Inactive or relic landslides are often difficult to identify in optical imagery because their visual characteristics depend on multiple factors such as the time elapsed since failure, vegetation regrowth, surface erosion, and local climatic conditions, which can collectively lead to low-contrast or blurred boundaries in visible-band imagery [3,4]. To address this, multi-stream and attention-based segmentation networks have been proposed. For example, feature-fusion segmentation network (FFS-Net) combines optical texture with DEM-derived topographic cues (i.e., elevation-based terrain attributes such as slope, aspect, and curvature), significantly improving the segmentation of blurred landslides over CNN-based baselines [3]. Similarly, hyper-pixel-wise contrastive learning augmented segmentation network (HPCL-Net) introduces hyper-pixel contrastive learning to enhance feature representation in small or ambiguous landslide patches [16]. Recently, Hu et al. [12] proposed a cross-attention landslide detector (CALandDet) framework that captures global context for improved segmentation of relic landslides, outperforming CNN-based baselines.

The fusion of multi-modal inputs such as RGB (Red-Green-Blue), DEM, and slope layers has consistently improved landslide delineation. DEMs contribute elevation gradients and slope breaks that can help clarify landslide boundaries obscured in spectral imagery, although the slope information accurately reflects landslide morphology only when the DEM is acquired after the failure event. Studies using dual-stream networks, such as the gated dual-stream convolutional neural network (GDSNet), demonstrate that gated feature fusion of RGB and DEM inputs achieves higher detection accuracy than single-modality models [3,17]. The Landslide4Sense benchmark further validates this, showing that networks incorporating DEM and slope layers outperform those using only Sentinel-2 optical data across multiple global test sites [8,18]. Other approaches combine mathematical morphology with DEM and orthophoto data to extract reliable landslide polygons [19]. More recently, Wu et al. [9] proposed a hybrid CNN–Transformer fusion network that integrates DEM with optical features, achieving superior accuracy on multi-regional benchmarks.

The scarcity of pixel-wise labeled landslide data presents a critical bottleneck for developing and validating deep learning models. Supervised segmentation networks rely on accurately annotated masks to optimize loss functions, learn discriminative spatial features, and evaluate model performance across heterogeneous terrain conditions. Insufficient labeled data often lead to overfitting and poor generalization, highlighting the need for large, well-annotated datasets to advance reliable and transferable landslide mapping. Manual delineation of landslides is costly and time-consuming, leading to limited training samples. To mitigate this, researchers employ data augmentation, transfer learning, and ensemble methods. For example, Faster R-CNN with aggressive augmentation improved detection recall under limited training data [5]. Transfer learning from large-scale image datasets (e.g., ImageNet) consistently boosts performance; Chandra et al. [4] demonstrated that U-Net with ResNet backbones achieved near-perfect landslide detection accuracy, far surpassing models trained from scratch. Beyond supervised learning, unsupervised and weakly supervised strategies such as multiscale adaptation and contrastive learning are emerging to reduce reliance on extensive ground truth [7].

Recent architectures enhance accuracy through attention mechanisms, multi-scale feature fusion, and advanced decoder designs. FFS-Net integrates multiscale channel attention to balance fine textures with semantic context [3], while the method proposed by Lu et al. [20] applies lightweight multi-scale fusion to boost pixel-level precision. Advanced U-Net variants incorporating multi-sensor inputs and refined loss functions consistently outperform vanilla U-Net baselines [4,19]. Ensemble strategies and hybrid CNN-Transformer models are also being explored for boundary refinement and context modeling [7,10,11,21]. The Landslide4Sense benchmark confirms that specialized fusion-based architectures can yield IoU and F1-score improvements of 5–10% compared to earlier CNNs [8,18]. These developments underscore the importance of multi-branch, attention-augmented, and transformer-based designs for advancing landslide segmentation performance.

The framework adopts a dual-stream encoder–decoder design to exploit complementary optical and topographic features. As illustrated in Fig. 2, two geographically co-registered inputs Stream A (optical RGB) and Stream B (DEM) are processed by a siamese EfficientNet-B4 backbone [22] with shared weights for consistent feature extraction. Each encoder produces five hierarchical feature maps {fiA}i=15 and {fiB}i=15 capturing fine-to-semantic representations.

Mid-level features (f3B,f4B) from Stream B are fused with optical counterparts through a GateFuse module, strengthening semantic alignment while preserving modality-specific shallow layer details (f1,f2). Two adaptive decoders progressively reconstruct high-resolution representations via cascaded TransUp and UpFlex modules. Decoder A operates on fused features for cross-modal representations, while Decoder B reconstructs Stream B independently to retain topographic specificity. Intermediate outputs (x3A,x4A) and (x3B,x4B) undergo late-stage gated fusion for enhanced boundary precision.

The fused decoder output is upsampled via SubPixelUp and projected through 1×1 convolution to obtain main prediction y^main∈RK×H×W. Two auxiliary predictions (y^aux2,y^aux3) from intermediate stages provide deep supervision, improving gradient flow and training stability. The architecture integrates gated fusion at encoder and decoder levels, attention-guided reconstruction, and multi-scale supervision for accurate landslide segmentation.

Both input streams are processed by a similar backbone trained on ImageNet. Each encoder extracts a five-level hierarchy of multi-scale feature maps, capturing fine textures at shallow layers and semantic context at deeper layers. Intermediate features from both streams are retained as skip connections to guide the decoders during reconstruction. Stream A features (a1,…,a4) feed the fused decoder branch, while Stream B features (b1,…,b4) support the modality-specific decoder. This shared pretrained backbone accelerates convergence and stabilizes training on limited landslide samples.

The adaptive decoder comprises three complementary modules SubPixelUp, TransUp, and UpFlex that collectively address global alignment, selective fusion, and efficient upsampling. Unlike prior architectures applying upsampling and attention sequentially, this unified design integrates these mechanisms coherently. TransUp performs sub-pixel upsampling with lightweight cross-attention for global context alignment between decoder and skip features. The attention mechanism is applied directly across the skip-feature space rather than through flattened spatial tokens, enabling modality- and depth-aware fusion without excessive computational overhead. UpFlex extends classical U-Net skip connections by integrating sub-pixel upsampling with attention-gated fusion for spatial refinement. Its flexible design operates efficiently across multiple decoder scales without parameter redundancy, adaptively emphasizing salient encoder features while suppressing irrelevant noise. SubPixelUp provides the core upsampling operation, ensuring artifact-free resolution recovery with reduced computational overhead compared to transposed convolutions.

The proposed fusion design is inspired by the gated multimodal unit introduced in [24] and adapted here for pixel-level feature integration. As illustrated in Fig. 2, the GateFuse module adaptively merges complementary information from the two input streams or their decoder branches. Here, a and b represent feature maps from different modalities, each of dimension RC×H×W. A spatial attention mask α is generated as α=σ(Conv1×1([a;b])), where [a;b] denotes channel-wise concatenation and σ(⋅) is the sigmoid activation. The fused feature is then computed as (4)f=α⊙a+(1−α)⊙b,where ⊙ indicates element-wise multiplication. To promote confident fusion decisions, a regularization term is added: (5)ℛ=1HW∑i,jαi,j(1−αi,j),which penalizes uncertain gating (α≈0.5) and stabilizes multimodal learning.

To address severe class imbalance and guide confident multimodal fusion, the network employs a compound loss combining Tversky segmentation loss [25] with the gating regularization term in Eq. (5). This formulation stabilizes training, improves recall of sparse landslide pixels, and enforces decisive gating across fusion stages. Following the deep supervision strategy in [26], the model generates a main prediction y^main and two auxiliary outputs (y^aux2,y^aux3) from intermediate decoder stages. The segmentation loss is then formulated as (6)ℒtversky=λ0ℓ(y^main,y)+λ1ℓ(y^aux2,y)+λ2ℓ(y^aux3,y),where ℓ(⋅) denotes the pixel-wise Tversky loss with hyperparameters α=0.3 and β=0.7 to emphasize recall over precision, particularly beneficial for detecting sparse landslide regions. To further stabilize the gating mechanism and prevent unreliable fusion weights, a regularization term with weight λr=10−3 is incorporated, yielding the total objective function (7)ℒ=ℒtversky+λr∑jℛj,where ℛj denotes the gating regularization at fusion stage j. This composite loss effectively balances segmentation accuracy with stable multimodal integration throughout the network hierarchy.

While large-scale datasets such as the Chinese Academy of Sciences (CAS) Landslide Dataset [27] provide multi-sensor coverage across diverse terrains, co-registered multi-modal datasets remain limited [28]. Optical data suffer from cloud cover and daylight constraints, whereas SAR can penetrate clouds [29], yet few datasets provide co-registered SAR or DEM for effective multi-modal fusion.

The Bijie dataset [30] is an optical benchmark from Bijie City, Guizhou Province, China, covering 26,853 km2 in the Tibetan Plateau transitional zone (altitude 457–2900 m). This landslide-prone region features steep terrain, unstable geology, and heavy rainfall (849–1399 mm annually). The dataset comprises 770 positive landslide samples (rock falls, rock slides, debris slides) and 2003 non-landslide samples from TripleSat imagery (May–August 2018, 0.8 m resolution). Each sample includes RGB imagery, manually delineated masks verified through field surveys, and DEM tiles (2 m elevation accuracy). Non-landslide samples represent diverse backgrounds (mountains, villages, rivers, forests, agricultural lands). Despite its value as a benchmark, coverage is limited to one region and acquisition period.

The LandSlide4Sense dataset [8] is a global benchmark comprising Sentinel-2 multispectral imagery (14 bands at 10 m resolution) and DEM-derived features from the Advanced Land Observing Satellite Phased Array type L-band Synthetic Aperture Radar (ALOS PALSAR). It provides 3799 training, 245 validation, and 800 test patches (128×128 pixels) with pixel-wise masks for training. Spanning multiple global regions (2015–2021), it captures shallow translational slides and debris flows caused by rainfall, erosion, or anthropogenic disturbances. While offering greater environmental diversity than site-specific datasets, its 10 m resolution may miss small landslides, and the patch-based design limits contextual information.

Landslide segmentation is formulated as binary classification with strong class imbalance. Performance is evaluated using four pixel-level metrics: (8)Precision=TPTP+FP,Recall=TPTP+FN,F1=2TP2TP+FP+FN,IoU=TPTP+FP+FN,where precision and recall measure accuracy and completeness, while F1 and IoU quantify overall segmentation quality. Two image-level metrics assess threshold-independent robustness: AUROC evaluates ranking consistency, and AUPRC summarizes precision–recall trade-offs: (9)AUROC=∫01TPR(t)dFPR(t),AUPRC=∫01Precision(r)dRecall(r).

Together, these metrics provide comprehensive evaluation of pixel-wise accuracy and model reliability under class imbalance.

To ensure a comprehensive and fair evaluation, we compared the proposed model with a diverse set of segmentation architectures representing the major evolutionary trends in semantic segmentation for landslides. The selected baselines include both classical convolutional networks and recent attention or transformer-based designs that have been widely adopted in remote sensing and landslide mapping. This diversity allows us to assess the relative contribution of multimodal fusion, attention mechanisms, and adaptive decoding within a unified benchmark.

Among convolutional encoder–decoder networks, U-Net [2], DeepLabv3+ [3] and LinkNet [31] serve as foundational CNN baselines known for their efficiency and strong localization performance. Dual-Stream U-Net [32,33] extends this paradigm to multimodal and multi-temporal settings, providing a reference for assessing the benefit of dual-stream fusion. More recent variants, such as the residual-multihead-attention unet (RMAU-NET) [34] and the dilated convolution and EMA attention with pixel attention-guided unet (DEP-UNet) [35] incorporate residual and dense connections, multi-head attention, and pyramid pooling to enhance multi-scale context and mitigate class imbalance. Together, these architectures establish a strong baseline for evaluating improvements introduced by our adaptive decoder and gated fusion mechanisms.

We further include state-of-the-art attention and transformer-based models to represent the current frontier of landslide segmentation. ShapeFormer [21] leverages vision transformers and shape priors for boundary-aware segmentation, while the enhanced multi-scale residual high-resolution network (EMR-HRNet) [11] fuses multi-resolution features through an efficient refinement module built upon HRNet. The global information extraction and multi-scale feature fusion segmentation network (GMNet) [36] adopts a multi-branch design combining global, local, and morphological features via a gated fusion mechanism. Although these models capture richer context through self-attention and multi-scale fusion, they typically operate on single-modality optical data and lack explicit cross-stream gating or deep supervision. Our framework complements these approaches by unifying attention, multimodal fusion, and adaptive decoding into a lightweight dual-stream architecture specifically tailored for landslide segmentation.

For the Bijie dataset, the landslide and non-landslide samples were merged and split into 70%, 20%, and 10% subsets for training, validation, and testing, yielding 1941, 554, and 278 samples, respectively. For the Landslide4Sense dataset, we adopted the official split of 3799 training, 245 validation, and 800 test images as defined by the dataset authors. All images were resized to 256×256 pixels for convolutional backbones and 224×224 for vision transformers, with standard augmentations (random flips, Gaussian noise, and Contrast Limited Adaptive Histogram Equalization (CLAHE)) applied during training.

In the Bijie dataset, each patch provides RGB and DEM layers. Stream A receives the RGB composite, while Stream B uses the DEM replicated three times, forming paired six-channel inputs that preserve architectural symmetry. For the Landslide4Sense dataset, Stream A uses RGB channels, and Stream B combines normalized difference vegetation index (NDVI), slope, and DEM features. This configuration efficiently combines optical, topographic, and vegetation information, balancing performance and computational cost. The NDVI is computed from the red (R) and near-infrared (NIR) bands as: (10)NDVI=NIR−RNIR+R.

This configuration allows both datasets to be processed in a consistent dual-stream manner, while leveraging complementary spectral and topographic information for landslide segmentation.

All experiments were conducted in Python 3.10 using PyTorch 2.6.10 on Ubuntu 22.04.4 LTS, with an Intel(R) Xeon(R) Platinum 8350C CPU @ 2.60 GHz, an NVIDIA RTX A6000 GPU (Graphics Processing Unit) (48 GB). The hyperparameters used are detailed in Table 1.

The Adam optimizer was employed with an initial learning rate of 3×10−4 and weight decay of 1×10−4. Models were trained for 100 epochs with a batch size of 32. Tversky loss was used to address class imbalance, with sigmoid activation in the final layer for probabilistic predictions, ReLU in convolutional modules, and softmax in the cross-attention mechanism.

This section presents quantitative and qualitative results on the Bijie and Landslide4Sense datasets. Training and validation curves are provided in Appendix A (Figs. A1 and A2).

We conduct ablation experiments to evaluate the contribution of individual components within the proposed framework, including the choice of backbone network, the effect of fusion strategies, the role of decoder components, and the impact of deep supervision. All experiments were conducted under identical settings for a fair comparison on Bijie dataset.