Landslide susceptibility mapping (LSM) is of great significance for the identification and prevention of geological hazards. LSM is based on convolutional neural networks (CNNs); CNNs use fixed convolutional kernels, focus more on local information and do not retain spatial information. This is a property of the CNN itself, resulting in low accuracy of LSM. Based on the above problems, we use Vision Transformer (ViT) and its derivative model Swin Transformer (Swin) to conduct LSM for the selected study area. Machine learning and a CNN model are used for comparison. Fourier transform amplitude, feature similarity and other indicators were used to compare and analyze the difference in the results. The results show that the Swin model has the best accuracy, F1-score and AUC. The results of LSM are combined with landslide points, faults and other data analysis; the ViT model results are the most consistent with the actual situation, showing the strongest generalization ability. In this paper, we believe that the advantages of ViT and its derived models in global feature extraction ensure that ViT is more accurate than CNN and machine learning in predicting landslide probability in the study area.

In semantic segmentation tasks, the transition from convolutional neural networks (CNN) to transformers is driven by the latter's superior ability to capture global semantic information in remote sensing images. However, most transformer methods face challenges such as slow inference speed and limitations in capturing local features. To address these issues, this study designs a hybrid approach that integrates knowledge distillation with a combination of CNN and transformer to enhance semantic segmentation in remote sensing images. First, this paper proposes the Dual-Path Convolutional Transformer Network (DP-CTNet) with a dual-path structure to leverage the strengths of both CNN and transformers. It incorporates Feature Refinement Module to optimize the transformer's feature learning, and Feature Fusion Module to effectively merge CNN and transformer features, preventing the insufficiently learning of local features by the transformer. Then DP-CTNet serves as the teacher model, and pruning and knowledge distillation are employed to create Efficient DP-CTNet (EDP-CTNet) with superior segmentation speed and accuracy. Angle Knowledge Distillation (AKD) is proposed to enhance the feature migration learning of DP-CTNet during knowledge distillation, leading to improved EDP-CTNet performance. Experimental results demonstrate that DP-CTNet thoroughly combines the respective advantages of CNN and Transformer, maintaining local detail features while learning extensive sequential semantic information. EDP-CTNet not only delivers impressive segmentation speed but also exhibits excellent segmentation accuracy following AKD training. In comparison to other models, the two models proposed in this paper notably distinguish themselves in terms of accuracy and result visualization.

Landslide susceptibility mapping (LSM) is an important decision basis for regional landslide hazard risk management, territorial spatial planning and landslide decision making. The current convolutional neural network (CNN)-based landslide susceptibility mapping models do not adequately take into account the spatial nature of texture features, and vision transformer (ViT)-based LSM models have high requirements for the amount of training data. In this study, we overcome the shortcomings of CNN and ViT by fusing these two deep learning models (bottleneck transformer network (BoTNet) and convolutional vision transformer network (ConViT)), and the fused model was used to predict the probability of landslide occurrence. First, we integrated historical landslide data and landslide evaluation factors and analysed whether there was covariance in the landslide evaluation factors. Then, the testing accuracy and generalisation ability of the CNN, ViT, BoTNet and ConViT models were compared and analysed. Finally, four landslide susceptibility mapping models were used to predict the probability of landslide occurrence in Pingwu County, Sichuan Province, China. Among them, BoTNet and ConViT had the highest accuracy, both at 87.78%, an improvement of 1.11% compared to a single model, while ConViT had the highest F1-socre at 87.64%, an improvement of 1.28% compared to a single model. The results indicate that the fusion model of CNN and ViT has better LSM performance than the single model. Meanwhile, the evaluation results of this study can be used as one of the basic tools for landslide hazard risk quantification and disaster prevention in Pingwu County.

The extraction of disaster geospatial intelligence (DGI) from social media data with spatiotemporal attributes plays a crucial role in real-time disaster monitoring and emergency decision-making. However, conventional machine learning approaches struggle with semantic complexity and limited Chinese disaster corpus. Recent advancements in large language models (LLMs) offer new opportunities to overcome these challenges due to their enhanced semantic comprehension and multi-task learning capabilities. This study investigates the potential application of LLMs in disaster intelligence extraction and proposes an efficient, scalable method for multi-hazard DGI extraction. Building upon a unified ontological framework encompassing core natural disaster elements, this method employs parameter-efficient low-rank adaptation (LoRA) fine-tuning to optimize open-source Chinese LLMs using a meticulously curated instruction-tuning dataset. It achieves simultaneous identification of multi-hazard intelligence cues and extraction of disaster spatial entity attributes from unstructured Chinese social media texts through unified semantic parsing and structured knowledge mapping. Compared to pre-trained models such as BERT and ERNIE, the proposed method was shown to achieve state-of-the-art evaluation results, with the highest recognition accuracy (F1-score: 0.9714) and the best performance in structured information generation (BLEU-4 score: 92.9649). Furthermore, we developed and released DGI-Corpus, a Chinese instruction-tuning dataset covering various disaster types, to support the research and application of LLMs in this field. Lastly, the proposed method was applied to analyze the spatiotemporal evolution patterns of the Zhengzhou “7.20” flood disaster. This study enhances the efficiency of natural disaster monitoring and emergency management, offering technical support for disaster response and mitigation decision-making.

The segmentation of crevasses in remote sensing images plays a pivotal role in diverse domains, including crevasse change monitoring, analysis of ice shelf surface water systems, and investigations into ice shelf stability. In response to the limitations in existing crevasses segmentation methods, which struggle to concurrently capture global structures while preserving local details, this letter introduces CREVNet. CREVNet is designed to achieve precise crevasse segmentation, comprising two integral components: the Transformer Path for enhanced local and global feature extraction, and the Convolutional Path for detailed depiction of crevasses. Evaluation on crevasses dataset, created through the integration of optical remote sensing imagery and laser altimetry data, reveals impressive results. CREVNet achieves F1-score, MIoU, and OA values of 80.40%, 80.98%, and 95.24%, respectively. Notably, CREVNet surpasses the performance of prominent deep learning methods, including Unet, DeepLabV3Plus, DFANet, FPN, MobileViT, and TransUnet. These outcomes underscore CREVNet’s practical potential for effective crevasses segmentation.

Very high-resolution remote sensing images is a key data source for space–air–ground collaborative monitoring, supplying fine-grained spatial information crucial to urban management, disaster response and ecological assessment. Reliable extraction of roads, buildings and vegetation therefore hinges on accurate semantic segmentation. Recently, the Mamba state-space model has been widely adopted for remotesensing semantic segmentation, markedly improving the globalcontext modelling capacity of existing approaches; nevertheless, the severe inter-class pixel imbalance among categories persists, continuing to hamper segmentation performance. To mitigate this issue, we present FPMamba, a hybrid network that couples frequency-domain cues with prompt learning to emphasize rare classes and suppress background clutter. On the frequency side, a Frequency Channel–Spatial Attention Module (FCSAM) is inserted into each VSS block to fuse spatial features with a frequency branch that captures high-frequency details, markedly sharpening edge and texture representations. A Frequency Fusion Attention Module (FFAM) is further deployed on every skip connection to adaptively reconcile highlevel semantics with low-level details. On the prompt side, we design a lightweight Prompt Text-Guided Strategy (PTGS) that aligns category prompts with image patches via Contrastive Language–Image Pre-training (CLIP), producing CLIP-guided attention mapsthat guide the network towards long-tail targets. PTGS and FCSAM are jointly embedded into a dual-enhanced VSS block, completing the FPMamba architecture. Comprehensive tests on the ISPRS benchmarks demonstrate that FPMamba reaches an overall accuracy of 88.97% and a MIoU of 75.61% on the Vaihingen dataset, while achieving 87.53% overall accuracy and 75.41% MIoU on the Potsdam dataset, outperforming current state-of-the-art methods. These results corroborate the effectiveness and transferability of integrating frequency-domain information with prompt learning for remote-sensing semantic segmentation. Code will be publicly released upon publication.