Hybrid lightweight transformer for efficient landslide change detection in remote sensing imagery Authors: Dhirendra Prasad Yadav, Bhisham Sharma, Shashank Singh, Panos Liatsis Published: 2025 ABSTRACT Remote sensing images may be used to detect and segment landslide areas. Recently, encoder and decoder-based convolutional neural networks (CNN) have been used to localise landslide regions. However, CNN-based methods extract shallow spatial features and lack the ability to capture long range dependencies. Due to this, edges and boundary regions of landslides are less explored. In this study, we present the change detection convolutional transformer network (CDCTNet), a hybrid model designed using CNN and transformer-based attention to identify and segment landslides. CDCTNet contains a lightweight convolutional encoder for spatial feature extraction. Local self-attention (LSA) and global self-attention (GSA) blocks are incorporated into the transformer to provide local attention and global information exchange with minimal overhead. The decoder consists of convolution blocks and a gated attention block that maintains long-range dependencies and provides for high processing speed to generate localized landslide region masks. The proposed model is experimentally evaluated on the Debrisflow, Landslide and Bijie datasets and compared with six state-of-the-art (SOTA) methods. CDCTNet achieved F1-scores of 76.56%, 72.06% and 94.85%, respectively, on the three datasets, surpassing previous methods with less computation overhead. In addition, CDCTNet requires lower flops and fewer trainable parameters compared to the transformer, ensuring feasibility in cloud-based real-time applications. Introduction Landslide incidents have increased dramatically due to human activities in mountainous regions and extreme weather occurrences due to global climate change1. Traditional methods such as field surveys, optical image-based remote sensing (RS) and global navigation satellite systems (GNSS) are helpful in landslide monitoring. However, these techniques face several challenges, including computation costs, low efficiency and inaccurate real-time monitoring2. Interferometric synthetic aperture radar (InSAR) is rapidly growing in geohazard recognition and surface observation due to its ability to provide high precision in all weather conditions in a short period3. Achache et al.4 proposed that SAR interferometry accurately monitors both large- and small-scale surface displacements, as demonstrated by ERS-1 SAR images of the 1991 Saint-Etienne-de-Tinée landslide. The technique provides detailed, consistent results comparable to ground measurements without additional costs. Torres et al.5 used ESA’s Sentinel-1 radar observatory to provide all-weather observations. This supports applications such as emergency response, ice monitoring, marine surveillance, and interferometric tasks. The mission ensures free, open data access with confirmed quality and performance. While InSAR provides valuable data for monitoring surface deformations, machine learning (ML)-based approaches have emerged as an efficient automated alternative for landslide detection. However, ML often relies on handcrafted features and expert knowledge, which may introduce bias and limit generalizability6,7. In contrast, CNN approaches automatically extract hierarchical features from RS images to detect landslide regions, potentially improving performance over manual methods, by reducing the need for domain-specific expertise. While handcrafted features could capture important information, they are prone to high error rates due to subjectivity and reliability. CNN-based systems can lead to improvements in performance through their automatic feature extraction capability. Despite these advantages, CNN typically extract shallow or local features, lacking the ability to capture global correlations. To address these limitations, researchers designed hybrid models incorporating residual blocks with vision transformers (ViT), which enhance feature maps by fusing the strengths of local convolutions with global self-attention. However, these architectures require a substantial amount of training data, while accommodating a large number of trainable parameters, which, in conjunction with the computational requirements of ViT multi-head self-attention (MHSA), make them less suitable for real-time landslide monitoring. In light of the above challenges, we propose CDCTNet, a hybrid network comprising a lightweight convolutional encoder for efficient spatial feature extraction, LSA and GSA blocks integrated in the transformer module, providing for focused local attention and global information exchange with minimal computational overhead. The decoder incorporates convolutional blocks and a gated attention block that maintains long-range dependencies and supports high processing speeds to produce localized landslide region masks. CDCTNet is evaluated on the three datasets and its performance is compared to other CNN- and ViT-based methods. The experiments demonstrate that the proposed approach achieved improvements in terms of quantitative and visual results. The main contribution of this research are as follows: We designed a 2D convolutional encoder and decoder using depth-wise convolutions to reduce computational costs, enabling real-time application for problems such as landslide detection.Spatial features extracted from the convolutional blocks are flattened, followed by Conditional Position Encoding (CPE), applied via a Position-Encoding Generator (PEG), to dynamically generate position-aware tokens at each transformer stage, enabling processing of varying input image resolutions.The traditional MHSA in the ViT encoder is replaced by the LSA and GSA blocks, improving efficiency in capturing both local and global contextual information for landslide features.A gated attention module is integrated into the decoder blocks for feature map refinement and enhancement, improving overall detection accuracy, while maintaining processing efficiency. We designed a 2D convolutional encoder and decoder using depth-wise convolutions to reduce computational costs, enabling real-time application for problems such as landslide detection. We designed a 2D convolutional encoder and decoder using depth-wise convolutions to reduce computational costs, enabling real-time application for problems such as landslide detection. Spatial features extracted from the convolutional blocks are flattened, followed by Conditional Position Encoding (CPE), applied via a Position-Encoding Generator (PEG), to dynamically generate position-aware tokens at each transformer stage, enabling processing of varying input image resolutions. Spatial features extracted from the convolutional blocks are flattened, followed by Conditional Position Encoding (CPE), applied via a Position-Encoding Generator (PEG), to dynamically generate position-aware tokens at each transformer stage, enabling processing of varying input image resolutions. The traditional MHSA in the ViT encoder is replaced by the LSA and GSA blocks, improving efficiency in capturing both local and global contextual information for landslide features. The traditional MHSA in the ViT encoder is replaced by the LSA and GSA blocks, improving efficiency in capturing both local and global contextual information for landslide features. A gated attention module is integrated into the decoder blocks for feature map refinement and enhancement, improving overall detection accuracy, while maintaining processing efficiency. A gated attention module is integrated into the decoder blocks for feature map refinement and enhancement, improving overall detection accuracy, while maintaining processing efficiency. Literature survey Covello et al.8 designed COSMO-SkyMed, an earth observation system funded by ASI and the Ministry of Defence, featuring a constellation of four X-band SAR-equipped satellites. Launched in 2007 and 2008, three are operational, providing global services for risk management, scientific, commercial, and defence applications. The commissioning results of the first two satellites and the Scientific Mission Exploitation strategy were outlined. Kankaku et al.9 developed JAXA using ALOS-2 with the advanced PALSAR-2 L-band SAR, succeeding ALOS “Daichi”. PALSAR-2 improves spatial resolution and disaster management capabilities with enhancements such as maximum bandwidth observation, spotlight mode, and dual receiving antenna system. Zhang et al.10 successfully launched HJ-1-C, China’s first civil Synthetic Aperture Radar (SAR) satellite. It’s the world’s first operational S-band SAR satellite, demonstrating superior disaster reduction potential with detailed characterization capabilities compared to its predecessors, HJ-1-A and HJ-1-B. Ye et al.11 designed differential InSAR, a cost-effective method for precise centimetre-scale ground motion monitoring. A challenge is the temporal change in ground reflectivity. They address issues like obtaining the accurate phase of a corner reflector, co-registering low-coherence pixels, selecting an interpolation kernel for SAR image resampling, and computing interferometric phase without flat earth term and height contribution from corners. Using pixel-level classification, Danneels et al.12 identified landslides in the Gultcha region of Kyrgyzstan in the country’s southwestern corner. This method compared the classification performance of artificial neural networks (ANN) and the maximum likelihood ratio-based approach. To locate potentially hazardous areas for landslides on Penang Island, Pradhan and Lee et al.13 employed three ANN, logistic regression and the frequency ratio on the Haiti, Messina and Barcelonnette datasets to detect landslides. Stumpf and Kerle et al.14 jointly explored random forest (RF) and an object-oriented method. They extracted twenty object-oriented features, including grey level co-occurrence matrix, slope, hill shade and other relevant features. ANN, support vector machines (SVM), and object-oriented methods were utilized by Moosavi et al.15 to detect landslides in the Iranian city of Kermanshah. They improved classification accuracy using the Taguchi method by optimizing the selected features. Gorsevski et al.16 utilized LiDAR-based digital elevation model (DEM) derivatives to detect shallow landslides through ANN and geographical information systems (GIS). Their model accurately predicted landslides in the very high susceptibility class (42.6% accuracy, covering 1.56% of the area) but had lower accuracy in the very low susceptibility class (1.20% of known landslides in 82.68% of the area). Mezaal et al.17 utilized three feature selection techniques (CFS, RF, ACO) and a fuzzy-based segmentation parameter (FbSP) optimizer to differentiate landslide types. RF evaluated their performance, with correlation-based feature selection (CFS) ranking highest in distinguishing landslide types. The method proved suitable for differentiating between shallow and deep-seated landslides, showcasing excellent transferability. Hu et al.18 proposed cross-validated ML models for landslide detection in “off-site” areas. SVM and ANN showed reduced accuracy outside their development areas, while the RF model exhibited stable portability with superior performance in new areas. Tavakkoli et al.19 developed object-based image analysis (OBIA) with ML methods, including ANN and RF, for landslide detection. They optimized the features and tested in Nepal, achieving overall accuracies of 83.3% to 87.2% with optimal parameters. DST significantly increased overall accuracy to almost 90%. Dou et al.20 evaluate SVM hybrid ensemble machine learning algorithms for predicting rainfall-induced landslides in Northern Kyushu island, Japan. Models were validated using a landslide inventory map and 13 predictors with area under the curve (AUC) values ranging from 0.74 to 0.91. SVM boosting outperformed other models, highlighting the importance of model selection. Information Gain Ratio (IGR) identified rainfall as a crucial factor in landslide occurrence. Dias et al.21 compare image classifiers (Maximum Likelihood, RF, SVM) for landslide identification in Itaóca, São Paulo, using RapidEye-5 (5 m) and SRTM DEM (12.5 m). SVM with six bands achieved the best results, identifying a 399,325 m² landslide area. While successful, it primarily captured larger landslides, suggesting further analysis with finer spatial resolution images in tropical environments. Rajabi et al.22 applied ANN to assess landslide risk in the Manjil-Rudbar earthquake-affected area in Iran (1990, M = 7.7). A hazard map was developed using Newmark displacement analysis for a 309 km2 area near the epicentre, considering soil properties, distance from the hypocenter, slope, and earthquake magnitude. The hazard map was compared with an inventory map and other research findings. Wang et al.23 proposed an automated co-seismic landslide mapping approach with key advantages, i.e., eliminating manual labelling, introducing multi-scale extended morphological profiles, and employing an ensemble strategy for improved accuracy. Youssef et al.24 utilized SVM, 1D-CNN, and 2D-CNN for landslide susceptibility mapping in the Asir Region, Saudi Arabia. They developed a landslide inventory map with 181 points, dividing them into 70% training and 30% validation datasets, and employed thirteen landslide indicator factors (LIF) for mapping. Although these techniques have shown promise in land cover change detection (LCCD), they still need global feature extraction to enhance performance. These approaches often ignore spatial features and the links between high and low-level features. The derived features are typically less robust to pseudo-changes since they are sensitive to noise, angle, shadow, and context. Thus, many enhanced algorithms were proposed to encode image context in the time-space dimensions better and improve feature discrimination ability, such as stacking more convolution layers, using dilated convolutions25, combining GAN models26, RNN27, and attention mechanisms28. Agrawal et al.29 utilized deep learning (DL) and time series data to predict landslides in a 15-meter area. Kimothi et al.30 utilized Unmanned Aerial Vehicles (UAV) to capture glacier data and applied a DL model to develop a real-time monitoring system. Chen et al.31 utilized a Swin transformer block for the features from the land region. In addition, the feature map was improved with a multiscale attention block using a fusion technique. Li et al.32 utilized a hybrid model using CNN and ViT encoder for landslide detection. Their model achieved a precision of 80.30% on a landslide dataset. Gao et al.33 applied multiscale attention using the Swin transformer for precise landslide change detection. They test the model on the Hengduan Mountains and post-earthquake Hokkaido; their method outperformed seven recent approaches with an IOU value gain greater than 5.36%, confirming its effectiveness for large-scale landslide extraction. Pirasteh et al.34 utilized the Swin transformer and two other modules for local features to enhance the spatial feature map. Their model obtained an 89.90% precision value. Chen et al.35 utilized interferometric synthetic aperture radar (InSAR) for monitoring slow-moving landslides. They applied a multiscale Swin Transformer InSAR detection network (MSIDNet) incorporating spatiotemporal attention for improved feature extraction and fusion. Further, the model was trained on a newly constructed dataset that integrates deformation rate, phase gradient, and the gradient coherence index (GRCI); the model outperformed Faster R-CNN and YOLOv3, achieving 0.89 precision and demonstrating strong generalization. Lv et al.36 applied ShapeFormer, a shape-enhanced ViT, to address the multiscale and irregular shapes of landslides in remote sensing images. Using a pyramid vision transformer (PVT) backbone with a shape extraction branch and a deconvolution-based decoder, they achieved superior performance on the Bijie and Nepal datasets. Some researchers design their models using residual blocks with ViT for feature map enhancement. However, the high volume of training data, many trainable parameters and the computation burden of the MHSA in the classical ViT make them less suitable for real-time landslide monitoring. Nevertheless, such approaches immediately process the full-resolution change map through up-sampling, ignoring the significance of high-resolution shallow information, which causes low-level feature loss. The aforementioned works suggest that the complex nature of landslide areas is a challenging problem for computer vision-based systems. Proposed method In the proposed study, CDCTNet is developed, which combines convolution and transformer blocks. For the given image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:I\in\:{R}^{H\times\:D\times\:B}\:$$\end{document}having spatial resolution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:H\times\:D$$\end{document} and bands B, we generate a feature map to predict segmented images of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:H\times\:D\times\:1$$\end{document}. The architecture of CDCTNet is depicted in Fig. 1. CDCTNet consists of a convolutional block, which works as an encoder and extract high-level semantic features. After that, a transformer block is added that provides long-range dependencies on the spatial features. Finally, it contains decoder blocks mainly consisting of convolutional blocks that decode the features and generate the labelled feature map. Fig. 1Proposed CDCTNet architecture for landslide change detection. Proposed CDCTNet architecture for landslide change detection. Convolutional encoder block The Convolutional Encoder Block (CEB) incorporates four sequential down sampling blocks, conv1, conv2, conv3 and conv4, to extract spatial features. Each convolution block has a kernel size of 3 × 3 with increasing channel sizes of 32, 64, 128 and 256, respectively, followed by batch normalization and ReLU activation function, which reduces the spatial dimension from 224 × 224 to 14 × 14 at the bottleneck. The activated feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}_{j,k}^{x,y}$$\end{document} in the encoder module at spatial position (x, y) of the k-th layer is as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}_{j,k}^{x,y}={\Phi\:}\left(\sum\:_{1}\:\:\sum\:_{h=0}^{{H}_{j}-1}\:\:\sum\:_{w=0}^{{W}_{j}-1}\:\:{w}_{j,k,1}^{h,d}{A}_{j-\mathrm{1,1}}^{x+h,y+d}+{b}_{j,k}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{j,k,1}^{h,d}\:I\in\:{R}^{H\times\:D\times\:B}$$\end{document} is the weight parameters of the k-th connected feature map and H, W represents the height and width of the 2D convolution kernel. The symbol \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varPhi\:$$\end{document} represents the activation function ReLU used to ensure non-linearity in the model. We followed the architecture of U-Net to extract semantic global features. For the input, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:I\in\:{R}^{H\times\:D\times\:B}$$\end{document} the feature map obtained by CEB \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F\in\:{R}^{{H}_{0}\times\:{D}_{0}\times\:{B}_{0}}$$\end{document} is as follows:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F={E}_{\theta\:}\left(I\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document} are the parameters of the encoder (E). The extracted features are flattened and tokens are generated for the transformer block. Vision transformer The transformer block captures long-range dependencies within the contextual features of the landslide images. We incorporated a transformer module after the CEB to improve the spatial feature map. Because fixed-length input sequences and direct feature extraction pose limitations in absolute positional encoding, we utilised CPE on the inputs, dynamically generating tokens through PEG at each transformer stage, as shown in Fig. 2. Figure 2(a) depicts that the input feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{e}\in\:{R}^{{H}^{{\prime\:}}\times\:{W}^{{\prime\:}}\times\:C}$$\end{document}, where H’=W’=14 × 14 and C = 256 is partitioned into non-overlapping sub-windows of size 7 × 7. Each sub-window is treated as a local patch and conditional adaptive positional encoding is added to generate tokens \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:T\in\:{R}^{N\times\:d}$$\end{document}, where, N = H’W’ and d = C. The generated tokens are capable of handling varying resolutions and provide a long-range sequence for the model’s training. Additionally, PEG, as shown in Fig. 2 (b), incorporates a 2D depth-wise convolution without batch normalisation, which generates position-aware embeddings for each sub window. After this a shifted window mechanism similar to the Swin transformer is applied. This enables self-attention communication to occur exclusively within each section. Furthermore, we devised a communication pathway, similar to the Swin transformer, to facilitate communication between different sections. This is necessary because processing information only locally, results in a limited receptive field and significantly inferior performance. Fig. 2Illustration of (a) CPE and (b) PEG. Illustration of (a) CPE and (b) PEG. Locally grouped self-attention The LSA block is designed to maintain computational efficiency and capture local contextual information within the landslide image. We first partition the 2D feature space into smaller sub-windows so that SA interactions can only occur within it. By doing so, the model focuses on capturing fine-grained details within a local region of the image. In addition, it also captures better spatial dependency and complex structure details present in the landslide region. Furthermore, LSA weighs the relevance of different pixels based on their contextual relationships, allowing it to prioritize informative regions while suppressing noisy ones. Moreover, in classical ViT, MHSA is computed where pixels can communicate only through the channels in their own heads by dividing into p x q sub-windows. The SA computation cost in the ViT window is O(HDpqd), while the LSA overall cost is O(HD2d/pq), as each group comprises HD/pq components. Global sub-sampled attention The global sub-sampled attention (GSA) block captures the global contextual features of the landslide image. We selectively sample the subset in the GSA block to perform attention. Through this mechanism, computational costs for the self-attention mechanism are reduced. GSA is seamlessly integrated into the transformer architecture, replacing conventional self-attention mechanisms. This integration ensures that the global contextual information captured by GSA is appropriately incorporated into the model’s decision-making process, enhancing its overall performance. The exchange of cross-group information by adding additional conventional GSA attention layers after each block of local attention is possible. However, the computational complexity is O(H2D2d). We can reduce the cost to O(pqHDd) by using p x q sub-windows and interacting with other sub-windows having computational costs of O((HDk1k2)/(pq))2d), where k1, k2 are sub-sampling factors. At the same time, we treat LSA and GSA as separable convolutional layers (depth-wise + point-wise). Further, the variables used in the Eq. 3 are defined in Table 1. Table 1Variables and their meaning with dimension.VariableDimensionMeaning \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{jk}^{1-1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{{k}_{1}\times\:{k}_{2}\times\:B}$$\end{document} Input feature map of the (l-1)-th layer for j-th sub window and k-th group. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\prime }{Z}}_{jk}^{1}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{{k}_{1}\times\:{k}_{2}\times\:B}$$\end{document} Feature map after self-attention and residual connections \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\prime }{\mathrm{Z}}}^{1+1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{{k}_{1}\times\:{k}_{2}\times\:B}$$\end{document} Output local feature map block after applying feed forward network and residual connections \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\wedge\:}{Z}}^{l+1}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:\:\:R}^{H\times\:W\times\:B}$$\end{document} Intermediate global feature map obtained after applying GSA and residual connections \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{Z}}^{1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:\:R}^{H\times\:W\times\:B}$$\end{document} Reconstructed global feature map after stacking \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\wedge\:}{Z}}_{jk}^{l}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{Z}}^{1+1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\:{R}^{H\times\:W\times\:B}$$\end{document} Final feature map obtained after FFN and LN Table 1Variables and their meaning with dimension.VariableDimensionMeaning \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{jk}^{1-1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{{k}_{1}\times\:{k}_{2}\times\:B}$$\end{document} Input feature map of the (l-1)-th layer for j-th sub window and k-th group. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\prime }{Z}}_{jk}^{1}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{{k}_{1}\times\:{k}_{2}\times\:B}$$\end{document} Feature map after self-attention and residual connections \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\prime }{\mathrm{Z}}}^{1+1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{{k}_{1}\times\:{k}_{2}\times\:B}$$\end{document} Output local feature map block after applying feed forward network and residual connections \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\wedge\:}{Z}}^{l+1}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:\:\:R}^{H\times\:W\times\:B}$$\end{document} Intermediate global feature map obtained after applying GSA and residual connections \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{Z}}^{1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:\:R}^{H\times\:W\times\:B}$$\end{document} Reconstructed global feature map after stacking \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\wedge\:}{Z}}_{jk}^{l}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{Z}}^{1+1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\:{R}^{H\times\:W\times\:B}$$\end{document} Final feature map obtained after FFN and LN Variables and their meaning with dimension. The SSSA (spatially separable self-attention) is mathematically defined as follows.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\stackrel{\wedge}{Z}}_{jk}^{\ell}= LSA\left( LN\left( Z_{jk}^{\ell-1} \right) \right) + Z_{jk}^{\ell-1}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Z}_{jk}^{\ell}=FFN\left(LN\left({\stackrel{\wedge }{Z}}_{jk}^{\ell}\right)\right)+{\stackrel{\wedge }{Z}}_{jk}^{\ell},$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\wedge }{Z}}^{\ell+1}=\mathrm{G}\mathrm{S}\mathrm{A}\left(\mathrm{L}\mathrm{N}\left({Z}^{\ell}\right)\right)+{Z}^{\ell},$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Z}^{1+1}=FFN\left(LN\left({\stackrel{\wedge }{Z}}^{\ell+1}\right)\right)+{\stackrel{\wedge }{Z}}^{\ell+1},$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j\in\:\{\mathrm{1,2},3,\dots\:\text{}m\},k\in\:\{\mathrm{1,2},3\dots\:\dots\:n\}$$\end{document}. where GSA is the global sub-sampled attention by interacting with the representative keys from each sub-window zˆij ∈ R k1×K2×B, LN = Layer normalization, FFN = Feed forward network. Consider a post-CEB feature map of dimensions 14 × 14 × 256, where the height and width are both equal to 14 and the feature dimension is 256. This map is partitioned into 7 × 7 sub-windows, each of size 2 × 2 tokens. With pooling parameters set as k1 = k2 ​=2, each zˆij ∈ R k1×k2×B aggregates the corresponding 2 × 2 grid into 4 averaged tokens using a stride of 1. As a result, the global attention module processes a total of 196 tokens (49 windows with 4 tokens each), leading to a computational complexity of O ((196)2.64), which is approximately 2.45 million operations when assuming dk=d/4 = 64. In comparison, a standard ViT requires O ((196)2.256), equivalent to about 9.8 million operations, thereby reducing the cost by a factor of 4. The SA mechanism in the classical ViT is calculated based on all the tokens of the input sequence. In addition, to attaining global attention, every token interacts with other tokens, which leads to quadratic complexity, and it may miss local spatial information. The LSA block divides the spatial feature map into sub-windows, and attention is calculated within this to local contextual information. At the same time, the GSA block selectively samples representative tokens and performs attention computations among them to provide global contextual information. The computational cost of self-attention in ViT is O (N²d + Nd²), where N is the number of tokens and d is the embedding dimension. The combined complexity of LSA and GSA is O (N²d/P + pqNd + Nd²), where P is the partition size for local attention and pq represents the sub window size. Attention gate In computer vision, the adaptability of attention mechanisms to shift emphasis on essential portions of an image, while ignoring unimportant ones has long been established. We designed a two-level embedding into the residual connections for the decoder, as illustrated in Fig. 3. Fig. 3Illustration of gated attention block for each decoding path. Illustration of gated attention block for each decoding path. The schematic illustrates the integration of a gated attention mechanism at each decoder stage. In this design, upsampled features are progressively refined by combining encoder skip connections with transformer outputs, thereby enhancing feature representation and improving the accuracy of the reconstruction process. From the encoded CNN features \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F\in\:{R}^{{H}_{0}\times\:{D}_{0}\times\:{B}_{0}}\:\:$$\end{document}and the transformer encoder spatial features \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Z\in\:{R}^{{H}_{0}\times\:{D}_{0}\times\:{B}_{0}}\:$$\end{document}the 1D channel attention map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{b}\in\:{R}^{B\times\:1\times\:1}$$\end{document} is calculated that dynamically recalibrates the object recognition features for the two level attention maps. Further, a 2D spatial normalization (2DSN) map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{S}\in\:{R}^{1\times\:{H}_{0}\times\:{D}_{0}}\:$$\end{document}is calculated and fed to the transformer module. The complete mathematical process for the attention map is summarized as follows:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{}F}_{cn}={w}_{b}\left(f\right)\otimes\:f,$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{}F}_{sn}={w}_{s}\otimes\:{\text{}F}_{cn}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\otimes$$\end{document} = Element wise multiplication, f = Features obtained from the CNN and Transformer. To effectively leverage the inter-channel connection of feature maps, we develop channel normalisation\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{cn}$$\end{document} which recalibrates the weight of each channel. We applied global average pooling on each layer followed by a fully connected layer, after which we scale each channel appropriately:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{}F}_{cn}=\sigma\:\left(MLP\right(AvgPool\left(f\right)\left)\right),$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{cn}=\sigma\:\left(w_2\left(w_1\left({f}_{\mathrm{avg}}^{b}\right)\right)\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_2$$\end{document} are MLP weights and σ = sigmoid activation function. Gated convolutional decoder In the decoder, only a CNN architecture has been utilized. The decoder block upsamples the feature map to generate a space with the same dimensionality as the original image, i.e.,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:I\in\:{R}^{H\times\:D\times\:B}$$\end{document}. Furthermore, transpose convolution is applied to recover the image resolution from the layer below it, so that pixel-level segmentation can be carried out. In addition, the feature map obtained from the spatial normalization is passed through 2DSN for effective global information exchange. These feature maps contain high-level semantic information, and we use attention gates to access this information in the upsampling layers, as shown in Fig. 1. Research suggests that skip-connections can result in higher feature fusion without a considerable increase in computational costs. The loss of each model is calculated using a hybrid loss function, i.e., binary cross entropy, dice and focal loss function. Dice loss helps to distinguish the overlapping region, while focal loss mitigates class imbalance, i.e., landslide vs. non-landslide regions. The hybrid loss function is defined as:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{HLoss}\text{}=\alpha\:BL+\beta\:DL+\gamma\:FL$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{B}\mathrm{L}=-\frac{1}{\mathrm{N}}\sum\:_{\mathrm{j}=1}^{\mathrm{N}}\:\:\left[{\mathrm{L}}_{\mathrm{j}}\mathrm{log}\left({\mathrm{p}}_{\mathrm{j}}\right)+\left(1-{\mathrm{L}}_{\mathrm{j}}\right)\mathrm{log}\left(1-{\mathrm{p}}_{\mathrm{j}}\right)\right],$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{D}\mathrm{L}=1-\frac{\sum\:_{\mathrm{j}=1}^{\mathrm{N}}\:\:{\mathrm{p}}_{\mathrm{j}}{\mathrm{L}}_{\mathrm{j}}+{\upepsilon\:}}{\sum\:_{\mathrm{j}=1}^{\mathrm{N}}\:\:{\mathrm{p}}_{\mathrm{j}}+\sum\:_{\mathrm{j}=1}^{\mathrm{N}}\:\:{\mathrm{L}}_{\mathrm{j}}+{\upepsilon\:}},$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:FL=-\alpha\:{\left(1-{p}_{t}\right)}^{\lambda\:}log\left({p}_{t}\right),$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{t}=\left\{\begin{array}{c}P,y=1\\\:1-P,y=0\end{array}\right\}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_j$$\end{document} =label of jth pixel, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_j$$\end{document} =Probability value of the jth pixel, N = Total number of pixels in a batch, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:,\beta\:,\gamma\:,\lambda\:,\epsilon\:\:$$\end{document}= hyperparameters, BL = Binary cross entropy loss, DL = Dice loss and FL = Focal loss. The algorithm for the land slide change detection is shown in Algorithm 1. Algorithm 1Proposed method for landslide change detection. Proposed method for landslide change detection. Experiments In this section, we present the description of the dataset, experimental settings, quantitative and visual results. Datasets and data augmentation In the proposed study, we performed experiments using three datasets to test the method’s superiority. Datasets DS1 (Debrisflow) and DS2 (Landslide) are high-precision aerial imagery of the debris flow disaster in Sichuan and surrounding areas since 200837. The high-resolution digital orthophoto map (0.2 m ~ 0.9 m) datasets of landslides and debris flows are generated. They contain 107 representative images, annotation data, and description files for earthquake, rainfall, gully, and slope debris flows. The datasets encompass the “5·12” Wenchuan, “4·20” Lushan, “8·8” Jiuzhaigou, and Jinsha and Dadu River earthquakes. High-precision image data sources, geohazard expert interpretation and annotation, and thorough disaster information make up this high-quality data collection. Next, we use the Bijie landslide dataset (DS3), which is an open remote sensing landslide dataset. DS3 contains optical images from satellites, shape files that show the edges of landslides, and digital elevation models. It has 770 images of landslides (red points) and 2003 other images, which were cropped from the TripleSat satellite38. DS1 and DS2 contain a smaller number of images, which may lead to overfitting, while DS3 contains 700 images. We randomly partitioned each of the datasets into 80% for training, and 20% for testing. Next, we applied data augmentation on the training images of DS1 and DS2 to generate additional samples. This strategy avoids potential concerns for data leakage. For DS1 and DS2, we used horizontal flip (HP), vertical flip (VP) and rotations. Following data augmentation in each of these datasets, 1000 images were retained for training. Experimental setup and performance metrics We conducted an experiment on NVidia Quadro TRX 4000 GPU, with Windows 10 operating system and 128 GB RAM using Python 3.8. For each experiment, the model is trained in a mini-batch size of 16 for 100 epochs using images of 224 × 224 × 3 pixels. The initial learning rate for the Adam optimizer is set to 0.0001. We randomly split the datasets into training and testing sets, considering ratios of 80% and 20%, respectively. Both CDCTNet and baseline models are evaluated in the context of landslide detection and segmentation by using standard metrics commonly employed in semantic segmentation tasks. These include precision, recall, F1-score, intersection over union (IoU), and average inference time (AIT). Each metric assesses different aspects of model performance, such as accuracy in identifying landslide regions, the balance between false positives and negatives, the degree of overlapping with ground truth, and computational efficiency. Precision measures the proportion of correctly predicted positive instances (true positives) out of all instances predicted as positive. This helps to assess over-prediction of landslide areas and is given by:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{Precision}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{T}\mathrm{P}$$\end{document} is true positives and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{F}\mathrm{P}$$\end{document} is false positives. On the other hand, recall measures the proportion of correctly predicted positive instances out of all actual positive instances, which supports model evaluation in detecting all relevant landslide regions, minimizing missed detections:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{Recall}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{F}\mathrm{N}$$\end{document} is false negatives. The harmonic mean of precision and recall is the F1-score, which provides a balanced measure of a model’s accuracy that accounts for both FP and FN. This metric is particularly useful in imbalanced datasets:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{F1-score}=2\times\:\frac{\mathrm{Precision}\times\:\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}$$\end{document} The overlap between the predicted segmentation mask and the ground truth mask for a single class (e.g., landslide) is evaluated using intersection over union (IoU):11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{IoU}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}$$\end{document} Beyond segmentation accuracy metrics, we measure the average time in ms to process a single image during inference using the average inference time (AIT). This metric is important in real-time applications, where low latency is essential for feasibility in cloud-based or edge devices. The above performance metrics are used on the test sets of the Debrisflow, Landslide, and Bijie datasets, ensuring a comprehensive evaluation of both segmentation quality and efficiency. Quantitative and qualitative results analysis For the quantitative results, CDCTNet was trained and tested on the three datasets. We performed comparisons with the baseline models of BiSeNet39, DANet40, SegNet41, DeepLabv342, PSPNet43, Unet44, Transfuse45, SwinUNet++46 and GeoViT47. The quantitative results of DS1, DS2 and DS3 are shown in Tables 1, 2 and 3, respectively. Table 1 shows that the precision, F1-score and IOU of BiSeNet, DeepLabv3 and PSPNet are very close to each other. Moreover, the AIT value of DeepLabv3 is very close to BiSeNet and PSPNet. The SegNet F1-score is 70.47%. Meanwhile, Transfuse obtained the second highest classification performance in Table 2. In addition, Transfuse has the highest recall value of 75.10%. Moreover, the hybrid models SwinUNet + + and GeoViT achieved notable performance on the DS1. SwinUNet + + obtained the second highest precision value of 74.16%. Furthermore, the proposed CDCTNet precision, IOU, and F1-score values are 78.57%, 52.45% and 76.56%, respectively. In addition, the AIT value for the proposed model is 32.75, which is a bit higher than that of DANet at 27.64. Table 2Performance comparison on DS1(bold/italics refer to best/second-best).ModelPrecision (%)↑Recall (%)↑IoU (%)↑F1-score (%)↑AIT↓BiSeNet65.2352.1634.5057.9647.34DANet45.4626.3423.4233.45 27.64 SegNet72.6468.3546.2770.4772.57DeepLabv366.3244.2738.4853.0964.78PSPNet67.8442.1335.2451.1045.32Unet35.6843.2823.6839.12102.87Transfuse73.58 75.10 52.16 74.33 83.12SwinUNet++ 74.16 73.5451.3773.8242.18GeoViT73.2874.1650.0973.7264.80CDCTNet 78.57 74.64 52.45 76.56 32.75 Table 2Performance comparison on DS1(bold/italics refer to best/second-best).ModelPrecision (%)↑Recall (%)↑IoU (%)↑F1-score (%)↑AIT↓BiSeNet65.2352.1634.5057.9647.34DANet45.4626.3423.4233.45 27.64 SegNet72.6468.3546.2770.4772.57DeepLabv366.3244.2738.4853.0964.78PSPNet67.8442.1335.2451.1045.32Unet35.6843.2823.6839.12102.87Transfuse73.58 75.10 52.16 74.33 83.12SwinUNet++ 74.16 73.5451.3773.8242.18GeoViT73.2874.1650.0973.7264.80CDCTNet 78.57 74.64 52.45 76.56 32.75 Performance comparison on DS1(bold/italics refer to best/second-best). Table 3 shows that SegNet has the highest precision and IoU values at 78.37% and 52.62%, respectively, followed by CDCTNet at 75.23% and 52.12%, respectively. However, Segnet’s recall is 67.72%, while CDCTNet demonstrates the highest recall at 69.16%. Transfuse obtained good performances in terms of precision, recall, IoU and F1-score, however, CDCTNet achieved the highest recall and F1-score. In terms of processing speed, DANet demonstrates the fastest AIT of 25.68, while BiSeNet and CDCTNet come next with AIT of 44.54. Moreover, the hybrid models SwinUNet + + and GeoViT obtained precision values of 73.81% and 70.15%, respectively, on the DS2. SwinUNet + + obtained the second highest F1-score value of 70.81%. Table 3Performance comparison on DS2 (bold/italics refer to best/second-best).ModelPrecision (%)↑Recall (%)↑IoU (%)↑F1-score (%)↑AIT↓BiSeNet68.2349.1635.1257.14 44.54 DANet47.1856.3428.2751.35 25.68 SegNet 78.37 67.72 52.62 69.6773.14DeepLabv370.1564.3750.9467.1363.15PSPNet64.1941.2834.8450.2446.37Unet37.5448.6228.5342.36103.24Transfuse74.2567.2351.1670.5784.76SwinUNet++73.8168.0550.67 70.81 55.26GeoViT70.1566.7647.1268.4178.14CDCTNet 75.23 69.16 52.12 72.06 44.54 Table 3Performance comparison on DS2 (bold/italics refer to best/second-best).ModelPrecision (%)↑Recall (%)↑IoU (%)↑F1-score (%)↑AIT↓BiSeNet68.2349.1635.1257.14 44.54 DANet47.1856.3428.2751.35 25.68 SegNet 78.37 67.72 52.62 69.6773.14DeepLabv370.1564.3750.9467.1363.15PSPNet64.1941.2834.8450.2446.37Unet37.5448.6228.5342.36103.24Transfuse74.2567.2351.1670.5784.76SwinUNet++73.8168.0550.67 70.81 55.26GeoViT70.1566.7647.1268.4178.14CDCTNet 75.23 69.16 52.12 72.06 44.54 Performance comparison on DS2 (bold/italics refer to best/second-best). The structure of DS3 is comparatively less complex. Due to this, overall model performance improved by and large compared to DS1 and DS2, shown in Table 4. Table 4 shows that, Unet and DANet demonstrated the lowest performances in terms of classification metrics. CDCTNet leads across all classification metrics achieving precision of 94.14%, recall of 95.56%, IoU of 85.83% and F1-score of 94.85%. Transfuse and SegNet are close contenders with precisions of 87.67% and 85.21%, recalls of 91.52% and 75.34%, IoUs of 78.97% and 67.13%, and F1-scores of 89.55% and 79.93%, respectively. DeepLabv3, BiSeNet and PSPNet produce mid-range performance. When it comes to AIT, DANet maintains its leading performance at 24.52, but fails to meet expectations in terms of classification metrics. This is followed by the proposed CDCTNet with a slightly increased AIT of 26.57, representing a balanced compromise between real-time and classification capabilities. Moreover, SwinUNet + + provided the second highest precision value of 90.50% and GeoViT achieved a precision value of 88.25%. Table 4Performance comparison on DS3 (bold/italics refers to best/second-best).ModelPrecision (%)↑Recall (%)↑IoU (%)↑F1-score (%)↑AIT↓BiSeNet75.4754.1542.8663.0542.36DANet52.3565.8438.5758.33 24.52 SegNet85.2175.3467.1379.9367.84DeepLabv377.3668.2561.2767.8562.24PSPNet72.1446.4237.8556.4943.58Unet42.5654.6828.5347.8498.89Transfuse87.67 91.52 78.97 89.5562.18SwinUNet++ 90.50 89.7467.80 90.19 38.86GeoViT88.2586.3157.6987.2660.23CDCTNet 94.14 95.56 85.83 94.85 26.57 Table 4Performance comparison on DS3 (bold/italics refers to best/second-best).ModelPrecision (%)↑Recall (%)↑IoU (%)↑F1-score (%)↑AIT↓BiSeNet75.4754.1542.8663.0542.36DANet52.3565.8438.5758.33 24.52 SegNet85.2175.3467.1379.9367.84DeepLabv377.3668.2561.2767.8562.24PSPNet72.1446.4237.8556.4943.58Unet42.5654.6828.5347.8498.89Transfuse87.67 91.52 78.97 89.5562.18SwinUNet++ 90.50 89.7467.80 90.19 38.86GeoViT88.2586.3157.6987.2660.23CDCTNet 94.14 95.56 85.83 94.85 26.57 Performance comparison on DS3 (bold/italics refers to best/second-best). Next, we provide a qualitative comparison of the segmentation results of CDCTNet with respect to the baseline models. The visual segmentation maps of the proposed method and BiSeNet, DANet, SegNet, DeepLabv3, PSPNet, Unet and Transfuse on DS1, DS2 and DS3 are shown in Figs. 4, 5 and 6, respectively. In Fig. 4, it is evident that the DANet and Unet cannot maintain the details of the objects. In addition, the resulting segmentations suffer from noise. However, BiSeNet and PSPNet demonstrate reduced noise performance while improving the preservation of structural details. Importantly, DeepLabv3 and Transfuse are able to capture the fine details of the land covers. Both SegNet and CDCTNet can maintain the objects’ coarse details and contain less noise. Furthermore, Fig. 5 shows that BiSeNet, DANet and PSPNet visual segmentation maps suffer from noise and have less detailed information about the objects. At the same time, SegNet and Transfuse demonstrate reductions in noise and provide more detailed information regarding land cover. CDCTNet’s visual map is the closest to GT and has rich information about the objects. We can notice in Fig. 6 that the BiSeNet, DANet, PSPNet and Unet cannot maintain the detailed edge information of the land covers and have a large amount of noise. On the other hand, SegNet, DeepLabv3, and Transfuse contain less noise. Lastly, the proposed model’s segmentation map has the least amount of noise and contains rich structural information. Fig. 4The visual results of DS2, (a) Original image, (b) ground truth image (c) BiSeNet (d) DANet (e) SegNet (f) DeepLabv3 (g) PSPNet (h) Unet (i) Transfuse (j) CDCTNet. The visual results of DS2, (a) Original image, (b) ground truth image (c) BiSeNet (d) DANet (e) SegNet (f) DeepLabv3 (g) PSPNet (h) Unet (i) Transfuse (j) CDCTNet. Fig. 5The visual results of DS1, (a) Original image (b) ground truth image (c) BiSeNet (d) DANet (e) SegNet (f) DeepLabv3 (g) PSPNet (h) Unet (i) Transfuse (j) CDCTNet. The visual results of DS1, (a) Original image (b) ground truth image (c) BiSeNet (d) DANet (e) SegNet (f) DeepLabv3 (g) PSPNet (h) Unet (i) Transfuse (j) CDCTNet. Fig. 6The visual results of DS3, (a) Original image (b) ground truth image (c) BiSeNet (d) DANet (e) SegNet (f) DeepLabv3 (g) PSPNet (h) Unet (i) Transfuse (j) CDCTNet. The visual results of DS3, (a) Original image (b) ground truth image (c) BiSeNet (d) DANet (e) SegNet (f) DeepLabv3 (g) PSPNet (h) Unet (i) Transfuse (j) CDCTNet. Comparison with SOTA methods As discussed in Sect. 2, land cover change detection using ML and DL has been the topic of much recent research. The complex structure of the landslide area and the limited availability of data makes this task even more challenging. ML-based methods are less efficient in real-world applications due to the use of handcrafted features. In contrast, DL-based approaches enhance change detection accuracy, but at the expense of higher computational costs. The comparison of different state-of-the-art methods on the Bijie data is shown in Table 5. Hou et al.48 applied ResNet-50-based Unet and transformer in LCCD. The number of trainable parameters of the ResNet-50 is 23 × 106, and the additional trainable parameters of the transformer block make this method computationally expensive. Yang et al.49 compared the computational efficiency of three DL methods, namely, Unet, DeeplabV3 and PSPNet, in change detection. These methods have high AIT, which makes them challenging to implement in real-time applications. Further, Wang et al.50 modified the Unet architecture and developed the Unet + + model using ViT and block attention to improve landslide segmentation performance. Zhou et al.51 proposed residual block-based encoder and decoder for change detection. Their DL model, DeenNet, extracts multi-scale landslide features to improve accuracy but demonstrates increases in computational costs with an AIT of 84. Liu et al.52 utilized ResNet-101 as an encoder and a CNN-based decoder for landslide segmentation. Their model obtained a precision of 87.60%. Xi et al.53 utilized the segment anything model (SAM) as encoder and a cross-feature fusion decoder and achieved a mIoU value of 85.93%. In another research, Li et al.54 applied dilated convolutions and enhanced multi-scale attention in their Unet implementation. Their model obtained a precision of 80.26%. In CDCTNet, we utilize 2D convolutional layers for spatial feature extraction, while long-range dependencies are provided by the LSA and GSA blocks in the transformer. The LSA provides local attention, and the GSA block maintains global feature attention. This improves the correlation among spatial features leading to notable performance increases. CDCTNet respectively achieves precision, IOU and mIoU values of 94.14%, 85.83% and 92.45% on DS3. Table 5State-of-the-art performance comparison on DS3 (bold/italics refer to best/second-best).MethodPrecision (%)↑ Recall (%)↑IOU (%)↑F1-score (%)↑mIoU (%)↑Hou et al48.93.14 93.69 78.83 93.41 88.10 Yang et al49.87.2488.2378.1587.7387.91Wang et al50. 95.05 90.38 86.31 92.65-Zhou et al51.92.4038.40-54.2423.70Liu et al52.87.6085.70-86.7087.10Xi et al53.93.7792.1877.1392.9685.93Li et al54.80.2688.06-83.9883.86CDCTNet 94.14 95.56 85.83 94.85 92.45 Table 5State-of-the-art performance comparison on DS3 (bold/italics refer to best/second-best).MethodPrecision (%)↑ Recall (%)↑IOU (%)↑F1-score (%)↑mIoU (%)↑Hou et al48.93.14 93.69 78.83 93.41 88.10 Yang et al49.87.2488.2378.1587.7387.91Wang et al50. 95.05 90.38 86.31 92.65-Zhou et al51.92.4038.40-54.2423.70Liu et al52.87.6085.70-86.7087.10Xi et al53.93.7792.1877.1392.9685.93Li et al54.80.2688.06-83.9883.86CDCTNet 94.14 95.56 85.83 94.85 92.45 State-of-the-art performance comparison on DS3 (bold/italics refer to best/second-best). To further support the promising performance of CDCTNet, we provide the validation loss curves for the three datasets utilized in this study in Fig. 7. In Fig. 7(a), related to validation on DS1, the loss starts high and exhibits a gradual downward trend, interspersed with notable peaks and valleys, likely attributable to the dataset’s complex landslide structures and potential class imbalance. This behavior may indicate moderate training instability, possibly from challenging edge cases, yet the overall convergence suggests effective learning stabilization over epochs. Similar patterns emerge in Figs. 7(b)-(c), with initial high losses decreasing progressively. Across all datasets, the model’s ability to reduce loss despite fluctuations demonstrates robust training stability, enhanced by the hybrid CNN-transformer architecture’s handling of long-range dependencies. This supports CDCTNet’s reliability for real-time landslide detection, though further regularization could mitigate peaks in complex scenarios. Fig. 7Validation loss curves for (a) DS1 (b) DS2 and (c) DS3. Validation loss curves for (a) DS1 (b) DS2 and (c) DS3. Ablation study and analyses To validate the contributions of key components in CDCTNet, i.e., LSA, GSA and their integration, the gated attention mechanism, we conducted an ablation study on the Bijie landslide dataset. These experiments systematically isolate and evaluate individual elements to quantify their impact on performance metrics. Moreover, we use class activation maps (CAM) to support the explainability of the proposed system. We also provide further statistical analyses, including paired t-tests, hyperparameter selection and precision-recall curves, to assess the significance of improvements and ensure robustness across datasets. This section also explores model complexity, failure cases, and training dynamics, providing deeper insights into CDCTNet’s efficiency and explainability for landslide segmentation tasks. Model component ablations and statistical validation In Table 6, we can notice that CNN as encoder and decoder with attention gate has a lower IoU value of 82.73%. Further, with the inclusion of classical ViT (MHSA) with CNN as encoder, the IoU value improves to 83.21%. At the same time, CNN + ViT (MHSA + GSA) obtained a 84.76% IoU value, higher than CNN + ViT (LSA). The highest IoU value is obtained when using CNN + ViT (LSA + GSA) in the encoder and CNN + Attention gate in the decoder of the proposed model.Table 6Effect of different components on performance for the Bijie landslide dataset.DatasetEncoderDecoderIoU (%)CNNCNN + Attention Gate82.73CNN + ViT (MHSA)CNN83.21Bijie landslideCNN + ViT (MHSA)CNN + Attention Gate83.69CNN + ViT (MHSA + GSA)CNN + Attention Gate84.76CNN + ViT (LSA)CNN + Attention Gate84.02CNN + ViT (LSA + GSA)CNN + Attention Gate85.83 Table 6Effect of different components on performance for the Bijie landslide dataset.DatasetEncoderDecoderIoU (%)CNNCNN + Attention Gate82.73CNN + ViT (MHSA)CNN83.21Bijie landslideCNN + ViT (MHSA)CNN + Attention Gate83.69CNN + ViT (MHSA + GSA)CNN + Attention Gate84.76CNN + ViT (LSA)CNN + Attention Gate84.02CNN + ViT (LSA + GSA)CNN + Attention Gate85.83 Effect of different components on performance for the Bijie landslide dataset. To statistically validate the performance improvement introduced by the LSA and GSA blocks over the classical Vision Transformer (ViT) with Multi-Head Self-Attention (MHSA), we conducted a paired t-test across the three benchmark datasets, presented in Fig. 8. This test assesses whether the improvements observed in evaluation metrics are statistically significant rather than due to random chance, thereby strengthening the reliability and generalizability of the proposed model’s superiority. The results are shown in Fig. 8. In DS1, we obtained a mean difference of 4.57 and a t-statistic of 4.04, the p-value is 0.027 < 0.05, which indicates that the null hypothesis can be rejected. Furthermore, the mean difference and t-statistic values are 1.55 and 9.11, respectively, on DS2. Moreover, the p-value is 0.0028 < 0.05. Furthermore, on DS3, the mean difference, t-statistic, p-values are 2.14, 9.09, and 0.0023 < 0.05, respectively, which indicates the null hypothesis can be rejected, supporting the hypothesis that the model outperforms on all three datasets. Fig. 8Paired t-tests on (a) DS1 (b) DS2 and (c) DS3. Paired t-tests on (a) DS1 (b) DS2 and (c) DS3. Explainability and model interpretations An interesting feature of CDCTNet is its enhanced explainability, achieved through the integration of LSA and GSA blocks, which allow for interpretable visualizations of the areas where the model focuses in its decision making. In this context, we employed class activation maps (CAM), which highlight regions of the input image that most influence the model’s predictions. As shown in Fig. 9, CAM before and after applying the ViT (LSA + GSA) module demonstrate a marked improvement in capturing landslide boundaries and distinguishing foreground (landslide areas) from background. For instance, on DS1 and DS2, the post-ViT maps exhibit precise localization of complex structures, with heightened activation on edges, addressing limitations in boundary exploration, typical in traditional CNNs. This interpretability not only validates the model’s attention to long-range dependencies but also supports real-time analysis by clarifying the model’s decision-making process. Fig. 9Class activation maps of the model on DS1: (a) Original image, (b) Input features to ViT, (c) Feature output from ViT, (d) Ground truth. On DS2: (e) Original image, (f) Input features to ViT, (g) Feature output from ViT, (h) Ground truth. On DS3: (i) Original image, (j) Input features to ViT, (k) Feature output from ViT, (l) Ground truth. Class activation maps of the model on DS1: (a) Original image, (b) Input features to ViT, (c) Feature output from ViT, (d) Ground truth. On DS2: (e) Original image, (f) Input features to ViT, (g) Feature output from ViT, (h) Ground truth. On DS3: (i) Original image, (j) Input features to ViT, (k) Feature output from ViT, (l) Ground truth. However, to assess robustness, we analyzed failure cases, as illustrated in Fig. 10. These reveal challenges where textures and shapes in non-landslide regions mimic landslide features, leading to misclassifications. Additionally, very small landslide areas pose detection difficulties due to insufficient contextual cues, even with augmented attention. These insights underscore areas for refinement, such as incorporating multi-scale enhancements, while reinforcing CDCTNet’s overall reliability in diverse scenarios. Fig. 10Sample images of the Bijie-landslide dataset. Sample images of the Bijie-landslide dataset. Hyperparameter sensitivity and performance profiles We performed an ablation study using the initial LR (learning rate) and batch size on DS1, DS2 and DS3, shown in Fig. 11. We observe in Fig. 11(a) that with an LR of 1e-4, the precision values on the datasets are lower. This increases with a LR of 2e-4 and reaches maximum precision values using a LR 3e-4. When attempting to further increase the value of the LR, the model showed degradation in precision. Figure 11 (b) shows that the model precision value reduces when using a batch size of 4. Some improvement is observed for a batch size of 8. Moreover, the highest value is obtained with a batch size of 16. Larger batch sizes (e.g., 32) slightly diminish performance, potentially due to less stable gradient updates with increased batch normalization challenges for smaller datasets. Fig. 11Ablation study on (a) Initial learning rate and (b) Batch size. Ablation study on (a) Initial learning rate and (b) Batch size. We performed PR (precision recall) curve-based analysis of the BiSeNet, DANet, SegNet, DeepLabv3, PSPNet, Unet, Transfuse and CDCTNet on the DS1, DS2 and DS3, shown in Fig. 12. Figure 12(a) shows that Unet has the least PR-AUC (area under the curve) on DS1. At the same time, DANet and PSPNet obtained a PR-AUC values of 0.3586 and 0.5830, respectively. The PR-AUC values of the SegNet and Transfuse are very close. Moreover, the proposed CDCTNet has a PR-AUC value of 0.768. Figure 12(b) shows reduced PR-AUC values for DANet and Unet on DS2. The PR-AUC values for DeepLabV3 and PSPNet are 0.6526 and 0.5217, respectively. Meanwhile, the PR-AUC value of the proposed CDCTNet is 0.7268. Figure 12 (c) shows that the PR-AUC values of the DANet and PSPNet are 0.5449 and 5708, on DS3, respectively. At the same time, DeepLabV3 and SegNet have PR-AUC values less than 0.75. Moreover, the Transfuse obtained the second highest value, and the proposed CDCTNet achieved a PR-AUC value of 0.939 on DS3. These analyses collectively underscore the model’s robustness to hyperparameter variation and its high effectiveness across datasets, affirming CDCTNet’s practicality for operational landslide detection scenarios. Fig. 12Illustration of PR-curve analysis (a) DS1 (b) DS2 and (c) DS3. Illustration of PR-curve analysis (a) DS1 (b) DS2 and (c) DS3. Computational efficiency and complexity CDCTNet’s design balances high segmentation accuracy with reduced computational overhead, making it suitable for real-time applications. This is achieved by leveraging a lightweight convolutional encoder with depthwise convolutions and efficient transformer modules incorporating LSA and GSA blocks. We performed complexity analysis using the parameters in millions, flops in Giga floating point operations per second of each model and precision values on DS1, DS2 and DS3. Table 7 shows that Transfuse has 47.8 M parameters and achieved a precision value of 73.58% on DS1. At the same time, DANet has the highest number of trainable parameters, and the precision value is 52.35% on the DS3. SegNet has 286.4 flops and the highest precision value on the DS2. DeepLabV3 contains 46.5 M trainable parameters, with a precision value of 77.36% on the DS3. The Unet has the second lowest number of trainable parameters, and the precision value is the lowest of the three datasets. Moreover, the proposed CDCTNet has 16.86 M parameters and obtained the highest precision values on the DS1 and DS3. The transformer based model Swin-UNet + + and GeoViT have relative large numbers of parameters and flops. Table 7Complexity analysis of the models.ModelParameters (M)↓Flops (G)↓Precision (DS1)↑Precision (DS2)↑ Precision (DS3)↑BiSeNet49.00 55.30 65.23%68.23%75.47%DANet66.00165.0045.46%47.18%52.35%SegNet29.40286.4072.64%78.37%85.21%DeepLabV342.5085.5066.32%70.15%77.36%PSPNet46.50123.0067.84%64.19%72.14%Unet 24.50 181.4035.68%37.54%42.56%Transfuse47.8088.60 73.58% 74.25%87.67%SwinUNet++52.6095.0474.1673.81 90.50 GeoViT83.50173.2573.2870.1588.25CDCTNet 16.86 84.4 78.57% 75.23% 94.14% Table 7Complexity analysis of the models.ModelParameters (M)↓Flops (G)↓Precision (DS1)↑Precision (DS2)↑ Precision (DS3)↑BiSeNet49.00 55.30 65.23%68.23%75.47%DANet66.00165.0045.46%47.18%52.35%SegNet29.40286.4072.64%78.37%85.21%DeepLabV342.5085.5066.32%70.15%77.36%PSPNet46.50123.0067.84%64.19%72.14%Unet 24.50 181.4035.68%37.54%42.56%Transfuse47.8088.60 73.58% 74.25%87.67%SwinUNet++52.6095.0474.1673.81 90.50 GeoViT83.50173.2573.2870.1588.25CDCTNet 16.86 84.4 78.57% 75.23% 94.14% Complexity analysis of the models. Performance evaluation on UAV dataset The model was also evaluated on the UAV-based image dataset, obtained from literature55. The dataset contains 5925 high-resolution images of Jiuzhai valley, having a resolution of 512 × 512, stored in TIFF format with a spatial resolution of 0.2 m per pixel. In addition, the label and mask of each image is provided to test the model performance for landslide detection. We kept the same experimental conditions, as discussed in Sect. 4.2. After training, the performance of the CDCTNet was evaluated and is depicted in Table 8. Table 8 shows that CDCTNet obtained a precision of 97.18% and an IOU value of 90.08%. In addition, it has an F1-score of 96.72% on the UAV dataset. Table 8Performance evaluation on the UAV based landslide dataset.MethodPrecision (%)↑Recall (%)↑IOU (%)↑F1-score (%)↑mIoU (%)↑CDCTNet97.1896.2790.0896.7294.80 Table 8Performance evaluation on the UAV based landslide dataset.MethodPrecision (%)↑Recall (%)↑IOU (%)↑F1-score (%)↑mIoU (%)↑CDCTNet97.1896.2790.0896.7294.80 Performance evaluation on the UAV based landslide dataset. Further visual results of some samples are depicted in Fig. 13. In Fig. 13, the top row presents the original image, the middle row presents the ground truth, and the bottom row presents the predicted images. It can be observed that the CDCTNet is capable of retaining most of the information in the predicted images. However, some noise can be seen in the complex landslide patterns. Fig. 13Visual results on the UAV dataset. Visual results on the UAV dataset. Conclusion Landslide incidents are rapidly growing globally due to heavy rains, earthquakes, volcanic eruptions, and droughts. Large area landslide is a challenging and complex problem. In the past, several computer vision-based methods have been developed to report landslide events. These methods’ performance is acceptable. However, real-time applications require highly efficient architectures and lower computational costs. Most methods utilized residual networks as a base to design the encoder and decoder. These methods have many trainable parameters, and shallow features are extracted from the landslide images. We utilized four depth-wise convolution blocks in the encoder and decoder to overcome high computational costs. In addition, to maintain the long-range dependency of spatial features, we included LSA and GSA blocks in the encoder and a gated convolution block in the decoder. By doing this, enhanced features are utilized, and LCCD performance is improved. The CDCTNet performance is evaluated on three datasets, and promising results are obtained. Furthermore, a paired t-test is performed on the three datasets utilized in the study. The p-value of the model is less than 0.05 on all the datasets, which confirms the model’s consistent performance. The CDCTNet model obtained precision, F1-score and IoU values of 78.57%, 76.56% and 52.46%, respectively, on DS1. Meanwhile, on DS2 and DS3, IoU values are 72.06% and 94.85%, respectively. Due to the coarse sub-window partitioning used in the local self-attention module, the model struggles to detect small-scale landslides efficiently. In addition, some misclassifications occur in non-landslide regions that share similar textures to landslide areas. Further, the global self-attention block introduces considerable computational overhead, constraining scalability for high-resolution images. Fourth, the framework does not currently incorporate multi-modal information, such as topographic or radar data, which could enhance robustness in complex terrains. In future studies, multi-scale attention with varied window sizes may improve the detection of small features. Adversarial training with generative adversarial networks can help distinguish subtle texture differences. Further, integrating multi-modal inputs through a parallel encoder branch can strengthen performance in challenging conditions. Additionally, nature-inspired optimization techniques may be applied to prune parameters while maintaining accuracy.