Search

Unmanned aerial vehicle-based photon-counting ocean bathymetric light detection and ranging (LiDAR) systems rapidly acquire topographic data from islands, reefs, and shallow waters. However, the bathymetric ability of seabed topography is affected by high backscattering from the sea surface owing to its proximity, and no suitable imaging models are available. Herein, we designed a novel ray approach for bathymetry based on a dual-polarization photon-counting LiDAR. Based on the transmission characteristics of light, a dual-polarization channel strategy was proposed, and the data from two channels in vegetation, sand, and shallow and medium-depth waters were compared. Based on the ray tracing method, imaging models of the water surface and depth for light and small photon-counting LiDAR were established. Shallow-water experiments were conducted near Jiajing Island, Hainan, China, to verify the accuracy of the LiDAR bathymetry data by shipborne single-beam sounding data. The results indicate that the vertical polarization channel data had a high signal-to-noise ratio in the terrestrial part, while the horizontal polarization channel had a better water surface backscatter suppression effect and strong bathymetry ability in the water part. The detectable depth was approximately 8 m in the experimental area. The MAEs of the depth values of the LiDAR point cloud before and after refraction correction relative to the single beam depth measurement values were 1.08 and 0.09 m, respectively. And the RMSEs before and after correction were 1.12 and 0.11 m, respectively.

The interferometric synthetic aperture radar (InSAR) is able to reconstruct the Earth's surface topography with a meter-level accuracy when two-dimensional phase unwrapping (PU) is properly implemented. The two-stage programming approach (TSPA) can convert the ill-posed PU problem into a well-posed problem by integrating perpendicular baseline diversity in multiple ($\ge$2) interferograms, and is currently among the most commonly used multibaseline (MB) PU algorithms. Nevertheless, TSPA still faces two challenges in real-world applications: first, TSPA cannot ensure exceptional performance for any complex topographic scenarios, and second, the PU error of short-baseline interferometric pair tends to propagate into the PU solution of long-baseline interferometric pair, degrading height accuracy. To overcome these issues, a refined TSPA (R-TSPA) MB PU algorithm is proposed in this article. R-TSPA contains two PU procedures under the framework of TSPA, where procedure 1 unwraps the flattened interferograms with TSPA, and procedure 2 re-estimates and reunwraps the erroneous ambiguity number gradients with TSPA. It is demonstrated that R-TSPA outperforms the conventional single-baseline PU algorithms and TSPA with actual InSAR datasets in western Sichuan Province and Tibet Autonomous Region of China, revealing its potentials in accurately mapping topography and broadening application scopes of InSAR.

Remote sensing image captioning (RSIC) is a crucial task in interpreting remote sensing images (RSIs), as it involves describing their content using clear and precise natural language. However, the RSIC encounters difficulties due to the intricate structure and distinctive features of the images, such as the issue of rotational ambiguity. The existence of visually alike objects or areas can result in misidentification. In addition, prioritizing groups of objects with strong relational ties during the captioning process poses a significant challenge. To address these challenges, we propose the visual rotated position encoding transformer for RSIC. First of all, rotation-invariant features and global features are extracted using a multilevel feature extraction (MFE) module. To focus on closely related rotated objects, we design a visual rotated position encoding module, which is incorporated into the transformer encoder to model directional relationships between objects. To distinguish similar features and guide caption generation, we propose a feature enhancement fusion module consisting of feature enhancement and feature fusion. The feature enhancement component adopts a self-attention mechanism to construct fully connected graphs for object features. The feature fusion component integrates global features and word vectors to guide the caption generation process. In addition, we construct an RSI rotated object detection dataset RSIC-ROD and pretrain a rotated object detector. The proposed method demonstrates significant performance improvements on four datasets, showcasing enhanced capabilities in preserving descriptive details, distinguishing similar objects, and accurately capturing object relationships.

This article proposes an unsupervised domain adaptation (UDA) method by transferring knowledge from rich labeled optical domain to unlabeled synthetic aperture radar (SAR) domain, tackling the issue that current deep-learning-based SAR target detection methods rely on abundant labeled SAR images. Specifically, we gradually encode the dependencies across different granularity perspectives including domain invariant representations (DIR) learning based on feature decomposition and domain discriminative representations (DDR) learning based on uncertainty-guided self-training. First, existing methods usually learn the DIR by directly minimizing domain discrepancy between two domains, which is difficult to achieve in practice. Due to the huge difference between the optical and SAR images, rich domain-specific characteristics bring great challenges to learn the DIR. To alleviate the above difficulty, we explicitly model the domain-invariant and domain-specific features in the representations by constructing a network with feature decomposition to better extract the DIR across domains, where only the DIR extracted from optical images and their labels are used to train the domain-shared detector in this stage. Second, even DIR can be extracted, the domain-shared detector will lose some discriminative and valuable features of the SAR domain while minimizing the distribution discrepancy between the SAR and labeled optical domain. In order to achieve the better detection performance for SAR images, a self-training method based on pseudolabels is proposed to learn DDR and train the SAR-dedicated detector. Furthermore, for ensuring the reliability of pseudolabels, we present a novel uncertainty-guided pseudolabel selection strategy, which contains two phases: one is instance uncertainty guided selection, the other is image uncertainty guided selection. Finally, based on measured optical and SAR datasets, we conduct extensive empirical evaluation to verify the effectuality of our proposed method.

Extreme precipitation events greatly threaten people's daily lives and safety, making accurate and timely precipitation estimation especially critical. However, common methods like radar and satellite remote sensing have limitations due to coverage and environmental factors. Existing deep learning models struggle with complex scenarios and multisource data correlations. These make the precipitation estimation tasks challenging. This article proposes a Multiscale Dual Cross-Attention UNet (MS-DCA-UNet) model for near-real-time precipitation estimation. It integrates Doppler weather radar and FY-4A satellite data to overcome single-source data limitations. To narrow the semantic gap among the encoder feature maps, the MS-DCA-UNet model introduces a dual-cross attention (DCA) module at the skip connections of the backbone network U-Net. The DCA module mainly employs a channel cross-attention and a spatial cross-attention to capture remote dependencies and enable multiscale feature fusion. A multiscale convolution module is designed to reduce the risk of the model falling into local optima. It is a multibranch upsampling strategy that runs parallel to the decoder. Experimental results show that the Critical Success Index (CSI), Root Mean Square Error (RMSE), and Pearson's Correlation Coefficient (CC) of MS-DCA-UNet are 0.6033, 0.5949 mm/h, and 0.8460, respectively, with the hourly CMPAS precipitation as the benchmark. These outperform the other comparisons, such as FY-4A QPE, GPM IMERG, U-Net, Attention-UNet, and DCA-UNet on the CSI, RMSE, and CC metrics. MS-DCA-UNet reduces the RMSE of Attention-UNet, UNet, and DCA-UNet by a margin of 34.68% (0.5949 mm/h versus 0.9107 mm/h), 10.24% (0.5949 mm/h versus 0.6628 mm/h), 6.96% (0.5949 mm/h versus 0.6394 mm/h), respectively.

Remote sensing semantic segmentation tasks aim to automatically extract land cover types by accurately classifying each pixel. However, large-scale hyperspectral remote sensing images possess rich spectral information, complex and diverse spatial distributions, significant scale variations, and a wide variety of land cover types with detailed features, which pose significant challenges for segmentation tasks. To overcome these challenges, this study introduces a U-shaped semantic segmentation network that combines global spectral attention and deformable Transformer for segmenting large-scale hyperspectral remote sensing images. First, convolution and global spectral attention are utilized to emphasize features with the richest spectral information, effectively extracting spectral characteristics. Second, deformable self-attention is employed to capture global-local information, addressing the complex scale and distribution of objects. Finally, deformable cross-attention is used to aggregate deep and shallow features, enabling comprehensive semantic information mining. Experiments conducted on a large-scale hyperspectral remote sensing dataset (WHU-OHS) demonstrate that: first, in different cities including Changchun, Shanghai, Guangzhou, and Karamay, DTSU-Net achieved the highest performance in terms of mIoU compared to the baseline methods, reaching 56.19%, 37.89%, 52.90%, and 63.54%, with an average improvement of 7.57% to 34.13%, respectively; second, module ablation experiments confirm the effectiveness of our proposed modules, and deformable Transformer significantly reduces training costs compared to conventional Transformers; third, our approach achieves the highest mIoU of 57.22% across the entire dataset, with a balanced trade-off between accuracy and parameter efficiency, demonstrating an improvement of 1.65% to 56.58% compared to the baseline methods.

Deep learning methods have achieved great success in the field of remote sensing image segmentation in recent years, but building a lightweight segmentation model with comprehensive local and global feature extraction capabilities remains a challenging task. In this article, we propose a convolutional neural network (CNN) and cross-shaped transformer hybrid network (CCTNet) for semantic segmentation of high-resolution remote sensing images. This model follows an encoder–decoder structure. It employs ResNet18 as an encoder to extract hierarchical feature information, and constructs a transformer decoder based on efficient cross-shaped self-attention to fully model local and global feature information and achieve lightweighting of the network. Moreover, the transformer block introduces a mixed-scale convolutional feedforward network to further enhance multiscale information extraction. Furthermore, a simplified and efficient feature aggregation module is leveraged to gradually aggregate local and global information at different stages. Extensive comparison experiments on the ISPRS Vaihingen and Potsdam datasets reveal that our method obtains superior performance compared with state-of-the-art lightweight methods.

Fine-grained remote sensing ship detection is crucial in a variety of fields, such as ship safety, marine environmental protection, and maritime traffic management. Despite recent progress, current research suffers from the following three major challenges: insufficient features representation, conflicts in shared features, and inappropriate anchor labeling strategy, which significantly impede accurate fine-grained ship detection. To address these issues, we propose FineShipNet as a solution. Specifically, we first propose a novel blend synchronization module, which aims to facilitate the coutilization of semantic information from top-level and bottom-level features and minimize information redundancy. Subsequently, the blend feature maps are fed into a novel polarized feature focusing module, which decouples the features used in classification and regression to create task-specific discriminating features maps. Meanwhile, we adopt the adaptive harmony anchor labeling and propose a novel metric, harmony score, to choose high-quality anchors that can effectively capture the discriminating features of the target. Extensive experiments on four fine-grained remote sensing ship datasets (HRSC2016, DOSR, FGSD2021, and ShipRSImageNet) demonstrate that our FineShipNet outperforms current state-of-the-art object detection methods, achieving superior performance with mean average precision scores of 81.3%, 68.5%, 85.7%, and 63.9%, respectively.

Ship detection needs to identify ship locations from remote sensing scenes. Due to different imaging payloads, various appearances of ships, and complicated background interference from the bird's eye view, it is difficult to setup a unified paradigm for achieving multisource ship detection. To address this challenge, in this article, leveraging the large language models powerful generalization ability, a unified visual-language model called Popeye is proposed for multisource ship detection from RS imagery. Specifically, to bridge the interpretation gap across the multisource images for ship detection, a novel unified labeling paradigm is designed to integrate different visual modalities and the various ship detection ways, i.e., horizontal bounding box and oriented bounding box. Subsequently, the hybrid experts encoder is designed to refine multiscale visual features, thereby enhancing visual perception. Then, a visual-language alignment method is developed for Popeye to enhance interactive comprehension ability between visual and language content. Furthermore, an instruction adaption mechanism is proposed for transferring the pretrained visual-language knowledge from the nature scene into the RS domain for multisource ship detection. In addition, the segment anything model is also seamlessly integrated into the proposed Popeye to achieve pixel-level ship segmentation without additional training costs. Finally, extensive experiments are conducted on the newly constructed ship instruction dataset named MMShip, and the results indicate that the proposed Popeye outperforms current specialist, open-vocabulary, and other visual-language models in zero-shot multisource various ship detection tasks.

The Environmental Trace Gases Monitoring Instrument 2 (EMI-02) payload, embarked on a hyperspectral observation satellite, was launched on September 7, 2021. The satellite's transit time is 10:30 local time. This article, drawing on data from EMI-02, introduces a methodology for measuring the tropical tropospheric ozone column (TOC) using the convective-cloud-differential method. We successfully applied this method to derive TOC in tropical regions from the EMI-02 data. The results indicate a general consistency between the EMI-02 TOC and TROPOspheric Monitoring Instrument (TROPOMI) TOC. However, owing to the difference in transit times, the TOC observed by TROPOMI at 13:30 is higher than that observed by EMI-02 at 10:30. Comparisons with ozonesonde data revealed discrepancies attributed to varying observation times, but nonetheless presented a high degree of consistency. These findings affirm the efficacy of our algorithm in providing reliable and precise TOC data. This capability is instrumental in understanding the diurnal variations of tropical TOC, where concentrations are higher in the afternoon than in the morning. The characteristics of tropical TOC varied by month. From June to October, the tropical TOC predominantly accumulated over the South Atlantic region, whereas, it tended to disperse and distribute near the 20°N latitude from November to May.

Water hyacinth, an invasive species originating from South America, has become a significant concern since its introduction in Lake Victoria (Kenya), particularly in the Winam Gulf, where large annual blooms are observed. Monitoring the occurrence and location using in situ methods is expensive and challenging due to the lake's vastness. Remote sensing monitoring methods offer an alternate option due to the ability to cover vast areas. This study explores the potential of polarimetric synthetic aperture radar (PolSAR), specifically utilising Sentinel-1 VV-VH data to map and monitor water hyacinth cover. The change detection method based on optimization of power difference and minimum eigenvalue selection achieves a remarkable accuracy of 98.89% in separating clear and water hyacinth-infested water. Using polarimetric data offered better separability, enabling spatial and temporal monitoring. The analysis reveals that in 2018 water hyacinth cover peaked, spanning over 200 km$^{2}$. Temporal variability showcases a seasonal rise and peak from September to December. This research demonstrates the capability of using PolSAR data to accurately map and monitor water hyacinth's spatial and temporal dynamics, offering valuable insights for effective management strategies.

Lake surface water temperature (LSWT) is a crucial ecological indicator, impacting water quality, and aquatic life. Understanding its spatiotemporal trends and driving mechanisms is fundamental for lake water environment protection and management. Previous research has been limited by low-resolution satellite data and numerical simulations, hindering in-depth understanding of LSWT. This article fills the research gap by reconstructing a high-resolution LSWT dataset spanning 2000 to 2020. Employing data fusion techniques, we combined moderate resolution imaging spectroradiometer (MODIS) and Landsat observations, achieving a spatial resolution of 30 m and a revisit cycle of eight days. Seven major lakes in Yunnan Province, China, varying in urbanization intensity, were selected to investigate the impacts and mechanisms of urbanization and climate change on LSWT. The results showed that: First, the high spatiotemporal LSWT dataset reconstructed on the ubESTARFM data fusion model outperformed the existing product datasets in terms of accuracy evaluation and spatial details. Over the past 20 years, all LSWT in the study area exhibited a warming trend in both temporal and spatial dimensions; lakes in basins with higher urbanization intensity had significantly higher warming rates than the warming rates of near-surface air temperature, and the lakes showed a global warming trend. Second, the warming trend of LSWT is not only related to lake morphology and climate change, but also closely associated with urbanization; higher spatiotemporal resolution LSWT data revealed better spatiotemporal correlations between urbanization and LSWT. Third, active ecological management and enhanced watershed vegetation coverage could effectively mitigate the rate of lake warming.

Semantic segmentation is a fundamental task in remote sensing image analysis. The accurate delineation of objects within such imagery serves as the cornerstone for a wide range of applications. To address this issue, edge detection, cross-modal data, large intraclass variability, and limited interclass variance must be considered. Traditional convolutional-neural-network-based models are notably constrained by their local receptive fields, Nowadays, transformer-based methods show great potential to learn features globally, while they ignore positional cues easily and are still unable to cope with multimodal data. Therefore, this work proposes an adjacent-scale multimodal fusion network (ASMFNet) for semantic segmentation of remote sensing data. ASMFNet stands out not only for its innovative interaction mechanism across adjacent-scale features, effectively capturing contextual cues while maintaining low computational complexity but also for its remarkable cross-modal capability. It seamlessly integrates different modalities, enriching feature representation. Its hierarchical scale attention (HSA) module bolsters the association between ground objects and their surrounding scenes through learning discriminative features at higher level abstractions, thereby linking the broad structural information. Adaptive modality fusion module is equipped by HSA with valuable insights into the interrelationships between cross-model data, and it assigns spatial weights at the pixel level and seamlessly integrates them into channel features to enhance fusion representation through an evaluation of modality importance via feature concatenation and filtering. Extensive experiments on representative remote sensing semantic segmentation datasets, including the ISPRS Vaihingen and Potsdam datasets, confirm the impressive performance of the proposed ASMFNet.

Soil particle size is an important indicator in soil systems, it can provide important assistance for the agricultural work. In order to address the weakness of traditional soil particle size measuring work, which are time-consuming, labor-intensive, and have limited applicability. This study utilizes the Mie theory and the Kubelka–Munk theory as the precondition, establish an empirical formula between the scattering power and the soil particle size. The study collected surface soil samples from Nong'an, Changchun City, Jilin Province, including black soil, brown soil, sandy soil, and each saline sample, based on visible and near-infrared spectroscopy. Prepare soil samples with a particle size range of 2.5–0.15 mm through drying, grinding, and sieving operations, combining scattering power parameters in the K-M theory to construct an empirical formula for it and soil particle. After verified by comparing different empirical formulas are suitable for the measured data, assume the inverse proportion formula added correction term is the most appropriate. The conclusion is there is a strong linear relationship between the scattering power and the reciprocal of particle size. The average fitting accuracy of the 400–2400 nm wavelength band reaches 94.45%, root mean square error ($\text{RMSE}$) reaches 0.0354 mm. After removing outliers, the fitting accuracy can reach up to 95.77%, $\text{RMSE}$up to 0.0337 mm. Proved there is a very high analytical relationship between soil particle size and scattering power parameters in K-M theory. The empirical formula also can find supported by Mie theory and S-shape R(D) function, and has a high transferability from the laboratory to Landsat8 satellite board, the accuracy can reach to about 90% on SWIR band, showed good generalization ability.

Hyperspectral image (HSI) is pivotal in ground object classification, owing to its rich spatial and spectral information. Recently, convolutional neural networks and graph neural networks have become hotspots in HSI classification. Although various methods have been developed, the problem of detail loss may still exist when extracting complex features within homogenous regions. To solve this issue, in this article, we proposed a double-hop graph attention multiview fusion network. This model is adept at pinpointing precise attention features by integrating a double-hop graph with the graph attention network, thereby enhancing the aggregation of multilevel node information and surmounting the limitations of a restricted receptive field. Furthermore, the spectral-coordinate attention module (SCAM) is presented to seize more nuanced spectral and spatial attention features. SCAM harnesses the coordinate attention mechanism for in-depth pixel-level global spectral–spatial view. Coupled with the multiscale Gabor texture view, we forge a multiview fusion network that meticulously highlights edge details across varying scales and captures beneficial features. Our experimental validation across four renowned benchmark HSI datasets showcases our model's superiority, outstripping comparative methods in classification accuracy with limited labeled samples.

The rapid pace of urbanization underscores the importance of waste monitoring and management in urban planning and environmental conservation. Remote sensing technology enables the aerial observation of terrestrial and marine features, with high-resolution images revealing diverse objects. Deep learning techniques have gained prominence for enhancing waste monitoring precision and efficiency. This article surveys deep learning approaches for waste monitoring in remote sensing images, focusing on relevant datasets. It reviews existing remote sensing datasets, including those from uncrewed aerial vehicles and satellites, for monitoring solid waste and marine debris. Nine publicly available datasets are described in detail, highlighting their origins and applications. The monitoring methods include two kinds of methods: 1) semantic segmentation; and 2) object detection. Semantic segmentation focuses on pixel-level classification and boundary delineation, while object detection targets object-level localization and shape. Representative methods within these categories are explored, and benchmark results from recent studies are summarized to evaluate the performance of various techniques. The discussion addresses current limitations and suggests future research directions, aiming to assist researchers and professionals in environmental monitoring.

High resolution remote sensing imagery plays a crucial role in monitoring coastal wetlands. Coastal wetland landscapes exhibit diverse features, ranging from fragmented patches to expansive areas. Mainstream convolutional neural networks cannot effectively analyze spatial relationships among consecutive image elements. This limitation impedes their performance in accurately classifying coastal wetlands. In order to tackle the above issues, we propose a Vision Transformer based UNet (ViT-UNet) model. This model extracts wetland features from high resolution remote sensing images by sensing and optimizing multiscale features. To establish global dependencies, the Vision Transformer (ViT) is introduced to replace the convolutional layer in the UNet encoder. Simultaneously, the model incorporates a convolutional block attention module and a multiple hierarchies attention module to restore attentional features and reduce feature loss. In addition, a skip connection is added to the single-skip structure of the original UNet model. This connection simultaneously links the output of the entire transformer and internal attention features to the corresponding decoder level. This enhancement aims to furnish the decoder with comprehensive global information guidance. Finally, all the extracted feature information is fused using Bilinear Polymerization Pooling (BPP). The BPP assists the network in obtaining a more comprehensive and detailed feature representation. Experimental results on the Gaofen-1 dataset demonstrate that the proposed ViT-UNet method achieves a Precision score of 93.50$\%$, outperforming the original UNet model by 4.10$\%$. Compared with other state-of-the-art networks, ViT-UNet performs more accurately and finer in the extraction of wetland information in the Yellow River Delta.

Ground-penetrating radar (GPR) data are often contaminated by hardware and environmental clutter, which significantly affects the accuracy and reliability of target response identification. Existing supervised deep learning techniques for removing clutter in GPR data require generating a large set of clutter-free B-scans as labels for training, which are computationally expensive in simulation and unfeasible in real-world experiments. To tackle this issue, we propose a two-stage unsupervised learning-based clutter removal scheme, called ULCR-Net, to obtain clutter-free GPR B-scans. In the first stage of the proposed scheme, a diffusion model tailored for GPR data augmentation is employed to generate a diverse set of raw B-scans from the input random noise. With the augmented dataset, the second stage of the proposed scheme uses a contrastive learning-based generative adversarial network to learn and estimate clutter patterns in the raw B-scan. The clutter-free B-scan is then obtained by subtracting the clutter pattern from the raw B-scan. The training of the two-stage network only requires a small set of raw B-scans and clutter-only B-scans that are readily available in real-world applications. Extensive experiments have been conducted to validate the effectiveness of the proposed method. Results on simulation and measurement data demonstrate that the proposed method has superior clutter removal accuracy and generalizability and outperforms existing algebraic techniques and supervised learning-based methods with limited training data by a large margin. With its high clutter suppression capability and low training data requirements, the proposed method is well-suited to remove clutter and restore target responses in real-world GPR applications.

Hyperspectral pansharpening is to fuse a high spatial resolution panchromatic image (PAN) with a low spatial resolution hyperspectral image (LR-HSI) and generate high resolution hyperspectral image (HR-HSI). However, most existing deep learning-based pansharpening methods have some issues, such as spectral distortion and insufficient spatial texture enhancement. In this work, we propose a novel multiscale pansharpening network based on the Dual Gaussian-Laplacian Pyramid (DGLP) and Transformer, named MDTP-Net. Specifically, the DGLP module is designed to obtain feature maps at multilevel scales, which effectively learn global spectral information and spatial detail texture information. Then, we design a corresponding Transformer module for each scale feature and utilize the multihead attention mechanism to guide the extraction of spatial information from LR-HSI and PAN images. This enhances the stability of pansharpening and improves the fusion of spectral with spatial information across feature spaces. In addition, the feature extractors are inserted to connect DGLP and Transformer, making the spatial feature map smoother and richer in channel and texture features. The feature fusion and multiscale feature connection blocks are used to connect multiscale information together to generate HR-HSI images with more comprehensive spatial and spectral features. Finally, extensive experiments on three classic hyperspectral datasets are conducted. The experimental results demonstrate that our proposed MDTP-Net outperforms conventional methods and existing deep learning-based methods.

High-spatial resolution (HSR) remote sensing images present significant challenges due to their highly complex backgrounds, a large number of densely distributed small targets, and the potential for confusion with land targets. These characteristics render existing methods ineffective in accurately segmenting small targets and prone to boundary blurring. In response to these challenges, we introduce a novel multilevel feature interaction network (MFIN). The MFIN model was designed as a dual-branch U-shaped interactive decoding structure that effectively achieves semantic segmentation and edge detection. Notably, this study is the first to address ways to enhance the performance for HSR remote sensing image analysis by iteratively refining features at multilevels for different tasks. We designed the feature interaction module (FIM), which refines semantic features through multiscale attention and interacts with edge features of the same scale for optimization, then serving as input for iterative optimization in the next scale's FIM. In addition, a lightweight global feature module is designed to adaptively extract global contextual information from different scales features, thereby enhancing the semantic accuracy of the features. Furthermore, to mitigate the semantic dilution issues caused by upsampling, a semantic-guided fusion module is introduced to enhance the propagation of rich semantic information among features. The proposed methods achieve state-of-the-art segmentation performance across four publicly available remote sensing datasets: Potsdam, Vaihingen, LoveDA, and UAVid. Notably, our MFIN has only 15.4 MB parameters and 34.2 GB GFLOPs, achieving an optimal balance between accuracy and efficiency.

Road extraction from images captured via remote sensing is a pivotal task across multiple domains, encompassing urban planning and intelligent transportation systems. In the realm of high-resolution remote sensing, traditional approaches to road extraction confront obstacles pertaining to reduced accuracy and resilience. This study introduces an innovative methodology for road extraction tailored to high-resolution remote sensing data. The devised algorithm integrates a feature enhancement attention module alongside parallel feature fusion. Specifically, the feature enhancement attention module is introduced to augment the network's capacity in discerning road-related information by analyzing feature maps produced at varying resolutions. Subsequently, during feature map extraction, the parallel feature fusion technique is employed to merge shallow and deep features sharing the same resolution, thus effectively leveraging the strengths of both to enhance the model's precision. Moreover, the network undertakes the computation of correlations among feature maps of differing resolutions as well as the entire feature map, thereby facilitating a holistic grasp of the global structure and semantic information embedded within the image. Experimental evaluations conducted on the CHN6-CUG and Massachusetts datasets substantiate that the proposed approach outperforms prevailing mainstream methods for road extraction in terms of both accuracy and processing speed.

Road network plays an important role in the fields of navigation, urban planning, and transportation. Extracting road network data from imagery based on machine learning models is an efficient and economical method for obtaining road network data. In order to save labor costs, crowdsourced data can be employed to automatically acquire the labels for model training. In response to the current challenges in road extraction, such as the limited number of labeled samples, low precision of sample labels generated from crowdsourced data, and difficulty in obtaining accurate road label data, which lead to low-quality, incomplete, and inaccurate road extraction, this study proposes a label refining framework based on a road matching and integrate algorithm. Labels are generated from OpenStreetMap (OSM) vector data, and roads are extracted from very high resolution orthoimage using the U-net model. The extracted roads are then matched and integrated with the original data to generate refined labels, which are employed for further model training and road extraction. Experimental results demonstrate that this process can overcome the poor quality of samples directly generated from the OSM data, i.e., the label refining framework led to significant improvements with respect to the completeness, accuracy, and quality of the road network extraction results.

Sea surface temperature (SST) is a vital oceanic parameter that significantly influences air–sea heat flux and momentum exchange. SST datasets are crucial for identifying and describing both short-term and long-term climate perturbations in the ocean. This article focuses on cloud detection and SST retrievals in the Western Pacific Ocean, using observations obtained by the Chinese Ocean Color and Temperature Scanner (COCTS) onboard the Haiyang-1C satellite. To distinguish between clear-sky and overcast regions, reflectance after sun glint correction and brightness temperature are used as inputs for an alternative decision tree (ADTree). The accuracy of cloud detection is 93.85% for daytime and 91.98% for nighttime, respectively. Application of the cloud detection algorithm improves the accuracy and data availability (spatiotemporal coverage) of SST retrievals. We implement a nonlinear algorithm to retrieve the SST and validate these retrieved values against buoy measurements of SST. Comparisons are conducted for measurements within ±1 h and 0.01° × 0.01° of the retrieval. During the day, the bias and standard deviation (SD) are −0.01 °C and 0.63 °C, respectively, while at night, they stand at −0.08 °C and 0.71 °C, respectively. Furthermore, the intercomparison between the SST products derived from the moderate-resolution imaging spectroradiometer (MODIS) onboard Terra and the results are conducted. During the day, the bias and SD are 0.03 °C and 0.42 °C, respectively, whereas at night, they are 0.25 °C and 0.76 °C, respectively. This article improves the accuracy and applicability of the SST retrieved from the COCTS thermal infrared channels.

Low-frequency ultrawideband synthetic aperture radar (UWB SAR) has high-resolution imaging and foliage-penetrating ability, which can detect the foliage-concealed target. However, due to the jungle detection environment and the low-frequency UWB SAR characteristics, there are often some nontarget strong scattering points in low-frequency UWB SAR images, which may increase the difficulty of foliage-concealed target change detection. To improve the change detection rate of the foliage-concealed target, a foliage-concealed target change detection strategy based on image stacks in low-frequency UWB SAR images is proposed. In image preprocessing, a relative radiometric correction method based on the bidirectional linear regression model is presented, which can eliminate the low-frequency UWB SAR image changes caused by nontarget factors. Besides, in change detection processing, multiple difference images are first obtained by subtracting the image to be detected from multiple reference images. Then, the Gaussian probability density function is used to model the distribution of the amplitude of these difference images. Finally, the generalized likelihood ratio test is used for target change detection, which effectively suppresses the interference, such as tree trunk clutter. Experimental results tested on the CARABAS-II SAR dataset demonstrate the correctness and effectiveness of the proposed strategy, which can improve the change detection probability of the foliage-concealed target with the lower false alarm rate.

Total precipitable water (TPW) is defined as the vertically integrated column water vapor from the earth's surface to the top of the atmosphere. TPW is a key element of the hydrological cycle and is responsive to changes in global climate related to greenhouse-gas-induced warming. In this research, we focus on trend analysis using the TPW retrieval product from the recently reprocessed Microwave Integrated Retrieval System (MiRS) Suomi National Polar-Orbiting Partnership (SNPP) Advanced Technology Microwave Sounder (ATMS) data and compare it with ERA5 reanalysis. The primary results show that the global TPW trend during 2012–2021 from reprocessed SNPP ATMS is 0.46 mm/decade, in relatively good agreement with the trend from ERA5 of 0.39 mm/decade. Trends for tropical and mid-latitude subregions are also in good agreement, with essentially the same trend of 0.43 mm/decade seen in both datasets in the mid-latitudes. Both the datasets show a large positive anomaly associated with the strong El Nino event in 2015–2016, which increased TPW amounts in the tropics. We also found that the TPW trend is not uniformly distributed spatially, with significant regional variations in both sign and amplitude. Nevertheless, the spatial patterns from MiRS SNPP ATMS retrievals and ERA5 analyses are in very good agreement. Both the datasets show that positive TPW trends in terms of relative percentage in the polar regions were on par with those seen in lower latitudes. The results suggest that water vapor observations from a single polar-orbiting microwave instrument with only two local observation times daily may be sufficient to characterize trends in TPW.

Recently, a sparse super-resolution method based on $L_{1}$ iterative reweighted norm (IRN) has been proposed to improve the azimuth resolution of forward-looking radar. However, this method suffers from poor adaptability and high computational complexity due to its noise-sensitive user-parameter and the necessity for high-dimensional matrix inversion. To this end, a fast adaptive $L_{1}$-IRN sparse super-resolution method is derived in this article, allowing for the user-parameter-free and efficient sparse imaging of forward-looking radar. First, we establish the super-resolution model of forward-looking radar and analyze the user parameter selection problem in the conventional $L_{1}$-IRN method. Second, based on Bayesian theory, adaptive iterative weights of different azimuths are derived by transforming the sparse estimation problem into a maximum posterior (MAP) estimation problem. Finally, by using QR decomposition and Sherman–Morrison formula, the dimensionality of the echo and antenna pattern involved in the iteration is reduced to further diminish the computational complexity. Compared to the existing $L_{1}$-IRN method, the proposed method eliminates the need for any user parameters, and the computational complexity has been reduced from ${O}({JN}^{3})$ to ${O}({JN}^{2}{a})$. Simulation and measured data demonstrate the superiority of the proposed method.

Red–green–blue (RGB) images (or videos) captured by consumer-level uncrewedaerial vehicle (UAV) cameras are widely used in high-resolution remote observations. However, digital number (DN) values of these RGB images usually have a nonlinear relationship with the incident radiance, which reduces the accuracy of quantitative remote sensing of macroalgae. To solve this problem, we proposed an improved processing procedure for UAV RGB images (or videos) based on camera response functions (CRFs). The CRF was utilized to convert the DN values into energy values (E values), which demonstrate a linear relationship with the incident radiance. When the DN values were replaced by their corresponding E values to calculate the reflectance of green macroalgae under different illumination intensities, the errors in reflectance were reduced by ∼21%; for the corresponding green macroalgae indices, such as the red–green band virtual baseline floating green algae height (RG-FAH), the E-value-based RG-FAH demonstrates more resistance to the impacts of sun glints; and the E values were further applied to estimate the coverage portion of macroalgae (POM, %) in RGB videos; the illumination-induced deviations of the POM were effectively reduced by up to 33.06%, showing an advantage in quantitative estimation of macroalgae biomass. The results of applications to UAV RGB images show that the E values have significant suitability in estimating POM across diverse green macroalgae species and various algae indices, suggesting promising potentials of the proposed processing procedure with E-based photo and/or video RGB images in monitoring aquatic plants and environment.

With the accelerating impact of global warming, the changes of Arctic sea ice has become a focal point of research. Due to the spatial heterogeneity and the complexity of its evolution, long-term prediction of Arctic sea ice remains a challenge. In this article, a spatial attention U-Net (SAU-Net) method integrated with a gated spatial attention mechanism is proposed. Extracting and enhancing the spatial features from the historical atmospheric and SIC data, this method improves the accuracy of Arctic sea ice prediction. During the test periods (2018–2020), our method can skillfully predict the Arctic sea ice up to 12 months, outperforming the naive U-Net, linear trend models, and dynamical models, especially in extreme sea ice scenarios. The importance of different atmospheric factors affecting sea ice prediction are also analyzed for further exploration.

In geospatial image analysis, domain shifts caused by differences between datasets often undermine the performance of deep learning models due to their limited generalization ability. This issue is particularly pronounced in hyperspectral imagery, given the high dimensionality of the per-pixel reflectance vectors and the complexity of the resulting deep learning models. We introduce a semisupervised domain adaptation technique that improves on the adversarial discriminative framework, incorporating a novel multiclass discriminator to address low discriminability and negative transfer issues from which current approaches suffer. Significantly, our method addresses mode collapse by incorporating limited labeled data from the target domain for targeted guidance during adaptation. In addition, we integrate an attention mechanism that focuses on challenging spatial regions for the target mode. We tested our approach on three unique hyperspectral remote sensing datasets to demonstrate its efficacy in diverse conditions (e.g., cloud shadows, atmospheric variability, and terrain). This strategy improves discrimination and reduces negative transfer in domain adaptation for geospatial image analysis.

Coastal wetland ecosystem is vital for carbon sequestration, making the accurate carbon sink estimation essential for its protection and management. Traditional carbon sink estimation methods have overlooked the influence of moist soil on sparse vegetation, resulting in the inaccurate estimation of net primary productivity (NPP), especially in coastal areas with mixed wetlands and vegetation. To address this challenge, this study proposed an improved Carnegie–Ames–Stanford approach model for NPP estimation, which utilizes the modified soil-adjusted vegetation index (MSAVI) to eliminate the background noise of moist soils and calculate the fraction of photosynthetically active radiation. By using MOD17A3 as reference data for comparative experiment, the accuracy of NPP results is improved by 89.6 gC·m−2. The proposed model was then used for carbon sink estimation and analysis of Shandong coastal area. The results indicate the following: First, the average NPPMSAVI across Shandong coastal area was improved by 99.12 gC·m−2, 36.17%, and 60.53 gC·m−2 in BIAS, relative bias, and root-mean-square error, respectively. Second, the spatial distribution of net ecosystem productivity (NEP) in Shandong coastal area is higher in the east and lower in the west, with mean values of approximately 210 gC·m−2 in the east and 60 gC·m−2 in the west. The seasonal differences in NEP among different land types are significant. Third, NEP exhibits a strong correlation with temperature, precipitation, and solar radiation, with mean r of 0.78, 0.8, and 0.84, respectively.

Ship detection achieves great significance in remote sensing of synthetic aperture radar (SAR) and many efforts have been done in recent years. However, distinguishing ship targets precisely from the interference of multiplicative non-Gaussian coherent speckle is still a challenging task due to the discreteness, variability, and nonlinearity of ship scattering features. A detection framework based on hierarchical sampling representation is introduced to alleviate the phenomenon in this article. First, ships in SAR images exhibit multiplicative non-Gaussian coherent speckle, which introduces nonlinear characteristics under the imaging mechanism of SAR. Therefore, a statistical feature learning module is proposed with a learnable design to describe the nonlinear representations and expand the feature space. Second, our method designs a convex-hull representation to fit the irregular contours of ships represented by strong scattering points. Third, in order to supervise and optimize the regression of convex-hull representation, a sparse low-rank reassignment module is employed to evaluate the positive samples with SAR mechanism and reassign ones of high quality, which produces better results. Furthermore, experimental results on three authoritative SAR-oriented datasets for ship detection application present the comprehensive performance of our method.

Utilizing massive spatio-temporal sequence data and real-time synthetic aperture radar (SAR) ship target monitoring technology, it is possible to effectively predict the future trajectories and intents of ships. While real-time monitoring technology validates and adjusts spatio-temporal sequence prediction models, it still faces challenges, such as manual anchor box sizing and slow inference speeds due to large computational parameters. To address this challenge, a SAR ship target real-time detection method based on CenterNet is introduced in this article. The proposed method comprises the following steps. First, to improve the feature extraction capability of the original CenterNet network, we introduce a feature pyramid fusion structure and replace upsampled deconvolution with Deformable Convolution Networks (DCNets), which enable richer feature map outputs. Then, to identify nearshore and small target ships better, BiFormer attention mechanism and spatial pyramid pooling module are incorporated to enlarge the receptive field of network. Finally, to improve accuracy and convergence speed, we optimize the Focal loss of the heatmap and utilize Smooth L1 loss for width, height, and center point offsets, which enhance detection accuracy and generalization. Performance evaluations on two SAR image ship datasets, HRSID and SSDD, validate the method's effectiveness, achieving Average Precision (AP) values of 82.87% and 94.25%, representing improvements of 5.26% and 4.04% in AP compared to the original models, with detection speeds of 49 FPS on both datasets. These results underscore the superiority of the improved CenterNet method over other representative methods for SAR ship detection in overall performance.

With a widespread adoption of synthetic aperture radar (SAR) observations in Earth sciences, the volume of annual data updates has soared to petabyte scales. Consequently, the accurate retrieval and efficient storage of SAR data have become pressing concerns. The existing data searching method exhibits significant redundancy, leading to wasteful consumption of bandwidth and storage resources. Aiming to address this issue, we present here an optimized retrieval method grounded in a greedy algorithm, which can substantially reduce redundant data by approximately 20–65% while ensuring comprehensive data coverage over the areas of interest. By significantly minimizing redundant data, the proposed method markedly enhances data acquisition efficiency and conserves storage space. Validation experiments with Sentinel-1 data, employing various keyhole markup language scope files as inputs, affirm the effectiveness and reliability of the method. The application of the proposed method is expected to pave the way for efficient data management and fully automatic InSAR processing.

With the help of interferometric synthetic aperture radar (InSAR) and global navigation satellite system (GNSS) technology, high spatial and temporal resolution surface deformation results can be generated, which can help to better understand the mechanism of surface deformation. The spatio-temporal Kalman-based InSAR and GNSS fusion method fully considers the spatio-temporal correlation of deformation and characterizes the potential spatio-temporal process of deformation through spatial modeling and Kalman filter or smooth. However, when focusing on deformation with typical subsidence spatial characteristics, existing studies did not fully consider the spatial distribution of deformations, resulting in the loss of spatial detail information. This article proposes an adaptive spatial modeling optimization method that takes into account the spatial distribution of deformation, which can capture the optimal spatial basis layout scheme under a limited number of spatial bases, establish a more accurate spatial model, and improve the accuracy of deformation fusion. The effectiveness and reliability of this method are verified through simulation experiments and the deformation monitoring results in Datong City, China. The results of the two experiments show that the proposed method can improve the accuracy of InSAR interpolation results by 27.7% and 11.5% on average, respectively.

Multiple modal data fusion can provide valuable and diverse information for remote sensing image segmentation. However, the existing fusion methods often lead to feature loss during the fusion of various modal data, and the complementarity among multimodal features is insufficient. To address these problems, we propose a multimodal gated segmentation network for remote sensing images based on the frequency decomposition. Complementary information from multimodal features is extracted by establishing a long-distance correlation between the low-frequency components of different modal data. In addition, high-frequency detailed features of different modal data are preserved by residual connection. The adaptive gated fusion method is then used to control the information flow between the complementary information and each modality feature map, enabling adaptive fusion between multimodal features. These operations can effectively improve the adaptability of the proposed method in various scenarios and data changes. Extensive experiments demonstrate that the proposed method has good effectiveness, robustness, and generalization and achieved state-of-the-art performance in several remote sensing image semantic segmentation tasks.

Building change detection on remote sensing images is an important approach to monitoring the urban expansion and sustainable development of natural resources. In conventional building change detection tasks, only changed regions between two time phases are typically concerned. The relevance and trend of spatiotemporal changes between multiple time phases are neglected in most cases. In this article, we propose a two-stage continual urbanization detector (COUD) for time series urban building change detection task. The COUD method employs self-supervised pretraining for feature refinement, and performs optimization through temporal distillation approach. Consequently, multitemporal feature extraction and changing regions localization of urban building complexes are conducted. Considering the gap in available dataset for time series change detection task, we produce and release a time series dataset named “TSCD”. Chengdu region of China is selected as the study area in this research, which is partially covered by the proposed TSCD dataset. By applying the proposed COUD method to the selected study area for exploring the changing pattern from 2016 to 2022, a comprehensive analysis is conducted in conjunction with actual planning policies published by the management department. Extensive experimental results confirm the reliability of our proposed method.

Most current researches primarily focus on improving experimental accuracy using large models, often neglecting the deployment challenges. There is a growing need for lightweight algorithms in certain remote sensing devices. Moreover, remote sensing images (RSIs) often contain numerous small, densely distributed targets, which pose significant detection challenges. To address these issues, we have improved the YOLOv8s network and developed a lightweight remote sensing object detection (RSOD) network based on multiscale features fusion and context information (MFFCI–YOLOv8). This network combines multiscale feature fusion and contextual information to accurately detect objects in RSIs. First, we introduce the lightweight CSP bottleneck with attention module, which utilizes partial convolution calculation and SimAM attention mechanisms to decrease the number of parameters and computational complexity while enhancing feature extraction capabilities. Second, we design the gate spatial pyramid pooling fast module to enhance the model's perception of scale and contextual information, thus improving the detection of small objects. Last, we employ the multiscale fusion lightweight neck module for more efficient multiscale feature fusion, preventing the loss of small objects. Compared to YOLOv8s, our overall model reduces the number of parameters by 7.7% and FLOPs by 11.9%. We validated the accuracy of MFFCI–YOLOv8 on two remote sensing datasets, NWPU VHR-10 and VisDrone. The experimental results demonstrate that our model offers a low computational cost and high detection accuracy compared to other RSOD models and other YOLO models.

Fine-grained image recognition (FGIR), unlike traditional coarse-grained recognition, is centered on distinguishing fine-level subclasses within broader semantic categories. It holds significant scientific research value, particularly in remote sensing, where the precise identification of specific objects—such as ships, buildings, and land use categories—is critical for tasks like boundary security, environmental monitoring, and urban planning. Recent advancements in FGIR have notably improved feature representation and generalization, especially under the diverse imaging conditions typical of remote sensing. However, challenges remain, including the heavy reliance on high-quality large-scale fine-grained image data and difficulties in extracting subtle image features. Efficiently utilizing limited data and enhancing feature extraction capabilities have thus become key focus areas in current FGIR research. This article systematically reviews the advancements in FGIR, covering its foundational principles, key methodologies, and the latest research developments, while providing a comprehensive comparative analysis of their performance in remote sensing image applications. In addition, it addresses the specific challenges posed by fine-grained recognition in remote sensing imagery and explores potential directions for future research in this field.

3-D building models are one of the most intuitive and widely used forms for understanding urban buildings. Generating 3-D building models based on a single off-nadir satellite image is an economical and rapid method, particularly valuable in large-scale 3-D reconstruction scenarios with limited time. In this article, we propose a novel pipeline for automatically reconstructing level of detail 1 (LoD1) 3-D building models based on a single off-nadir satellite remote sensing image. Our pipeline is built upon a multitask neural network called off-nadir building reconstruction network (ONBuildingNet), which extracts building roof polygons and offsets from the image. Using this information, the pipeline computes the building footprint polygons and heights, constructs LoD1 building models, and then extract textures from the off-nadir image. ONBuildingNet introduces our proposed cross-field auxiliary task and multiscale mask head to extract building roof polygons with accurate shapes. We have demonstrated through extensive experiments that our pipeline can automatically and rapidly construct LoD1 3-D urban building models. In addition, our proposed ONBuildingNet outperforms current state-of-the-art methods in extracting more shape accurate building roof polygons, thereby enhancing the accuracy of the final 3-D models produced by our pipeline. Experimental results demonstrate that our method for reconstructing 3-D models of urban building scenes has strong visualization effects, with an average height error of 3.3 m.

Semantic segmentation models experience a significant performance degradation due to domain shifts between the source and target domains. This issue is particularly prevalent in remote sensing imagery, where a semantic segmentation model trained on images from one satellite is tested on images from another. Previous research has often overlooked the role of data augmentation in enhancing a model's adaptability to target domains. In contrast, this article proposes a novel self-training framework that incorporates data augmentation at both the input and feature levels, yielding excellent results. Specifically, we introduce a regularized online self-training framework that effectively addresses the challenges of overconfidence and class imbalance inherent in self-training. Based on this framework, we implement two robust data augmentation strategies at the input and feature levels to facilitate the learning of cross-domain invariant knowledge. At the input level, we employ a large-scale domain mixing strategy, termed multidomain mixing, to enhance the model's generalization capability. At the feature level, we introduce masked feature augmentation, a masking-based perturbation technique applied to the semantic features of the target domain. This approach enhances the consistency of teacher–student network predictions in the target domain feature space, thereby improving the robustness of the model's recognition of target domain features. The integration of the proposed self-training framework with dual-level data augmentation culminates in our innovative self-training-based dual-level data augmentation (STDA) method. Extensive experimental results on the ISPRS semantic segmentation benchmark demonstrate that STDA outperforms existing state-of-the-art methods, showcasing its effectiveness.

Given the inherent limitations of camera-only and LiDAR-only methods in performing semantic segmentation tasks in large-scale complex environments, multimodal information fusion for semantic segmentation has become a focal point of contemporary research. However, significant modal disparities often result in existing fusion-based methods struggling with low segmentation accuracy and limited efficiency in large-scale complex environments. To address these challenges,we propose a semantic segmentation network with camera–LiDAR cross-attention fusion based on fast neighbor feature aggregation (MFSA-Net), which is better suited for large-scale semantic segmentation in complex environments. Initially, we propose a dual-distance attention feature aggregation module based on rapid 3-D nearest neighbor search. This module employs a sliding window method in point cloud perspective projections for swift proximity search, and efficiently combines feature distance and Euclidean distance information to learn more distinctive local features. This improves segmentation accuracy while ensuring computational efficiency. Furthermore, we propose a cross-attention fusion two-stream network based on residual, which allows for more effective integration of camera information into the LiDAR data stream, enhancing both accuracy and robustness. Extensive experimental results on the large-scale point cloud datasets SemanticKITTI and Nuscenes demonstrate that our proposed algorithm outperforms similar algorithms in semantic segmentation performance in large-scale complex environments.

In recent years, the development of continuous-scale super-resolution (SR) methods in the field of remote sensing (RS) has garnered significant attention. These innovative methods are capable of delivering arbitrary-scale image SR through a single unified network. However, the majority of these methods employ the same feature extractor for different SR scales, which constrains the enhancement of network performance. Furthermore, the utilization of a multilayer perceptron for image reconstruction results in the loss of a substantial quantity of high-frequency information, which is of particular significance in the context of RS images. This, in turn, gives rise to the generation of blurred SR results. In order to address the aforementioned issues, the scale-frequency dual-modulation network (SFMNet) is proposed as a means of achieving RS image continuous SR. First, scale modulation feature fusion can modulate different levels of feature fusion according to different scale factors, thereby fully integrating the scale information into the feature extraction process of the network. Subsequently, frequency modulation reconstruction can modulate the frequency-domain information at the root of the image reconstruction process, thereby enhancing the ability of the network to learn high-frequency information. The experimental results demonstrate that the proposed SFMNet outperforms existing RS image continuous SR methods in terms of quantitative indices and visual quality.

Earth observation utilizes multisource satellite data to enhance photogrammetry mapping. This study introduces a novel method to improve the geometric positioning accuracy of large-scale optical satellite imagery by combined adjustment with NASA's Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) laser altimetry data. Although ICESat-2 is known for its high vertical accuracy, its potential to improve horizontal accuracy has been limited due to the difficulty in matching caused by temporal inconsistencies across large survey areas. To address this, our method employs a robust terrain profile elevation sequence similarity matching technique, refined with two-dimensional Gaussian fitting to achieve enhanced position extraction. We also propose a weighted adjustment strategy that uses matching confidence to enhance the precision of the combined adjustments. Large-scale tests across various terrains showed that our approach has significantly reduced horizontal and vertical positioning errors to 4.3 and 1.7 m, respectively, outperforming existing methods.

We developed a cross-calibration method to improve sea surface temperature (SST) retrieval from remotely sensed thermal infrared data from space. The cross-calibration method is based on a physical SST method that determines SST by solving the infrared radiative transfer equation. We calibrated the thermal infrared data from Himawari-9, a Japanese geostationary meteorological satellite, using the developed method and the reference observation obtained with the Second-generation Global Imager (SGLI) onboard the Global Change Observation Mission-Climate (GCOM-C) satellite. As a result, the bias and the standard deviation of Himawari-9 SST were improved from $-$0.63 to $-$0.028 K and from 1.2 to 0.73 K, respectively, compared to the buoy data. Consistency between the Himawari-9 SST and SGLI SST was also improved after this calibration. Meanwhile, negative bias remained in the Himawari-9 SST, which was determined using 3.9 $\mu$m data if the observation zenith angle was larger than 50°. This article discusses the cross-calibration method and the validation results of SST retrieved from the calibrated Himawari-9 data.

Nano and microsatellites have expanded the acquisition of satellite images with higher spatial, temporal, and spectral resolutions. Nevertheless, downlinking all this data to the ground for processing becomes challenging as the amount of remote sensing data rises. Custom onboard algorithms are designed to make real-time decisions and to prioritize and reduce the amount of data transmitted to the ground. However, these onboard algorithms frequently require cloud-free bottom-of-atmosphere surface reflectance (SR) estimations as inputs to operate. In this context, this article presents the data transformations and autocalibration for Sentinel-2 (S-2) network (DTACSNet), an onboard cloud detection and atmospheric correction processor based on lightweight convolutional neural networks. DTACSNet provides cloud and cloud shadow masks and SR estimates 10× faster than the operational S-2 L2A processor in dedicated space-tested hardware: 7 mins versus 1 h for a 10 980 × 10 980 scene. The DTACSNet cloud masking, based on a lightweight neural network, obtains the highest F2-score (0.81), followed by the state-of-the-art KappaMask (0.74), Fmask (0.72), and Sen2Cor v.2.8 (0.51) algorithms. In addition, validation results on independent datasets show that DTACSNet can efficiently replicate Sen2Cor SR estimates, reporting a competitive accuracy with differences below 2%.

In this work, we address the bidirectional reflectance distribution function (BRDF) characterization of homogeneous surfaces by means of multiangular datasets acquired with an unmanned aerial system (UAS) carrying a multispectral sensor (MAIA) replicating the spectral characteristics of the multispectral instrument onboard Copernicus Sentinel-2 satellite. The UAS field campaign was performed in clear-sky conditions over two different test sites, a vegetation cover and an asphalted area, exhibiting different behaviors in terms of surface reflectance anisotropy. A dual angular approach for the processing of the reflectance measurements is examined: a conical configuration considering a cone angle of 10° (hemispherical conical-reflectance distribution) and a directional configuration (hemispherical directional-reflectance distribution) considering a cone angle of 3°. Afterward, the retrieval of the parameters of the Ross–Li–Maignan BRDF model was implemented by a least-squared fitting of the UAS reflectance measurements for each MAIA band. The accuracy of the modeled reflectances was evaluated and the overall relative root-mean-square error between the measured and modeled reflectances was less than 10% for both test sites. The outcomes of the present study go toward the definition of a standard approach for UAS-based measurements with high angular resolution features for BRDF modeling, avoiding the well-known issues related to the use of ground-based and satellite-based instruments, and proving the UAS effectiveness in supporting calibration and validation activities of satellite missions.

In this article, we consider the issue of change detection (CD) for heterogeneous remote sensing images. Existing deep learning-based methods for CD usually utilize square convolution receptive fields, which do not sufficiently exploit the contextual and boundary information in heterogeneous images. To address the aforementioned issue, we propose an enhanced and unsupervised Siamese superpixel-based network for CD in heterogeneous remote sensing images. Our newly proposed method innovatively combines superpixels with the square receptive fields to generate the boundary adherence receptive fields and better capture the contextual information than existing methods only with the regular square receptive fields. Furthermore, we utilize an adaptive superpixel merging module to prevent the oversegmentation of superpixels and strengthen the robustness of our method in terms of superpixel sizes. Experiments based on four real datasets demonstrate that the proposed method achieves higher accuracy than other commonly used CD methods in heterogeneous remote sensing images.

Remote sensing image spatiotemporal fusion (STF) aims to generate composite images with high-temporal and spatial resolutions by combining remote sensing images captured at different times and with different spatial resolutions (DTDS). Among the existing fusion algorithms, deep learning-based fusion models have demonstrated outstanding performance. These models treat STF as an image super-resolution problem based on multiple reference images. However, compared to traditional image super-resolution tasks, remote sensing image STF involves merging a larger amount of multitemporal data with greater resolution difference. To enhance the robust matching performance of spatiotemporal transformations between multiple sets of remote sensing images captured at DTDS and to generate super-resolution composite images, we propose a feature fusion network called the multiscale deformable convolution distillation generative adversarial network (DCDGAN-STF). Specifically, to address the differences in multitemporal data, we introduce a pyramid cascading deformable encoder to identify disparities in multitemporal images. In addition, to address the differences in spatial resolution, we propose a teacher–student correlation distillation method. This method uses the texture details' disparities between high-resolution multitemporal images to guide the extraction of disparities in blurred low-resolution multitemporal images. We comprehensively compared the proposed DCDGAN-STF with some state-of-the-art algorithms on two landsat and moderate-resolution imaging spectroradiometer datasets. Ablation experiments were also conducted to test the effectiveness of different submodules within DCDGAN-STF. The experimental results and ablation analysis demonstrate that our algorithm achieves superior performance compared to other algorithms.

We propose a novel multilevel unsupervised domain adaptive framework for single-stage object detection in remote sensing images. Our framework combines pixel-level adaptation together with feature-level adaptation in a progressive learning scheme. Pixel-level adaptation usually suffers from the imperfect translation problem with respect to local region deformation. To address this problem, we introduce a semantically important region-attentive pixel-level domain adaptation based on a cycleGAN-like translation design, which incorporates an attention module and a learnable normalization function to facilitate shape transformation and image style transfer across domains. Moreover, to adapt single-stage detector while removing the need for explicit local features, we introduce the attention-guided multiscale feature-level domain adaptation, which employs multiple domain discriminators at different scales to perform multiscale feature alignment for objects of different sizes. This alignment process is guided from global to local by exploiting a self-attention mechanism that allows the model to gradually recognize local regions. The experimental results on several remote sensing datasets demonstrate the validity of our proposed framework. Compared with the baseline detector trained on the source dataset, our approach consistently improves the detection performance on the target dataset by 9.1%–16.0% mAP and achieves state-of-the-art results under various datasets.

Deploying image target detection algorithms on embedded devices is critical. Previous studies assumed that fewer model parameters and computations improved the inference speed. However, many models with few parameters and computations have slow inference speeds. Therefore, developing a remote sensing image target detection model that can perform real-time inference on embedded devices is required. We propose a fast target detection model for remote sensing images leveraging roofline analysis on edge computing devices (FTD-RLE). It comprises three parts: (1) We analyze the hardware characteristics of embedded devices using RoofLine and incorporate global features to design a model structure based on the operational intensity (OI) and arithmetic intensity (AI) of embedded devices. (2) The mirror ring convolution (MRC) is designed for extracting global features. The global information-aware module (GIAM) extracts local features from key areas using the global feature guidance model. The global-local feature pyramid module (GLFPM) is proposed to combine global and local features. (3) Additionally, hardware deployment and inference acceleration technologies are implemented to enable the model's deployment on edge devices. TensorRT and quantization methods are used to ensure fast inference speed. The proposed algorithm achieves an average detection accuracy of 92.3% on the VHR-10 dataset and 95.2% on the RSOD dataset. It has 1.26 M model parameters, and the inference time for processing one image on Jetson Orin Nx is 8.43ms, which is 1.90 ms and 1.98 ms faster than the mainstream lightweight algorithms ShufflenetV2 and GhostNet, respectively.