Milena Radenković

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS)

A Spectral Sequence-Based Nonlocal Long Short-Term Memory Network for Hyperspectral Image Classification

Baoxiang Huang; Zhipu Wang; Jie Shang; Ge Chen; Milena Radenkovic

https://ieeexplore.ieee.org/document/9736454

Efficient classification for hyperspectral images (HSI), which assigns each pixel of the image to a specific category, has been a critical research topic in HSI analysis. Under supervised classification settings, deep learning approaches are very useful for label prediction. However, most deep learning modeling methods cannot fully exploit spectral information, which is critically important for object interpretation. Consequently, a spectral sequence-based nonlocal long short-term memory (LSTM) network for HSI classification is proposed in this article. To strengthen the dominant role of spectral information, an LSTM network for sequential data processing is employed. Furthermore, nonlocal diverse regions are exploited to learn contextual features for stronger discriminative ability. Meanwhile, an attention mechanism is adopted to improve classification performance. Experiments on Salinas, Indian Pines, Pavia University, and Houston University datasets demonstrate that spectral-spatial information is effectively fused to achieve accurate pixel-level classification and outperform state-of-the-art methods.

attention mechanism; hyperspectral image (HSI) classification; long short-term memory (LSTM); nonlocal diverse region; spectral information

Časopis

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS)

An Efficient Oceanic Eddy Identification Method With XBT Data Using Transformer

Hongfeng Zhang; Baoxiang Huang; Ge Chen; Linyao Ge; Milena Radenkovic

https://ieeexplore.ieee.org/document/9944871

Oceanic mesoscale eddies are relatively small, short-lived circulation patterns that are approximately in geostrophic balance. They are omnipresent and are commonly characterized by dynamic sea level anomalies and temperature anomalies, making sea level anomaly (SLA) a mainstream source for eddy identification. However, nearly 90% of sea level dynamic anomalies caused by oceanic eddies cannot be observed due to the insufficient resolution of satellite altimeters. By combining in situ Expendable Bathythermograph (XBT) profile data and sea surface temperature data calibrated by altimeters, this article proposes a deep neural network for identifying subsurface oceanic eddies and inverting the corresponding sea surface eddy properties. Experimental results show a classification accuracy of 98.22%, and reclassification recaptures about 36% of eddies outside altimeter-identified samples, demonstrating the effectiveness of vertical profiles in deep-learning-based eddy identification.

artificial intelligence; deep learning; deep neural network; oceanic eddy; vertical structure; XBT profile

Časopis

IEEE Geoscience and Remote Sensing Letters (GRSL)

Oceanic Eddy Identification Using Pyramid Split Attention U-Net With Remote Sensing Imagery

Nan Zhao; Baoxiang Huang; Jie Yang; Milena Radenkovic; Ge Chen

https://ieeexplore.ieee.org/document/10041953

Oceanic eddies are ubiquitous ocean flow phenomena and key factors in the transport of ocean energy and materials. This letter proposes a pyramid split attention (PSA) eddy detection U-Net architecture (PSA-EDUNet) for oceanic eddy identification from remote sensing imagery. The model builds on U-Net encoder-decoder architecture and introduces a PSA module to enhance feature extraction. The fusion data include sea surface temperature (SST) and sea level anomaly (SLA) as primary criteria for eddy identification. Experiments on the Kuroshio Extension and South Atlantic regions show that the proposed method outperforms other approaches, especially for eddy edges and small-scale eddies.

deep learning; oceanic eddy identification; pyramid split attention (PSA); U-Net network

Časopis

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS)

WildFishNet: Open Set Wild Fish Recognition Deep Neural Network With Fusion Activation Pattern

Xiaoya Zhang; Baoxiang Huang; Ge Chen; Milena Radenkovic; Guojia Hou

https://ieeexplore.ieee.org/document/10197176

Wild fish recognition is a fundamental problem in ocean ecology research and contributes to understanding biodiversity. Due to the large number of wild fish species and unrecognized categories, the task is essentially an open-set fine-grained recognition problem. This article proposes WildFishNet, a deep neural network for open-set wild fish recognition with a fused activation pattern. The model combines reciprocal inverted residual modules via neural architecture search for improved fine-grained feature extraction and introduces a fusion activation pattern of softmax and openmax functions to improve open-set recognition. Experiments on the WildFish dataset (54,459 unconstrained images; 685 known classes and one open-set unknown category) demonstrate the effectiveness of the proposed approach.

deep neural network; fusion activation pattern; neural structure search; open set fine-grained recognition; wild fish recognition

Časopis

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS)

SCMNet: Toward Subsurface Chlorophyll Maxima Prediction Using Embeddings and Bi-GRU Network

Ao Wang; Baoxiang Huang; Jie Yang; Ge Chen; Milena Radenkovic

https://ieeexplore.ieee.org/document/10288355

A subsurface chlorophyll maximum is an important ecological feature of planktonic ecosystems. Although vertical profiles can be obtained using biogeochemical (BGC)-Argo buoys, this approach does not meet the requirements for high-resolution spatiotemporal ocean observation. This study demonstrates that deep learning can help fill these observational gaps when combined with sea surface data from ocean color radiometry. A benchmark dataset is constructed by fusing sparse BGC-Argo vertical profiles with sea surface data. Inspired by dense numerical representations, a model combining embeddings and a bidirectional gated recurrent unit (Bi-GRU) network is proposed to infer vertical profiles from BGC-Argo and satellite data. Extensive experiments in the Northwest Pacific demonstrate the effectiveness of the methodology and show substantial improvements in predictive performance.

biogeochemical (BGC)-Argo; bidirectional gated recurrent unit (Bi-GRU) network; deep embedding; remote sensing; subsurface chlorophyll maxima

Časopis

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS)

ARU²-Net: A Deep Learning Approach for Global-Scale Oceanic Eddy Detection

Junmin Geng; He Gao; Baoxiang Huang; Milena Radenkovic; Ge Chen

https://ieeexplore.ieee.org/document/10572225

Ocean eddies significantly affect marine ecosystems and climate by transporting essential substances through the ocean. This study addresses global-scale eddy detection via pixel-by-pixel classification and generates a global eddy classification map with a resolution of 720 × 1440, improving data scale and generality. A high-precision attention residual U²-Net model, ARU²-Net, is proposed for mining eddy surface features from sea level anomaly (SLA) and sea surface temperature (SST) data in the global ocean. Experimental results demonstrate the effectiveness of the proposed approach on a global eddy dataset, achieving strong test performance and improving detection in several regions.

ARU²-Net architecture; attention mechanism; deep learning; global ocean eddy; intelligent detection

Časopis

IEEE Transactions on Geoscience and Remote Sensing (TGRS)

Intelligent Sparse2Dense Profile Reconstruction for Predicting Global Subsurface Chlorophyll Maxima

Yongjun Yu; Baoxiang Huang; Milena Radenkovic; Tingting Wang; Ge Chen

https://ieeexplore.ieee.org/document/10684819

Subsurface chlorophyll maxima (SCM) are crucial ecological indicators in marine ecosystems. Although biogeochemical Argo assimilation has produced promising results, sparse observations limit effective operational use. This article proposes an AT-GRU deep learning model based on gated recurrent units with an attention mechanism to reconstruct biogeochemical-Argo (BGC-Argo) profiles from sparse to dense (Sparse2Dense) and improve the estimation of subsurface chlorophyll-a concentration. The methodology builds a dataset from satellite remote sensing data and associated BGC-Argo profiles, reconstructs chlorophyll profiles from 1 to 300 m, and analyzes SCM characteristics across multiple regions. Extensive experiments demonstrate improved performance (including strong R-squared values), supporting the feasibility of the approach for global applications.

bidirectional gated recurrent unit (Bi-GRU); biogeochemical-Argo (BGC-Argo); remote sensing; residual connection module; self-attention; sparse to dense (Sparse2Dense); subsurface chlorophyll maxima

Časopis

IEEE Transactions on Geoscience and Remote Sensing (TGRS)

Global Oceanic Mesoscale Eddies Trajectories Prediction With Knowledge-Fused Neural Network

Xinmin Zhang; Baoxiang Huang; Ge Chen; Linyao Ge; Milena Radenkovic; Guojia Hou

https://ieeexplore.ieee.org/document/10497608

Efficient eddy trajectory prediction driven by multi-information fusion can facilitate oceanographic research, but complex dynamic mechanisms make the problem challenging. This article proposes a knowledge-fused neural network (EddyTPNet) inspired by ocean dynamics and guided by prior physical knowledge. The approach includes analysis of trajectory kinematics, decoupling longitude and latitude prediction, embedding directional dispersion prior knowledge into the decoder, and transfer learning for adaptation to local regions. Extensive experiments show that EddyTPNet can reliably forecast global eddy motion for the next seven days with low daily mean geodetic error, offering valuable insights for knowledge-based time-series neural networks in oceanography.

deep learning; directional divergence physical information; eddy trajectory prediction; knowledge-fused neural network

Časopis

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS)

Mesoscale Westward Eddy Trajectory Prediction With Memory Augmented Neural Network

Wanchuan Kan; Baoxiang Huang; Milena Radenkovic; Xinmin Zhang; Ge Chen

https://ieeexplore.ieee.org/document/10759624

Accurate prediction of oceanic eddy trajectories is crucial for monitoring ocean climate change, but complex dynamics and varying environmental effects make the problem difficult. This article proposes a parsimonious and interpretable network with external Rossby-wave memory for westward mesoscale eddy trajectory prediction. The method uses multilayer perceptrons (MLPs) for cross-variable feature extraction, gate recurrent units (GRUs) for temporal correction, and an external memory unit to store phase speed information of long Rossby waves across scales. Experimental results on benchmark datasets show that the proposed method outperforms baseline methods in both accuracy and computational efficiency.

eddy trajectory prediction; external memory; gate recurrent unit (GRU); multilayer perceptron (MLP); Rossby wave

Časopis

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS)

Long-Term Forecasting of Long-Lived Oceanic Eddies Using an Improved Informer Framework

Yimin Wang; Baoxiang Huang; Milena Radenkovic; Wanchuan Kan; Xinmin Zhang; Ge Chen

https://ieeexplore.ieee.org/document/11146589

Long-term prediction of long-lived mesoscale oceanic eddies is important for ocean dynamics research because prediction accuracy affects climate-change assessment, ocean energy transport analysis, and marine resource management. Due to the nonlinear and multiscale nature of vortex dynamics, this remains challenging. This paper proposes an Informer-based deep learning framework enhanced with two innovations: an adaptive feature selection module for dynamic prioritization of input features and a geospatial loss function based on the Haversine formula to optimize spatial accuracy. The study also analyzes the influence of dipole eddies and lifecycle stages on trajectory predictability. Experiments show that the proposed method outperforms seven baseline models and reduces error accumulation in long-term forecasting settings.

deep learning; dipole eddy interaction; feature selection; long-term trajectory prediction; ocean dynamics

Časopis

Deep Sea Research Part I: Oceanographic Research Papers

Deep Blue AI: A New Bridge from Data to Knowledge for the Ocean Science

Chen, G.; Huang, B.; Chen, X.; Ge, L.; Radenkovic, M.; Ma, Y.

https://www.sciencedirect.com/science/article/pii/S0967063722001984

The global ocean is the largest ecosystem on our planet, the scientific understanding of which is the first step toward the sustainable development of the perpetual ocean. With the continuous advance of observation technology, ocean science has entered the era of big data, which significantly facilitates its junction with artificial intelligence (AI). This paper addresses whether AI can serve as a new bridge from oceanic data to knowledge by summarizing a large number of investigations and implementing application examples. First, AI frameworks for feature engineering, classification and detection, and time-series forecasting in emerging AI oceanography are systematically discussed. Then, several case studies across different ocean depths demonstrate the effectiveness of these data-to-knowledge bridges, including ocean current reconstruction, Arctic sea ice prediction, oceanic eddy identification with profiling data, and deep-sea debris detection for ocean ecology. Finally, recommendations are proposed to strengthen ocean AI in typical application areas. The study emphasizes deep blue AI (DBAI) as a potentially universal methodology and highlights the urgent need to explore its capabilities for AI-based knowledge discovery in oceanography.

Perpetual ocean, big ocean data, artificial intelligence, knowledge discovery, feature engineering, classification and detection, time series forecasting

Časopis

Remote Sensing of Environment

Deep learning solver unites SDGSAT-1 observations and Navier–Stokes theory for oceanic vortex streets

Gao, H.; Huang, B.; Chen, G.; Xia, L.; Radenkovic, M.

https://www.sciencedirect.com/science/article/pii/S0034425724004516

The world’s first scientific satellite for sustainable development goals (SDGSAT-1) provides valuable data about offshore small-scale ocean phenomena, including the Karman vortex street phenomenon. Traditional simulation of oceanic vortex streets relies on strong theoretical assumptions of the Navier–Stokes equations. This paper proposes a self-supervised neural network with high generalization ability to implement Navier–Stokes equations and simulate realistic oceanic vortex streets. A physics-informed convolutional neural network is first used to determine the corresponding pressure and velocity fields, achieving accurate simulation with lower computational cost. Observational islands in SDGSAT-1 imagery are embedded as obstacles, while the marine background field (including wind and terrain) is incorporated to produce more realistic simulations than traditional methods. Morphological parameters of oceanic vortex streets are then calculated and analyzed to improve understanding of small-scale vortex street phenomena. The results demonstrate promising efficiency and suggest that combining observation and theory with deep learning PDE solvers can accelerate the understanding of oceanic phenomena.

SDGSAT-1, Karman vortex street, Navier–Stokes equations, deep learning

Časopis

Journal of Sensor and Actuator Networks (J. Sens. Actuator Netw.)

Resistance to Cybersecurity Attacks in a Novel Network for Autonomous Vehicles

Brocklehurst, C.; Radenkovic, M.

https://www.mdpi.com/2224-2708/11/3/35

The increased interest in autonomous vehicles has led to the development of novel networking protocols in VANETs. In this widespread safety-critical application, security is paramount. The paper examines autonomous vehicle edge networks as opportunistic networks bridging fully distributed vehicle-to-vehicle communication and cellular-based centralized infrastructure. Experiments use opportunistic networking protocols to provide data to autonomous trams and buses in a smart city, while attacking vehicles attempt to disrupt the network. By varying the number of vehicles and attack duration, the study analyzes impacts on delivery probability, delivery time, and computation expense. The results show that MaxProp performs strongly in many categories under normal conditions, while PRoPHET provides more reliable delivery probability under attack. Two attack types (black hole and grey hole) are evaluated, and the multidimensional analysis (including geospatial elements) demonstrates important tradeoffs in protocol resilience.

VANETs, opportunistic networks, security

Časopis