Recent advances in vision foundation models (VFMs) have demonstrated strong potential for diverse computer vision tasks; however, their practical applicability to high-resolution satellite imagery under operational constraints remains insufficiently explored. In particular, it is unclear whether a single parameter-efficient fine-tuned VFM can effectively support multiple remote sensing tasks while maintaining competitive performance relative to task-specific models. We evaluate the applicability of a parameter-efficient VFM for multitask analysis of KOMPSAT-3/3A high-resolution satellite imagery, focusing on image captioning, object detection, and referring-expression-based segmentation. To adapt the foundation model to the KOMPSAT domain with limited training data and computational resources, low-rank adaptation was employed. The fine-tuned model was systematically evaluated across the three tasks using consistent preprocessing and evaluation protocols, and its performance was compared with task-specific reference models, including a strong task-specific object detector (YOLOv11). Experimental results indicate that the parameter-efficient fine-tuned VFM substantially improves image captioning quality and achieves competitive performance in single-class object detection when patch-based inference is applied. By contrast, although traditional task-specific models remain preferable for high-throughput, simultaneous multiclass deployment, the parameter-efficient VFM offers distinctive operational utility as a flexible, prompt-driven framework capable of adaptive, sequential target querying without requiring task-specific retraining for diverse referring expressions. For referring-expression-based segmentation, the model demonstrates promising spatial grounding capabilities, although pixel-perfect delineation remains bounded by architectural constraints such as discrete location tokens and simplified polygon vertices. These findings suggest that parameter-efficient VFMs can serve as a practical and versatile solution for multitask satellite image analysis, particularly in applications requiring flexible semantic understanding and limited annotation efforts. Overall, we provide practical guidance for selecting and deploying foundation-model-based pipelines in real-world Earth observation applications under resource and annotation constraints.
Cross-scene classification of hyperspectral images (HSIs) is one of the most important tasks in the field of remote sensing image processing. Although previous work on HSI classification has achieved impressive results, it still faces challenges, such as spectral shifts and semantic discrepancies between domains, which severely degrade classification performance. To address the above problems, we propose a prior-guided adaptive attention and alignment network for cross-scene HSI classification, which ultimately enhances the model's cross-scene classification performance and stability. Specifically, the core design of this network consists of three parts. The first is a prior-guided spatial-spectral attention module, which integrates self-attention, spatial-spectral dual prior guidance, and a spatial-spectral dual attention mechanism. It enhances the ability to distinguish local details and enables the extracted features to cover diverse dimensions, making them more discriminative. The second one is the residual cross-domain feature interaction module, which adaptively adjusts the fusion weight of the source domain and the target domain features via a gated mechanism, effectively reducing the migration of invalid information. Finally, the multi-objective joint loss function combines classification loss, feature diversity loss, and domain alignment loss. Although ensuring the classification accuracy of the source domain, this function also improves the model's adaptability to the target domain. We evaluate our proposed model on three well-known public datasets, and the experimental results demonstrate the effectiveness.
We propose a parallel tile-based architecture for accelerating path-following phase unwrapping algorithms used in digital elevation model (DEM) generation from synthetic aperture radar (SAR) images via interferometric processing (InSAR). In the proposed approach, the input interferogram is partitioned into a configuration-dependent number of tiles, each of which is processed concurrently using a selected phase unwrapping algorithm. The resulting DEM tiles are then combined through a dedicated tile merging module, where continuity across tile boundaries is enforced by minimizing discontinuities and ensuring smooth transitions. The performance of the proposed framework is evaluated using both synthetic and real SAR datasets, considering execution time and generated DEM accuracy in terms of root mean square error. In addition, the obtained results are compared with existing parallelization and tile-merging approaches in the literature. Experimental findings demonstrate that the proposed architecture significantly improves computational efficiency while preserving the structural integrity of the reconstructed DEM, thereby enhancing the practical applicability of path-following phase unwrapping methods for large-scale InSAR data processing.
Along with the advancement of cross-domain contrastive learning techniques, few-shot learning frameworks rooted in prototype networks have been extensively adopted in hyperspectral image classification tasks. However, existing prototype network-based methods often deviate from the true distribution of prototypes when noise and outliers are present, which makes classification less accurate. To address this issue, we introduce a category prototype calibration-based cross-domain few-shot learning method (CPC-CFSL) tailored for hyperspectral image classification. The proposed method reduces the impact of noise on prototypes through a prototype calibration network, employs a class-covariance-based distance metric to capture inter-band correlations, and utilizes a memory bank to mitigate domain shift, thereby improving classification accuracy. Experiment results on four public datasets demonstrate that CPC-CFSL exhibits strong competitiveness compared with other state-of-the-art approaches.
Remote sensing image semantic segmentation remains challenging due to complex spatial distributions, inter-class spectral confusion, and extreme scale variations. Although convolutional neural networks capture local features effectively, they exhibit limitations in modeling the global context. Existing methods often underutilize frequency domain information, with simple fusion strategies failing to exploit spatial-frequency complementarity. We propose the adaptive spatial-frequency feature alignment network with multiscale pyramid fusion (ASFNet). Built on a ConvNeXt-Small backbone, ASFNet employs adaptive spatial-frequency feature alignment to adaptively align spatial and frequency domain features through content-aware gating and cross-modal attention, whereas multiscale feature pyramid fusion addresses scale variations through multirate dilated convolutions and dual attention. Evaluations demonstrate robust performance: 85.71% mIoU on ISPRS Vaihingen, 88.21% on ISPRS Potsdam, and 56.8% on LoveDA. The proposed method shows strong potential for practical deployment in large-scale land cover mapping and urban planning applications.
Aircraft-induced condensation trails are an important anthropogenic climate-forcing agent, yet their automated detection and time-series monitoring under complex cloud backgrounds remain challenging. The 0.65 mu m channel of the FY-4B Advanced Geostationary Radiation Imager provides high-temporal-resolution imagery at 500 m spatial resolution, but its single-band visible configuration constrains traditional multispectral detection schemes. To address this gap, we propose LACNeXt, a line-structure-aware segmentation network built on a ConvNeXt-V2 backbone. LACNeXt integrates a direction-aware depthwise convolution module to strengthen the representation of oriented linear features and a cross-scale cross-attention module to enhance multiscale context modeling. A feature pyramid network-based dual-head decoder with skeleton supervision further improves segmentation continuity and centerline stability for slender contrails. We construct a large-scale multiscene FY-4B contrail dataset spanning diverse underlying surfaces and cloud regimes and conduct tracking-ready time-series case studies based on frame-wise segmentation outputs to qualitatively analyze contrail evolution. Results show that LACNeXt outperforms representative baselines in overall segmentation accuracy and produces temporally stable delineations that support contrail monitoring under varying meteorological conditions. These capabilities make LACNeXt a reliable tool for contrail monitoring from geostationary satellite observations and provide a solid data basis for subsequent analyses of their climate impacts.
Balancing accuracy and efficiency continues to pose a key challenge in semantic segmentation of high-resolution remote sensing images. Existing models frequently face limitations due to either insufficient feature extraction or excessive computational cost. To address this limitation, we propose TransUMamba, an efficient lightweight network with a U-shaped architecture. The encoder is built upon ResT for rapid multiscale feature extraction with minimal parameters. The decoder is equipped with a Spatio-Frequency-Enhanced Mamba module, designed to capture both frequency-domain and spatial-domain information. A dual-branch wavelet fusion upsampling module enhances detail recovery. In addition, an auxiliary branch with a multiscale cross-level feature aggregation module integrates multilevel features to aid training. Comprehensive experimental results on the public datasets released by the International Society for Photogrammetry and Remote Sensing, including the Vaihingen and Potsdam datasets, demonstrate that TransUMamba achieves 91.28% mIoU and 95.38% mF1 on Vaihingen, and 82.77% mIoU and 90.30% mF1 on Potsdam, with only 13.2 M parameters, remarkably achieving an optimal balance between performance and efficiency.
Existing change detection methods for high-resolution remote sensing imagery still face challenges in accuracy and efficiency when addressing pseudo-changes caused by factors such as illumination and seasons, as well as multiscale targets. Moreover, mainstream deep learning approaches predominantly adopt a Siamese network architecture, which isolates bi-temporal information during the feature extraction stage, overlooking the necessity of early interaction and thus limiting detection accuracy. To address this, we introduce a "co-extraction, co-interaction" strategy and propose a hierarchical interaction-fusion network (HIFNet). HIFNet incorporates cross-temporal interaction mechanisms at multiple hierarchical levels, enabling effective early fusion of bi-temporal features and fundamentally overcoming the isolation inherent in traditional Siamese architectures. In addition, an improved temporal complementary fusion (ITCF) module deeply integrates the interacted features through a hybrid similarity metric and a gated fusion strategy, thereby effectively enhancing boundary integrity and internal consistency within change regions. Experiments on public datasets, including learning, vision and remote sensing laboratory change detection, Wuhan University change detection, Lebedev, and Sun Yat-sen University, demonstrate that the proposed method outperforms existing mainstream methods in both accuracy and robustness across diverse scenarios.
Flooding of urban areas causes great risk to lives and property. It may be detected using high-resolution synthetic aperture radar (SAR) sensors by measuring changes in double scattering and/or interferometric coherence between pre- and postflood images. Double scattering in a postflood image usually increases compared with a preflood image, whereas coherence generally decreases. Less attention has been paid to change detection techniques that use a high-resolution digital surface model (DSM) in addition to double scattering and coherence. The availability of a DSM may enable increased spatial resolution in the flood extent delineation and allow flood depth (useful for flood damage assessment) to be measured as well as extent. We aim to investigate how best to combine double scattering, coherence, and DSM data for urban flood mapping using change detection. Four urban floods from Asia and Europe were studied. Due to the limited training data available in these floods, a shallow learning approach was adopted, using separate training and test datasets. It was found that when classifying using double scattering and coherence only, an approach using a neural network (NN) was marginally superior to an established method. The DSM was applied to the resulting NN classification in a second-stage process. The results indicated that using the NN employing only double scattering and coherence, a weighted average classification accuracy of 84% could be achieved on test data. If the DSM was also included in the classification, the accuracy increased to 91%, spatial resolution improved, and flood depth maps could generally be generated. These findings should be of use in automating the detection of urban flooding as an aid to operational flood incident management and flood forecasting.
We present a data-efficient pipeline for semantic segmentation of cereal crops from Sentinel-2 imagery in Algeria. Leveraging two pretrained Vision Foundation Models (Prithvi-EO-2.0 and Satlas), we fine-tuned both under a few-shot regime (124 spatially disjoint image-mask pairs collected on site) and combined them through a lightweight Mixture-of-Experts (MoE) gating network. On a held-out validation split, we report mean cereal-class intersection-over-union (IoU)=95.6 +/- 0.9% and overall accuracy=97.8 +/- 0.6% (mean +/- SD over three independent runs). Detailed comparative analysis shows that the MoE improves IoU by +1.5-2.0% relative to the best single expert. We release code, model checkpoints, and a dataset sample to enable reproducibility. We demonstrate that foundation-model fusion via MoE yields robust cereal mapping under severe label scarcity, with potential for transfer to other crops and regions.
High-resolution remote sensing images play an important role in large-scale Earth observation, providing valuable information for land use analysis, urban planning, and environmental monitoring. However, semantic segmentation of such data remains highly challenging due to complex multi-scale object distributions, diverse imaging conditions, and blurred boundaries. Existing semi-supervised methods have attempted to address the scarcity of labeled data through consistency regularization, pseudo-labeling, and generative adversarial training, yet they often suffer from over-smoothed predictions, noisy label propagation, and limited generalization capability. To overcome these challenges, we propose Mamba-based contextual label realignment with denoising latent restoration, a semi-supervised segmentation framework that integrates a state-space-driven Mamba encoder with contextual label realignment and denoising latent restoration. The Mamba encoder effectively captures long-range dependencies and global semantic context while maintaining computational efficiency, the label realignment module refines noisy pseudo-labels via contextual similarity calibration, and the latent restoration module employs a conditional diffusion process to denoise feature embeddings, enhancing the robustness of representation learning. In addition, a polygonization-based boundary refinement step is introduced to improve the quality of object contours and ensure high-fidelity segmentation results. Extensive experiments demonstrate that the proposed framework achieves significant improvements over existing methods and provides an efficient and robust solution for semi-supervised remote sensing segmentation.
Deep learning has emerged as the dominant approach for extracting water bodies from high-resolution remote sensing imagery, yet its performance depends critically on large-scale manual annotations. To overcome this limitation, we propose a fully automated water body extraction method for high-resolution RGB imagery. Specifically, this method requires no manual annotation during the training phase by relying on normalized difference water index (NDWI)-derived pseudo-labels. This method adopts a cross-scale sample transfer framework integrated with weakly supervised learning. It implements a three-stage progressive strategy that transfers samples from medium-resolution NDWI imagery (10 m) to high-resolution RGB imagery (0.5 m). First, the framework transfers reliable label information from NDWI imagery (10 m) to high-resolution RGB imagery (0.5 m). Second, we construct a weakly supervised classification model to generate class activation maps (CAMs) and propose an NDWI-constrained CAM optimization strategy to accurately convert scene-level labels into pixel-level pseudo-labels. Finally, the method adopts a cross-region joint training strategy with a composite loss function to train a deep semantic segmentation model for accurate water body extraction. We validate the proposed method on a large-scale dataset covering five typical Chinese cities. Experimental results demonstrate that under annotation-free conditions, our method achieves an IoU of 77.54% and an F1 score of 87.27% for water body extraction in unseen test regions. This performance closely approaches that of fully supervised methods (IoU = 78.50%, F1=87.90%), with an IoU gap of only 0.96% and an F1 gap of only 0.63%. This study provides an effective solution for intelligent water body extraction under annotation-free conditions and offers important technical support for applications such as flood disaster emergency monitoring and large-scale water body dynamic monitoring.
The joint classification of hyperspectral imagery (HSI) and light detection and ranging (LiDAR) data holds immense potential by leveraging their complementary nature. However, effectively extracting and adaptively fusing their heterogeneous, multiscale features while simultaneously modeling complex long-range dependencies remains a significant challenge for existing methods. To address this challenge, we propose a novel framework, the Hierarchical MultiScale Features Mamba (HMSFMamba), for joint HSI-LiDAR classification. HMSFMamba first employs a hybrid convolutional residual module to extract multiscale spatial features. It then leverages a Mamba-based module to both enhance HSI-LiDAR spatial interaction through an optimized scanning mechanism and filter HSI spectral features using a cascaded attention mechanism. Finally, an attention-gated fusion module adaptively fuses these multimodal features to achieve robust classification. Extensive experiments on three benchmark datasets demonstrate that HMSFMamba consistently outperforms state-of-the-art methods.
The Visible Infrared Imaging Radiometer Suite (VIIRS) thermal emissive band (TEB, I4-I5, and M12-M16) sensor data record (SDR) is widely used for wildfire/volcano/gas flare monitoring and other thermal anomaly studies. We evaluate and intercompares S-NPP, NOAA-20, and NOAA-21 VIIRS TEB on-orbit saturation radiance (reliable upper limit radiance) changes over time. A method for estimating TEB half-angle mirror-side and detector level saturation radiances using the VIIRS thermal model, prelaunch characterization results, on-orbit degradation factors, instrument temperatures, and telemetry data was proposed and applied to all VIIRS TEBs, except for the fire band M13, due to challenges in accurately characterizing its on-orbit degradation. For I4 and I5, S-NPP exhibits higher saturation radiances than NOAA-20 and NOAA-21. Since March 2024, the saturation radiances of NOAA-20 are similar to 0.7 K (I4) and similar to 6.5 K (I5) lower than those of S-NPP, whereas those of NOAA-21 are similar to 1.3 K (I4) and similar to 8.6 K (I5) lower than those of S-NPP. S-NPP I5 detector level saturation radiances have increased by 2 to 5 K after similar to 14 years on-orbit due to gradual on-orbit responsivity degradations. Moreover, the I5 radiance limit used in the NOAA operational processing and reprocessing is suboptimal for S-NPP (especially in recent years) and NOAA-21 during its early mission (launch-March 3, 2023), but can be improved using results from this study. Validation results using operational/reprocessed SDRs and scheduled lunar observations suggest that the proposed method can estimate I4 to I5 saturation radiances reasonably well. We also applied the proposed method to M12 and M14 to M16. The estimated long-term on-orbit saturation radiance time series, as well as corresponding validation results, are presented as well. The methodology and results we presented are valuable for improving VIIRS on-orbit calibration by enabling more good-quality high radiance observations in the TEB SDR products and eventually benefit downstream applications.
We present a deep-learning-based framework for high-efficiency mesoscale eddy prediction by integrating convolutional long short-term memory (ConvLSTM) networks for spatiotemporal forecasting with the you only look once v11 (YOLOv11) object detection model for automated eddy identification. Haiyang-2 (HY-2) sea level anomaly (SLA) data spanning 2022 to 2025 are utilized as the training dataset for the ConvLSTM model, which captures both spatial patterns and temporal dependencies inherent in SLA evolution. The training process demonstrates robust convergence, achieving a validation root mean square error of less than 0.1 m, indicating that the model effectively learns the underlying physical dynamics without substantial overfitting. Building upon these predicted SLA fields, the YOLOv11 model successfully identifies mesoscale eddies with high accuracy, achieving test detection rates of 0.84 for anticyclonic eddies and 0.85 for cyclonic eddies, with a mean average precision at an intersection over union threshold of 0.5 (mAP@0.5) of 0.752. Furthermore, a systematic analysis of the rolling forecast strategy reveals that prediction errors associated with the ConvLSTM model gradually accumulate as lead time increases. This temporal degradation manifests as increasing irregularity in the predicted SLA fields, which consequently impacts eddy identification performance. Comparative analysis with the traditional vector geometry (VG) algorithm indicates that although YOLOv11 significantly outperforms VG under regular, low-error conditions, the VG method exhibits superior robustness when applied to highly irregular predicted fields.
For over 26 years, the MODIS instrument on the Terra platform has generated a broad range of scientific products, enabling the remote sensing community worldwide for the study of many geophysical parameters of the Earth's system. Together with the complementary observations from the MODIS instrument on the Aqua platform, a continuous and consistent data record dating back to the year 2000 has been generated. MODIS, a major advance over its heritage sensors in terms of its spectral, spatial, and temporal resolutions, provides Earth scene imagery via 36 spectral bands ranging from 0.4 to 14.4 mu m at three different spatial resolutions with a wide swath of 2330 km. In addition, significantly enhanced are dedicated calibration efforts to help produce and maintain MODIS data quality through its entire mission. We present a comprehensive summary of Terra MODIS operations, the calibration strategies developed and implemented, and its on-orbit performance over the 26 years of successful operation. Terra MODIS has maintained radiometric uncertainties of less than 2% for reflective solar bands and better than 0.1 K for thermal emissive bands over its 26-year mission, far exceeding its 6-year design life. The mission's success in maintaining long-term data quality demonstrates the critical importance of comprehensive on-board calibration systems and continuous performance monitoring for multiyear Earth observations. This calibration heritage provides essential guidance for designing future multidecadal Earth observing missions.
Synthetic aperture radar (SAR) automatic target recognition (ATR), a field that has seen widespread application and significant development in recent years through the leverage of deep learning technology. However, conventional deep learning-based supervised methods face significant limitations due to their heavy dependence on large-scale annotated datasets, a requirement that conflicts with the inherent complexities of SAR imaging mechanisms and the costly nature of SAR data acquisition, ultimately leading to few-shot learning bottlenecks. Moreover, SAR images typically exhibit high resolution but relatively small target dimensions, resulting in sparse target occurrences and substantial clutter interference. To mitigate these limitations, we introduce clutter-aware self-supervised learning (CASSL), a paradigm that reconceptualizes clutter as a latent useful information rather than mere interference. By integrating self-supervised learning (SSL) strategies, CASSL exploits unlabeled SAR imagery containing diverse target categories and spatial resolutions, enabling the model to extract discriminative features from both target-specific patterns and clutter contexts. This exploitation fosters a holistic understanding of SAR image semantics, thereby enhancing feature representativeness and generalization capability. Extensive experiments on benchmark datasets, including MSTAR, HRSID, and AIR-SARShip1.0, validate the superiority of CASSL. Notably, experimental results based on the MSTAR, HRSID, and AIR-SARShip1.0 datasets demonstrate the effectiveness of the proposed method. Specifically, the accuracy on the MSTAR dataset improves by over 10%, whereas the average precision (AP) on the HRSID and AIR-SARShip1.0 datasets increases by nearly 13% and 15%, respectively. These results underscore the potential of clutter-aware SSL for few-shot SAR ATR under real-world scarcity of annotated data.
The joint classification of hyperspectral image (HSI) and light detection and ranging (LiDAR) data has become a crucial topic in remote sensing, as these modalities provide highly complementary spectral and structural information. However, effectively integrating their heterogeneous characteristics while maintaining discriminative and balanced representations remains a fundamental challenge. We propose a wavelet-enhanced multiscale cross-modal fusion network (WMCFN) to address this issue. Compared with fixed wavelet decomposition strategies, WMCFN incorporates dataset-specific wavelet selection and energy-aware subband modulation to improve frequency representation quality. The framework is designed to capture both global spectral-spatial structures and local detail variations, enabling more comprehensive multimodal representation learning. Extensive experiments on two benchmark datasets demonstrate that the proposed method consistently outperforms existing approaches in classification accuracy and generalization. Beyond performance gains, we provide insights into adaptive multimodal feature fusion and contribute to advancing the interpretability and robustness of HSI-LiDAR joint analysis.
The recent launch of the Plankton, Aerosol, Cloud, ocean Ecosystem (PACE) mission in February of 2024 was a significant step forward for the observation of ocean science data from low Earth orbit. PACE has a payload of two polarimeters and the Ocean Color Instrument (OCI); OCI is the next-generation sensor for ocean color science, drawing heritage from sensors such as MODIS, SeaWiFS, and VIIRS, but with increased spectral coverage and improved accuracy. OCI is a grating spectrometer with hyperspectral coverage from the ultraviolet (similar to 315 nm) to the near-infrared (similar to 895 nm), with additional filtered channels in the short-wave infrared (940 to 2260 nm). The calibration of OCI on-orbit is vital to the maintenance of high-quality science data products for the ocean science community. This calibration requires some parameters that can only be measured in the laboratory prior to launch. One such parameter is the response versus scan angle (RVS). The RVS measures the scan-angle-dependent change in signal intensity of the OCI optics. With the RVS, the gain calibration measured on-orbit at a specific scan angle can be transferred to any scan angle. OCI underwent rigorous ground testing to measure the RVS; results indicate that the measurements were of high fidelity and contribute to the generation of high-quality science data now that OCI is on-orbit. Total RVS variation over the on-orbit scan range is as high as 7% but typically much lower; total uncertainties are below 0.1% except in the ultraviolet below similar to 340 nm and in the short-wave infrared for 1615, 2130, and 2260 nm.
Flood disasters pose severe threats to human lives, property, and ecological security. Existing research focuses primarily on hazard-forming environment analysis, short-term warning, and post-disaster loss assessment. In recent years, although machine learning techniques have been widely applied to enhance the analysis and prediction of flood disasters, there is still a lack of studies capable of directly generating regional flood risk indices from raw data in an end-to-end manner. To address this gap, this study proposes an end-to-end framework for predicting regional flood risk indices, aiming to directly forecast future flood risk levels using historical raw data. First, we construct a four-dimensional evaluation system that includes riskiness, sensitivity, vulnerability, and resilience to assess the historical flood risk of the study area, the evaluation results serving as label data for model training. Second, we extract positional encodings from the raster data of the study area to record spatial information. Finally, using positional encodings, we integrate all data into a three-dimensional tensor, enabling the model to predict flood disaster risks in large regions. This framework avoids intermediate manual intervention and the temporal limitations of single-series prediction. It enables direct prediction of regional flood risk indices from raw data, providing a solution for dynamic flood risk mapping and early warning.
