Artificial intelligence-powered health applications on smartphones and wearable devices are transforming consumer healthcare, enabling users to monitor skin conditions anytime and anywhere. However, these next-generation consumer technologies face a critical cyber risk: deep learning models are vulnerable to adversarial attacks. An attacker can inject tiny, imperceptible perturbations into skin images, causing the app to misdiagnose malignant moles as benign. This poses a direct safety threat to consumers. To mitigate this risk, we propose HRIDM, a Highly Robust Intelligent Diagnostic Model designed for secure deployment on consumer devices. Our framework has three parts. First, an Attack Defense Module creates a dual defense barrier at both data input and feature levels. It actively removes adversarial perturbations before they can affect the diagnosis. Second, a Contour Feature Extraction Module guides the network to focus on stable geometric features like lesion borders, rather than fragile textures that attackers can easily manipulate. Third, a Squeeze-and-Excitation Residual Network performs the final classification using these defended and enhanced features.We evaluate HRIDM on the ISIC 2019 skin disease dataset under strong adversarial attacks. Results show that at attack strength ϵ=0.0001, HRIDM achieves 22.77% higher accuracy, 7.31% higher AUC, 22.87% higher weighted precision, and 26.10% higher F1-score compared to mainstream baselines. Unlike standard models that collapse under attack, HRIDM maintains reliable performance while having reasonable computational cost for mobile deployment. This work directly addresses the cyber risk in consumer health technologies, providing a pathway toward secure and trustworthy AI-powered skin diagnosis for everyday users.

Federated learning (FL) has emerged as a promising paradigm for privacy-preserving collaborative training. However, under the intertwined challenges of severe data scarcity (e.g., 1\% sampling rate) and non-IID domain skew, existing FL approaches often experience substantial performance degradation. To address this issue, we introduce FedSHAPE (Federated Sharpness-aware Harmonized Aggregation and Parameter Equalization), a principled framework combining SPO and HAE. By systematically investigating the performance bottlenecks in such restrictive settings, we identify two coupled factors: standard local optimizers rapidly converge into domain-specific sharp minima (severe overfitting), rendering subsequent global feature alignment biased and suboptimal. Accordingly, FedSHAPE integrates a Stable Parameter Optimization (SPO) module for local flatness seeking with a Harmonized Aggregation and Equalization (HAE) module for robust consistency filtering and fair feature alignment. Extensive experiments demonstrate that FedSHAPE consistently improves both generalization robustness and cross-domain fairness under highly restricted data conditions.

Remote sensing image scene classification faces fundamental data challenges-such as the scarcity of labeled samples and severe class imbalance-that substantially constrain the performance of traditional deep learning methods. Although several recent reviews have documented isolated advances in generative modeling, they remain fragmented and lack a systematic examination of the paradigm shift from adversarial generation to iterative refinement mechanisms. Following the preferred reporting items for systematic reviews and meta-analyses methodology, we conducted a comprehensive review of 138 studies published between 2016 and September 2025, including 112 based on generative adversarial networks (GANs) and 26 based on generative diffusion models, to assess their applications in remote sensing image scene classification. Our analysis reveals key evolutionary trends in this field: GANs, which pioneered adversarial one-shot generation, have undergone three distinct stages-initial exploration (2016-2018), rapiddevelopment (2019-2021), and mature application (2022-2025). They have demonstrated strong performance in data augmentation, feature extraction, and direct classification tasks; however, persistent issues, such as training instability and mode collapse remain unresolved. Since 2023, generative diffusion models have rapidly gained prominence, offering remarkable stability and image quality through iterative denoising, particularly in complex feature learning and superresolution enhancement. Nonetheless, their computational complexity imposes significant costs, posing challenges for large-scale deployment. Our analysis of existing datasets reveals a concerning "generalization gap": approximately 78% of the studies rely on three benchmark datasets (Indian Pines, Pavia University, and Salinas Scene), which introduces a risk of overfitting to specific geographic scenes, while hyperspectral remote sensing images remains dominant. Future research should focus on lightweight model architectures, cross-modal feature fusion, and improved capabilities for low-resolution image processing, while also promoting the development of more diverse benchmark datasets to support broader applications in remote sensing scene classification. This review provides a critical comparative analysis and outlines feasible directions for future advancement, contributing to the development of efficient and robust generative models for global-scale remote sensing applications.

Sap flow can provide a critical foundation for water resource management, precision irrigation, and assessment of tree health. However, the generalisability and robustness of current sap flow prediction models remain limited. In this study, we propose a novel deep learning framework that integrates copula-entropy based feature selection with a CNN-FA-BiGRU architecture for sap flow prediction using environmental variables. The objective of this study was to predict sap flow at a future time step using historical environmental factors. Following data preprocessing, a novel feature selection method was proposed based on the assessment between reconstructed sap flow and environmental variables. Subsequently, a novel deep learning framework was developed by integrating a feature-attention mechanism into a combined CNN-BiGRU architecture establish the sap flow prediction model. Environmental factors and eight sap flow datasets were collected the SAPFLUXNET global database. Results show that the proposed method achieves the lowest mean absolute error, root mean square error, weighted mean absolute percentage error, and symmetric mean absolute centage error values and the highest coefficient of determination (R2 = 0.8809 +/- 0.0217) among 19 competing models, including four diffusion residual models, three traditional methods, and 11 deep learning-based models developed in this study. The R2 of the model using features selected by the proposed CE-based method exceeds that obtained using four other feature selection methods. Furthermore, the analysis reveals that the temporal patterns of couple entropy between environmental variables and sap flow are highly consistent within the tree species but exhibit notable variation across different species. In addition, the stability of sap flow in trees different ages within the same species and the hysteresis between sap flow and environmental factors analysed, and both were found to be significantly correlated with tree age and species. Overall, the proposed model exhibits good generalisability and robustness, making it suitable for sap flow prediction across tree species of varying ages.

Sixth-generation (6G) networks are projected to include massive machine-type communication (mMTC), ultra-reliable low-latency communications (URLLC), and enhanced mobile broadband (eMBB) to offer immersive, diverse, and latency-critical services. These heterogeneous service requirements demand intelligent and autonomous slicing mechanisms that remain reliable under dynamic traffic, multimodal context variations, and cross-layer uncertainties. Existing machine learning (ML) approaches achieve strong predictive performance but lack built-in Service-Level Agreement (SLA) validation, while rule-based strategies ensure constraint adherence yet struggle with rapidly varying 6G conditions. To address these limitations, this article proposes an Agentic Intelligence-assisted ML Fusion Framework that combines ensemble prediction with autonomous agentic reasoning.For example, an agentic controller can independently override the choice and reallocate resources to avoid service failure if an ML model predicts a slice appropriate for URLLC traffic but the expected latency exceeds the SLA level. A stacked ensemble of Light Gradient Boosting Machine (LightGBM), Random Forest (RF), and Logistic Regression (LR) captures nonlinear and multimodal dependencies in 6G beamforming data, whereas an agentic overlay enforces throughput, latency, and energy constraints in real time. Evaluation on a public 6G IoT beamforming dataset demonstrates near-perfect accuracy (99.9%) with consistent SLA compliance, improved interpretability, and enhanced stability under fluctuating service demands. The hybrid design provides a scalable foundation for multimodal, semantic-aware, and multi-agent orchestration in next-generation 6G slicing ecosystems.

Implementing internal tree hollow detection using the Internet of Things (IoT) technology is a crucial method for forestry conservation. Current methodologies primarily rely on stress wave sensors for such inspections. However, the inherently low signal acquisition density of these sensors leads to significant discrepancies between reconstructed wave velocity tomography and conventional optical imaging principles. This sparse signal distribution severely limits the ability of existing image processing algorithms to resolve internal hollow features, creating a bottleneck of insufficient precision in hollow localization and dimensional estimation for detection systems. To address the aforementioned problems, the research team independently developed a sensor for detecting internal tree defects and proposed a transformer-based model (RCE-DETR) for internal tree hollow detection. The sensor comprises three core modules: stress wave signal detection probes, a signal processor, and a display module. The proposed model in this study is deployed in the signal processor. By replacing the original convolution in the transformer framework with receptive-field attention convolution (RFAConv), the model retains important feature information, effectively fuses signal features, and further improves the accuracy of defect detection. Additionally, the model incorporates cascaded group attention (CGA) and efficient multiscale attention (EMA) to address the difficulty that the existing methods have in precisely determining the size of internal tree defects. Experimental results demonstrate that compared with the existing models, the RCE-DETR model increases the mean average precision (mAP) by 10.5%, 10.9%, 5.2%, and 6.3%, respectively, when compared to the commonly used YOLOv11, YOLOv8, EfficientDet, and CenterNet++ models.

In recent years, data generation has grown exponentially, and Big Data has emerged as a propelling force in the development of various machine learning advances and Internet of Things devices. In this regard, the analytical and learning tools that transport data from several sources to a central cloud for processing, training, and storage enable the realization of the potential of Big Data. Nevertheless, since the data may contain sensitive information like banking account information, government information, and personal information, these traditional approaches often raise serious privacy concerns. To overcome such challenges, Federated Learning (FL) has emerged as a sub-field of machine learning that focuses on scenarios where several entities (commonly termed as clients) work together to train a model while maintaining the decentralization of their data. Although significant research efforts have been dedicated to this area, a comprehensive review focusing on FL within the realm of Big Data services is still lacking. This paper, therefore, emphasizes the use of FL in handling Big Data and related services, which provides a comprehensive review of the potential of FL in Big Data acquisition, storage, Big Data analytics, and further privacy preservation. Subsequently, the potential of FL in Big Data applications, such as smart city, smart healthcare, smart transportation, smart grid, and social media are also explored. The paper also highlights various projects related to FL for Big Data and discusses the challenges associated with such implementations. These discussions provide a direction for further research, encouraging the development of plausible solutions.

By 2050, global food production must increase by 50%. Accurate farmland mapping is essential to meeting this target, yet the task remains technically challenging. Three persistent problems stand out: small and irregular field boundaries, dynamic crop phenology that alters spectral signatures, and the difficulty of fusing heterogeneous data sources. Existing reviews either address remote sensing broadly or focus on narrow algorithmic categories, leaving a gap in systematic treatment of segmentation-driven precision agriculture. This survey fills that gap by tracing the methodological evolution of farmland boundary extraction from traditional feature engineering to modern deep learning and foundation models, and by demonstrating how precise spatial delineation enables critical downstream applications. We introduce a 4D taxonomy that organizes methods along four dimensions: technical approach, data modality, spatial scale, and application use case. Our analysis covers 117 papers published between 2018 and 2026 across three developmental stages: traditional techniques, deep learning, and foundation models. From this review, three bottlenecks become clear. Even the most advanced models suffer substantial performance degradation in highly fragmented landscapes. Label scarcity remains a pervasive barrier to generalization. The complexity of multimodal data fusion also continues to resist robust solutions. To address these challenges, we outline a forward-looking roadmap built on three emerging paradigms. These include agricultural foundation models for self-supervised representation learning, physics-informed AI for trustworthy predictions, and federated learning for privacy-preserving cross-institutional collaboration. Together, these directions point toward reliable, scalable farmland mapping systems for next-generation precision agriculture.

Consumer-grade Agricultural AIoT (Agri-AIoT) systems increasingly rely on cloud-based intelligence, which introduces latency, privacy, and connectivity limitations. These limitations are particularly severe for heterogeneous consumer devices operating under strict cost, energy, and computational constraints. This review advocates a shift from cloud-centric architectures toward distributed, edge-centric intelligence. We examine lightweight model compression techniques, including pruning, quantization, and knowledge distillation, that enable real-time and on-device decision-making. We further analyze security threats such as data poisoning and adversarial attacks that arise in decentralized agricultural systems. Privacy-preserving learning mechanisms, including Federated Learning, are discussed as key enablers of collaborative intelligence without raw data sharing. By integrating lightweight Artificial Intelligence techniques with AI-native networking principles, this paper provides a unified perspective on distributed intelligence in consumer Agri-AIoT ecosystems. We conclude that the convergence of these approaches is essential for building sustainable, secure, and self-adaptive consumer-grade agricultural electronics.

Rapid and precise semantic segmentation of flash floods, such as landslides and debris flows, is critical for emergency response and post-disaster assessment. In recent years, Unmanned Aerial Vehicle (UAV) swarms offer flexible large-area coverage but face two major challenges for collaborative model training: (i) highly non-Independent and Identically Distributed (non-IID) data distributions that degrade global accuracy, and (ii) prohibitive communication costs in mobile ad hoc networks. To address these challenges, we propose Federated Multi-Level Knowledge Distillation (FedMLKD), a novel framework that enables UAVs to exchange both logits and adaptively pooled intermediate features. This design preserves critical spatial structure while drastically reducing transmission overhead. A client quality–aware aggregation strategy further ensures robust global knowledge construction. Experiments on the Landslide segmentation dataset demonstrate that compared to state-of-the-art baselines, FedMLKD improves mIoU by 9.19% and the average Boundary F1-score by 85.54%, while reducing communication cost by nearly 200 times. These results highlight FedMLKD's ability to balance accuracy and efficiency, making it a practical candidate for real-time disaster mapping in bandwidth-constrained UAV swarms.

The evolution of Artificial Intelligence (AI) has reached a critical point, where agentic AI systems demonstrate strong capabilities in goal formulation and planning but remain difficult to deploy in real-world settings due to their limited grounding in physical environments. These limitations arise from the challenges of partial observability, actuation uncertainty, and strict resource constraints that characterize the physical world. This survey argues that the Artificial Intelligence of Things (AIoT) provides the necessary foundation to embed agentic intelligence into such environments by enabling continuous interaction between sensing, reasoning, and action. We analyze the synergy between goal-driven agentic AI and distributed AIoT infrastructures and present a unified taxonomy of AIoT-enabled agentic architectures, highlighting trade-offs across centralized, edge-native, and hybrid deployment models. The survey further examines key enabling technologies, including edge intelligence, semantic communication, digital twins, and trust mechanisms, and discusses how they integrate into cognitive control loops. Through representative applications in smart cities, industrial automation, healthcare, and energy systems, we show how this convergence moves automation beyond rule-based behavior toward context-aware autonomy. Finally, we identify open challenges related to long-horizon safety, resource-aware intelligence, and ethical governance, and outline research directions toward robust, trustworthy, and socially embedded autonomous systems.

The meteorological communication networks provide critical data support for agriculture and environmental monitoring. However, covert gradient-based attacks persistently inject subtle perturbations, threatening data integrity and increasing the operational overhead for network operators. To achieve proactive service assurance and security-aware network management, this paper proposes a data integrity monitoring mechanism as a managed network function, named EK-IGNN. Unlike traditional passive detection, EK-IGNN functions as an active security service. It first employs the Empirical Mode Decomposition Kalman Filter (EMD-KF) to extract high-fidelity attack fingerprints, which are then analyzed by an Intrinsic Graph Neural Network (IGNN). The IGNN model captures complex dependencies and adaptively amplifies weak attack features, enabling closed-loop network security management. Experimental results demonstrate that the proposed algorithm achieving an average improvement of 16.07% in accuracy and 15.27% in F1-score over state-of-the-art benchmarks.

Vegetation, as a critical ecological feature and irreplaceable material resource of the Earth's surface, plays a crucial role in environmental protection, resource assessment, and urban planning. The rapid advancement of remote sensing platforms and sensor technologies has facilitated the acquisition of high-resolution remote sensing imagery, providing excellent conditions for the detailed analysis of vegetation. However, vegetation segmentation in remote sensing imagery poses distinct challenges, including extreme scale variance, spectral ambiguity, and complex boundary characteristics. The performance of convolutional neural network-based methods in vegetation segmentation is constrained by their limited ability to model long-range spatial dependencies. In contrast, although Transformer-based architectures are effective at capturing global context, their quadratic complexity makes them impractical for processing high-resolution remote sensing images. Recently, state-space models (SSMs) have emerged as a promising alternative due to their linear complexity and efficient long-range modeling capabilities. Nevertheless, existing vision SSMs have two critical limitations: 1) insufficient modeling of boundary information and 2) the limited performance improvements offered by multidirectional scanning strategies while increasing computational cost. To address these limitations, we propose BRSMamba, a boundary-aware network that integrates two novel modules with noncausal state-space duality (NC-SSD) to enhance both boundary preservation and global context modeling. Specifically, the boundary-subject fusion perception module extracts robust boundary features via a Laplacian-of-Gaussian convolutional block and fuses them with category-centric semantics to generate a boundary-subject map. This map then directs the boundary-body resolution module to inject boundary awareness into the NC-SSD state-transition matrix, enabling selective scanning that preserves fine details and captures long-range dependencies. Experiments on four benchmark datasets demonstrated that BRSMamba achieved 78.86% mean intersection over union on the Vaihingen dataset, 82.84% on the Potsdam dataset, 54.37% on the LoveDA dataset, and 72.24% on the GID dataset while using only 10.61M parameters and 26.28G floating-point operations. These results validate the effectiveness of the proposed method in balancing the accuracy and efficiency of complex vegetation segmentation tasks.

Precipitation nowcasting aims to predict short-term precipitation evolution over forecast lead times of 1–6 h and can support hydrological-risk and disaster-prevention applications when near-real-time observations are available. However, precipitation forecasting remains challenging because of rapid spatiotemporal evolution, spatial displacement, and the difficulty of representing localized high-intensity precipitation. To address these issues, this study proposes STAMP-GAN, a spatiotemporal attention-modulated generative adversarial network for regional precipitation sequence prediction. STAMP-GAN combines an AM-ConvLSTM temporal evolution module with spatial attention, efficient channel attention, large-receptive-field context modeling, and temporal-index-conditioned feature modulation. A spatially aligned two-dimensional digital elevation model (DEM) field is retained as static auxiliary geographical information. The STAMP-Net generator uses hierarchical multi-scale feature extraction to reconstruct precipitation structures at different spatial scales while a dual-branch temporal PatchGAN provides adversarial supervision for both the complete forecast sequence and the final three forecast frames. A hybrid objective combines regression, event-based, structural, temporal, and adversarial constraints. Experiments on the ERA5 and CMA-S datasets show that, compared with the best-performing baseline for each metric, STAMP-GAN achieves relative CSI improvements of approximately 6.5% and 9.5%, respectively. The proposed framework provides a data-driven approach for retrospective hourly regional precipitation sequence prediction under gridded meteorological-data conditions, rather than a fully validated operational real-time nowcasting system.

Human activity recognition (HAR) is vital to support day-to-day human activities and improve the interactions among users and external objects. HAR provides vital information about user personality and physiological traits, which can be analyzed through effective Artificial Intelligence (AI) based models. Different applications in computer vision and Machine Learning (ML), like video surveillance, healthcare, gesture recognition, person identification, and human-computer interaction, utilize the key principles of HAR. Recent surveys have centrally focused on the role of AI in HAR, where the computational requirements (both centralized and decentralized) are not discussed in detail. Owing to the gap, the article presents a detailed review of the fusion of AI and the cloud to address the dual objectives of model learning, resource requirements, and optimization. Based on research questions, we present a reference architecture that integrates AI and cloud for HAR in surveillance systems. A solution taxonomy is presented for diverse industrial verticals, and key challenges and future opportunities of HAR are discussed. A case study of Deep Learning (DL) based HAR is proposed for smart healthcare, where networking and security parameters are considered. The survey intends to assist industry and academia in the design of novel cloud and AI-based HAR systems with optimal control and resource management in various applicative verticals.

Night-time harvesting is an important strategy for improving the efficiency of agricultural automation; however, image degradation under low-light conditions severely limits the perception accuracy of machine vision systems. In complex orchard environments, existing methods often suffer from detail loss and structural distortion, which makes it difficult to meet the requirements of high-precision operations. To address these limitations, this study proposes a generative diffusion model for image enhancement built on a U-Net backbone. A Transformer module is incorporated to capture global context, and a convolutional block attention module is used to strengthen features of fruit as well as branches and leaves. In addition, pyramid-based resolution sampling and a global color correction module are employed to balance detail recovery with color consistency. Experimental results show that the proposed method achieves a PSNR of 27.1144 dB and an SSIM of 0.8891 on a synthetic apple dataset. To further assess practical applicability, a comparative analysis is conducted using both a synthetic dataset and a real low-light dataset. The results indicate that the model offers measurable generalization to real scenes and provides improved detail reconstruction. After enhancement by the proposed model, the YOLOv13 detector's mAP50-95 increases from 0.249 to 0.352. Overall, the proposed approach mitigates severe visual degradation under extreme lighting and provides technical support for all-day, high-precision automated agricultural operations.

Healthcare systems are increasingly data-rich but remain fragmented, reactive, and difficult to coordinate across institutions, devices, and clinical workflows. Healthcare 5.0 addresses these limitations through human-centricity, resilience, and symbiotic human-AI collaboration. This survey positions agentic AI, goal-oriented systems capable of contextual perception, persistent memory, multi-step planning, tool use, and bounded action, as the orchestration layer of Healthcare 5.0. We make four contributions: (i) a reference architecture that maps perception, reasoning, planning, action, and memory across device, edge, and cloud tiers; (ii) an autonomy-criticality taxonomy for four healthcare application domains; (iii) a structured analysis of safety, security, privacy, ethics, and regulatory challenges; and (iv) a phased roadmap for responsible clinical deployment. Our analysis shows that Healthcare 5.0 requires not only capable models, but also safety-governed, auditable, and ethically aligned agentic ecosystems.