated learnable capacity to each modality, ensuring that the InSAR branch develops specialized representations of displacement velocity gradients, coherence degradation signatures, and phase anomaly patterns rather than being implicitly forced to mimic the optical branch’s pre-trained feature hierarchy. Under an asymmetric architecture, in contrast, the InSAR auxiliary branch receives weaker gradient signals during training and tends to converge to a feature distribution that is dominated by the optical stream, thereby diluting the diagnostic value of deformation observations. Second, the dynamic sparse cross-modal fusion (DSCF) selectively exchanges complementary information only at spatially informative locations through a symmetric query–key–value formulation, preventing the optical features—which possess richer pre-trained statistics—from overwhelming the InSAR features during fusion. Together, these two mechanisms elevate InSAR from an auxiliary channel to an equal information source, allowing the network to exploit deformation signals that are essential for distinguishing actively deforming slopes from optically similar but stable terrain. The empirical recall improvement observed on the Nanning-HHLS dataset is fully consistent with this design rationale.

6.2. Effectiveness of Deformation Integrity Prior

The deformation integrity prior introduced in the DIPD addresses a critical weakness of existing segmentation methods: the production of spatially fragmented and boundary-inaccurate landslide masks. The ablation results (Configs 6–7 in Table 5) confirm the quantitative contribution of this design, with the deformation integrity loss contributing 1.71% improvement in mIoU and 4.22% in Boundary F1. Qualitatively, the deformation integrity loss serves two complementary functions. First, the intra-region consistency constraint penalizes high deformation variance within predicted landslide regions, effectively merging fragmented patches that belong to the same landslide body into spatially complete masks. Second, the boundary gradient alignment constraint guides the predicted boundaries toward deformation field discontinuities, producing physically meaningful delineation that aligns with actual geomorphological boundaries. A particularly important benefit is the suppression of false alarms from non-landslide deformation sources. Mining subsidence typically exhibits concentric deformation patterns with sharp circular boundaries, construction activities produce spatially confined and temporally abrupt deformation, and agricultural land settlement presents low-coherence diffuse patterns. These deformation signatures differ fundamentally from the continuous, gradient-consistent patterns of coherent landslide motion. The deformation integrity constraints implicitly encode these physical differences, enabling the network to correctly reject non-landslide deformation regions without requiring explicit negative class annotations. This physical interpretability distinguishes our approach from purely data-driven boundary refinement methods.

The deformation signals of potential and already-occurred landslides can be separated into three cases. (i) Post-failure deposits that have stabilized no longer produce coherent present-day deformation and are excluded by criterion (i) of the definition in Section 1; during annotation, historical-inventory polygons without a current InSAR signal were labeled as background. (ii) Reactivated old landslide bodies undergoing renewed creep satisfy all three criteria and are legitimately labeled as high-risk targets. (iii) Non-landslide deformation sources are distinguished jointly by the supervised segmentation objective and the learned cross-modal features, while the deformation-integrity regularizers refine spatial coherence and boundaries. Supplementary Figure S1 illustrates the InSAR-branch feature responses, dynamic sparse gate maps ${\mathbf{G}}_l$, and prediction evolution with and without the deformation-integrity loss for representative scenes; together with the ablation path in Table 5 (Configs 1→3→5→9: 74.12→76.42→81.03→83.57 mIoU), these results connect each performance increment to its architectural cause.

The same physical rationale informs our exclusion of DEM inputs, for two core reasons. First, for high-risk landslide extraction—as opposed to susceptibility mapping—dynamic deformation evidence of ongoing movement carries greater diagnostic value than static terrain attributes describing predisposition. Second, DEM quality is a practical concern: in Guangxi’s densely vegetated mountains, freely available DEMs suffer from substantial canopy-induced errors and temporal mismatch with the observation period. In a controlled ablation on the HAEFNet benchmark, we appended GLO-30 DEM as an extra input channel to DIGSFNet. The resulting mIoU of 79.05% (vs. 78.92% without DEM) is within the run-to-run standard deviation reported in Table 8, and the DEM ablation of the re-trained HAEFNet reported in Table 4 points in the same direction.

6.3. Limitations and Future Work

Despite the encouraging results achieved by DIGSFNet, several limitations warrant acknowledgment. First, the Nanning-HHLS dataset is geographically confined to a single region characterized by karst topography, subtropical monsoon climate, and specific landslide types. Although the experiments on the HAEFNet public benchmark—covering the Qinghai–Tibet Plateau and Sichuan active-landslide regions—show that the architecture, when independently retrained, also performs well on a geomorphologically distinct region, comprehensive validation across additional geological environments (e.g., loess plateau, volcanic terrain, coastal cliff systems) and a true zero-shot cross-region transfer experiment remain necessary to fully establish the method’s robustness. Second, the current approach represents the temporal InSAR record by a single line-of-sight summary map (cumulative displacement for Nanning-HHLS and mean velocity for the HAEFNet benchmark) rather than the full displacement time series, which may sacrifice valuable temporal dynamics of the deformation evolution process. Concretely, this compression discards acceleration signatures of progressive failure, seasonal and rainfall-driven periodicity, and the ability to distinguish steady creep from transient motion; it was nevertheless adopted here to provide a uniform input interface across datasets whose released products differ and to retain single-epoch inference for operational surveys. Modeling the full displacement time series through temporal query propagation or temporal attention mechanisms could potentially capture deformation acceleration patterns indicative of impending slope failure. Third, the proposed method performs binary segmentation (landslide vs. non-landslide) and does not predict landslide risk levels or activity stages. Future work will address these limitations through three directions: (1) temporal-aware architectures that model displacement time-series evolution for early warning applications; (2) multi-class risk-level segmentation that distinguishes low-risk, moderate-risk, and high-risk landslides; and (3) integration of geological prior knowledge graphs to enhance cross-region transferability.

7. Conclusions

This paper proposes DIGSFNet, a novel Deformation-Integrity-Guided Symmetric Fusion Network for high-risk landslide extraction from multi-source remote sensing images, which explicitly addresses the modality imbalance, spatial fragmentation, and computational efficiency challenges that hamper existing InSAR–optical fusion methods. The proposed framework introduces three key innovations: (1) the Symmetric Deformation-Aware Encoder (SDAE) treats InSAR and optical modalities as equal information sources through parameter-efficient symmetric adapters and dynamic sparse cross-modal fusion, effectively alleviating the representational imbalance inherent in asymmetric fusion approaches; (2) the Deformation Integrity Prior Decoder (DIPD) leverages InSAR deformation continuity and boundary gradient consistency as physically interpretable constraints, enabling the network to produce spatially complete and boundary-accurate landslide masks while suppressing false alarms from non-landslide deformation sources; (3) the Lightweight Deployable Student Network (LDSN) achieves efficient large-area deployment through cross-modal knowledge distillation and INT8 quantization with minimal accuracy degradation. Extensive experiments on the Nanning-HHLS dataset and the public HAEFNet multimodal active-landslide benchmark [5] covering the Qinghai–Tibet Plateau and Sichuan demonstrate that DIGSFNet consistently achieves state-of-the-art performance across mIoU, F1-score, Recall, and Boundary F1 metrics when independently trained on each of these geomorphologically distinct regions, while the lightweight student model enables practical deployment for large-area landslide hazard surveys. In future work, we will extend DIGSFNet to incorporate temporal InSAR time-series modeling and multi-class risk-level prediction, aiming toward a comprehensive landslide hazard assessment framework with enhanced temporal awareness and geological knowledge integration.