Internal Solitary Waves (ISWs) play a pivotal role in transporting energy and matter within the ocean and also pose substantial risks to ocean engineering, navigation, and underwater communication systems. Consequently, measures need to be adopted to alleviate their negative effects and minimize linked risks. An effective method entails extracting ISW positions from Synthetic Aperture Radar (SAR) data for precise trajectory prediction and efficient avoidance strategies. However, manual extraction of ISWs from SAR data is time-consuming and prone to inaccuracies. Hence, it is imperative to develop a high-precision, rapid, and automated ISW-extraction algorithm. In this paper, we introduce Middle Transformer U²-net (MTU²-net), an innovative model that integrates a distinctive loss function and Transformer to improve the accuracy of ISWs’ extraction. The novel loss function enhances the model’s capacity to extract bow waves, whereas the Transformer ensures coherence in ISW’s patterns. By conducting experiments involving 762 image scenes, incorporating ISWs, from the South China Sea, we established a standardized dataset. The Mean Intersection over Union (MIoU) achieved on this dataset was 71.57%, surpassing the performance of other compared methods. The experimental outcomes showcase the remarkable performance of our proposed model in precisely extracting bow wave attributes from SAR data.

Internal waves (IWs) play a crucial role in the transport of energy and matter within the ocean while also posing significant risks to marine engineering, navigation, and underwater communication systems. Consequently, effective segmentation methods are essential for mitigating their adverse impacts and minimizing associated hazards. A promising strategy involves applying remote sensing image segmentation techniques to accurately identify IWs, thereby enabling predictions of their propagation velocity and direction. However, current IWs segmentation models struggle to balance computational efficiency and segmentation accuracy, often resulting in either excessive computational costs or inadequate performance. Motivated by recent developments in the Mamba2 architecture, this paper introduces the state-space model meets linear attention (SMLA), a novel segmentation framework specifically designed for IWs. The proposed hybrid architecture effectively integrates three key components: a feature-aware serialization (FAS) block to efficiently convert spatial features into sequences; a state-space model with linear attention (SSM-LA) block that synergizes a state-space model with linear attention for comprehensive feature extraction; and a decoder driven by hierarchical fusion and upsampling, which performs channel alignment and scale unification across multi-level features to ensure high-fidelity spatial detail recovery. Experiments conducted on a dataset of 484 synthetic-aperture radar (SAR) images containing IWs from the South China Sea achieved a mean Intersection over Union (MIoU) of 74.3%, surpassing competing methods evaluated on the same dataset. These results demonstrate the superior effectiveness of SMLA in extracting features of IWs from SAR imagery.

Internal Solitary Waves (ISWs) are a significant oceanic phenomenon, particularly in the South China Sea. The propagation of ISWs influences the transport of matter, energy, and momentum within the ocean. They also pose a considerable risk to offshore engineering and submarine safety. To leverage the advantages of ISWs and to mitigate the associated risks to offshore activities, it is valuable to understand their continuous propagation process over a period of time. At present, there is a lack of tools capable of illustrating the continuous propagation of ISWs. Consequently, reconstructing the continuous propagation process of ISWs based on sparse remote sensing imagery is a valuable endeavor. To address this issue, this paper defines the completion of the ISW propagation process as an end-to-end visual generation task and proposes a novel Collaborative Diffusion Framework that directly generates high-fidelity visual sequences of ISW propagation. The core innovation of the framework is the disentanglement of static appearance and dynamic motion, together with the integration of physical information into both representations. It employs a dual-path architecture where an appearance model learns the static morphology of ISWs from a single image, while a motion model learns their propagation kinematics. These two components are synergistically constrained by a shared-weights mechanism and physical information, ensuring a physically coherent spatio-temporal representation. The proposed framework is validated using numerical simulation data, Synthetic Aperture Radar (SAR) imagery, and Geostationary High-speed Imager (GHI) imagery. The framework is assessed primarily by visual metrics, supplemented by quantitative metrics where continuous sequences are available. Quantitative results on both GHI and numerical simulations demonstrate the high fidelity of the proposed method. For the GHI data, the errors in propagation velocity remain within 0.7 m/s, while the predicted propagation direction is highly consistent with the observations. For the numerical simulations, the velocity error is only 0.1 m/s, with a direction deviation of 11◦. The results confirm that this framework provides a novel and effective tool for reconstructing the propagation process of ISWs over time from otherwise sparse observations.

The rich spectral information in hyperspectral images (HSIs) provides a significant advantage for fine-grained, multiclass change detection (CD), yet a robust and practical framework remains elusive. Existing deep learning methods are often fragmented, failing to simultaneously address a triad of critical challenges: i) spectral inconsistency between HSI pairs acquired from different sensors or times; ii) inefficient static fusion of hierarchical spectral-spatial features, which struggles to model complex changes; and iii) severe label scarcity for multiclass supervision, which limits model training. To systematically solve these interconnected problems, we propose the Deep Differential Feature Fusion (D²F²) framework, a holistic, end-to-end architecture. D²F² introduces two novel components: a Spectral Adaptive Sampling (SAS) module that adaptively remaps inputs with arbitrary spectral bands to a unified feature space, thereby alleviating spectral inconsistency caused by missing spectral channels and band-number mismatch. Concurrently, a Spatial Hierarchical Aggregation (SHA) module replaces static fusion with a dynamic, attention-based mechanism to generate highly discriminative, change-aware features. To address label scarcity, the entire framework is trained using an efficient semi-supervised strategy. This fully convolutional design also enables ultra-efficient map-to-map inference, which eliminates the massive computational redundancy of patch-based methods. Extensive experiments on three challenging HSI datasets demonstrate that D²F² not only achieves new state-of-the-art accuracy but is also orders of magnitude more computationally efficient than competing deep learning models. The code will be available at: https://github.com/dmhz/HSCD.