Wildfires pose significant challenges globally, including in Iran. This study analyzes wildfire occurrences in Iran from 2001 to 2022, using NASA FIRMS' active fire detections MCD14DL data. To enhance the reliability of this satellite-based dataset, particularly in data-scarce regions like Iran, we applied a multi-sensor cross-validation framework before modeling. We also aim to examine the country's wildfire dynamics over two decades, employing k-means clustering to categorize wildfires into ten clusters, delineating fire zones. Two random forest regression models explore the relationships between annual CO2 emissions, indicative of human activities, and average temperature, a proxy for climate variability, with wildfire occurrence. Our findings reveal a notable escalation in the frequency and intensity of wildfires across Iran during the study period. Specifically, the western and southwest regions, designated as Zone 05, emerge as highly affected areas, recording 162,734 fires despite their smaller size. The years 2015-2018 stand out as critical, marked by heightened wildfire activity and rapid annual fluctuations. Interestingly, the regression analysis shows a strong correlation between CO2 emissions and wildfire activity, which highlights the significant influence of human activities. In contrast, the weaker link with the average temperature suggests that climate variability plays a comparatively smaller role in shaping wildfire patterns in Iran during the study period . This study provides insights into Iran's wildfire patterns, revealing that the wildfire regime in Iran is evolving mainly through event frequency rather than fire intensity. These results emphasize the need for stakeholders to understand these dynamics thoroughly for effective mitigation strategies against the environmental and economic challenges posed by wildfires in the region.

It has been shown that early detection of Parkinson's disease (PD) using unobtrusive digital biomarkers can lead to better health outcomes. In this paper, an Edge-AI framework for PD screening in smart city environments is presented, emphasizing privacy, scalability, and explainability. For data acquisition and federated learning, the framework consists of (1) an edge-enabled architecture, (2) a hybrid pipeline, combining CNN-based feature extraction with XGBoost classification and score-level fusion, and (3) a SHAP-based explainability module for clinical interpretation. In this study, we conduct experiments using publicly available datasets for handwriting (HandPD), gait (PhysioNet Gait), and wearable sensors. Since there are no subject-aligned multimodal data available, we use a simulated late-fusion approach to evaluate the integration of multimodal data. This study presents a reproducible methodology, including preprocessing, model design, and cross-validation, and clearly distinguishes between simulated fusion and real multimodal evaluation. This framework provides the basis for future subject-level multimodal PD detection in real-world healthcare settings.

GenForge is an open-source Python package for interpretable symbolic modeling through multi-population genetic programming. It unifies regression, classification, and semantic feature partitioning into a single evolutionary learning framework. By integrating multi-gene symbolic regression, ensemble evolution, and Semantic-Preserving Feature Partitioning (SPFP), GenForge enables high-fidelity modeling while maintaining transparency and parsimony. The package provides modules for symbolic regression (gpregressor), classification (gpclassifier), and feature partitioning (SPFPPartitioner), each with reproducible example scripts and diagnostic visualization tools. GenForge supports reproducible research and educational use in explainable AI, symbolic learning, and multi-view ensemble modeling.

Wind energy plays a critical role in meeting global electricity demand and decarbonisation targets, accounting for a substantial and growing share of renewable power generation worldwide. However, the inherently intermittent, non-stationary, and spatially coupled nature of wind severely complicates the reliable integration of wind farms into power systems. In particular, short-term wind power forecasting at ultra-short (10-minute) and short (one-hour) horizons is essential for real-time dispatch, reserve allocation, and grid stability, yet remains challenging due to wake interactions, regime shifts in wind direction and speed, sensor noise in supervisory control and data acquisition, and dynamically evolving turbine-to-turbine dependencies. To address these challenges, a Physics-Informed directionality-aware Large Language Guided Dynamic Spatio-Temporal Graph Deep Model is proposed for short-term wind farm power forecasting. The proposed framework integrates physics-informed power-curve cleaning, dynamic graph construction guided by a large language model, and spatio-temporal graph convolutional networks to explicitly model regime-dependent, directed turbine interactions. Unlike conventional static or correlation-based graphs, the large language model infers context-conditioned adjacency structures that adapt to changing wind regimes while preserving interpretability and reproducibility through schema validation and fallback mechanisms. Extensive experiments are conducted on a large-scale, high-resolution dataset from a multi-turbine wind farm, benchmarking the proposed framework against 14 well-known recurrent deep learning, transformer-based, and hybrid ensemble models with two recent LLM-based forecasting baselines. For 10-minute-ahead forecasting, the proposed model (LLXSTG) substantially outperforms the JST-Transformer, achieving 66.5% and 67.9% reductions in mean absolute error and root mean squared error, respectively, together with a 19.3% improvement in R-value. When compared with the advanced ensemble learning baseline, the proposed model yields 32.6% lower mean absolute error, 26.3% lower root mean squared error, and a 4.1% increase in R-value, demonstrating clear gains even against strong non-transformer architectures. At the one-hour-ahead horizon, the proposed model maintains its superiority under increased forecast uncertainty, reducing average absolute error and root mean squared error by 55.2% and 54.2%, respectively, relative to JST-Transformer, while improving the R-value by 11.5%. Compared with the advanced ensemble, the proposed model achieves 35.0% lower average absolute error, 29.6% lower root mean squared error, and a 4.9% increase in correlation. These consistent improvements across both ultra-short and short-term horizons highlight the effectiveness of integrating causality awareness, large language-guided dynamic graph construction, and spatio-temporal learning for robust wind farm power forecasting.

Accurate segmentation of intermediate gastrointestinal bleeding in endoscopic images is essential for rapid diagnosis and treatment, yet remains challenging due to low contrast between bleeding regions and surrounding mucosal tissues. This paper introduces S2-KFCM (Double Spatial Kernelized Fuzzy C-Means), a novel framework that integrates fuzzy logic with deep learning for fine-grained identification of bleeding patterns. Our approach employs a coarse-to-fine segmentation strategy: the spatial kernelized fuzzy c-means algorithm first incorporates 22 spatial-intensity features and color space thresholding for preliminary localization, which is then refined by a DeepLabv3+ architecture with a ResNet50 backbone to produce precise segmentation maps. We conduct comprehensive comparisons against ten state-of-the-art models (2020–2026) spanning fuzzy clustering, deep learning, transformer-based, and knowledge-enhanced approaches. Experiments on the Kvasir-SEG dataset demonstrate that S2-KFCM achieves 98.95

A pioneering depth-specific water quality management method was developed and evaluated at Oman's largest and most strategic dam. Unlike conventional models that depend on costly deep-water sampling, this study introduces a novel surface-to-depth inference framework. We demonstrate a framework where surface measurements of water quality variables (WQVs) combined with geographical coordinates (latitude and longitude) were used to estimate dissolved oxygen (DO) and chlorophyll-a (Chl-a) at two intake-relevant depths within the reservoir. This novel method uses Bayesian optimization and four machine learning (ML) algorithms: Regression Trees (RT), Support Vector Regression (SVR), Gaussian Process Regression (GPR), and Ensemble of Trees (ET). The WQVs were comprehensively monitored at twenty sampling stations across the Wadi Dayqah reservoir, with measurements restricted to a maximum depth of 35 m. At depths of 20 and 25 m, which correspond to the installation locations of the inlet pipes, a stepwise regression analysis was conducted to identify the principal variables affecting DO and Chl-a. The performance of the GPR model in estimating Chl-a and DO at both depths was enhanced by reducing the number of input variables. For DO at a depth of 20 m, the GPR model yielded an Rsquared value of 0.88, an RMSE of 0.105, and an MAE of 0.070. Additionally, it demonstrated the highest level of accuracy in estimating Chl-a at the same depth, as supported by its R-squared value of 0.95 and the smallest RMSE value of 0.059. This study serves as a single-reservoir proof-of-concept; generalization of the proposed method requires further validation on additional water bodies with different morphometric characteristics. In addressing reservoir water quality concerns, the outcomes of this study may offer a fresh perspective on implementing water quality management strategies and interventions.

This paper introduces Multi-population Ensemble Genetic Programming (MEGP), a computational intelligence framework that integrates cooperative coevolution and the multi-view learning paradigm to address classification challenges in high-dimensional and heterogeneous feature spaces. MEGP decomposes the input space into conditionally independent feature subsets, enabling multiple subpopulations to evolve in parallel while interacting through a dynamic ensemble-based fitness mechanism. Each individual encodes multiple genes whose outputs are aggregated via a differentiable softmax-based weighting layer, enhancing both model interpretability and adaptive decision fusion. A hybrid selection mechanism incorporating both isolated and ensemble-level fitness promotes inter-population cooperation while preserving intra-population diversity. This dual-level evolutionary dynamic facilitates structured search exploration and reduces premature convergence. Experimental evaluations across eight benchmark datasets show that MEGP consistently outperforms a baseline GP model in fitness, diversity, and generalization, although convergence-rate and runtime results are more mixed. Comprehensive statistical analyses validate significant improvements in Log-Loss, Precision, Recall, $F_{1}$ score, and AUC. MEGP also exhibits stronger diversity retention and improved fitness progression, although the current implementation increases runtime and is evaluated with two views only. By unifying population-based optimization, multi-view representation learning, and cooperative coevolution, MEGP contributes a structurally adaptive and interpretable framework that advances emerging directions in evolutionary machine learning.

A Water Quality Index (WQI) is a widely used tool for assessing water quality. This study examines two primary research ideas regarding WQI development for dam reservoirs: 1) The feasibility of implementing a singular weighting method across multiple reservoir points at varying times, depths, and locations, and 2) The influence of biological effects on physical indicators related to water quality. This paper presents the following innovative methodologies to tackle these challenges: 1) A Conductivity-Temperature-Depth (CTD) sensor was utilized to assess water quality indicators at 31,322 points within a dam reservoir, succeeded by the creation of an Optimized Variational Auto Encoder (OVAE) to develop a time-varying 3-dimensional (3D) (Latitude, Longitude, and Depth ) sensor-based digital twin of the reservoir, 2) Explainable AI (XAI) algorithms were utilized to ascertain the impact of various water quality indicators, and 3) A contribution-based weighting scheme was implemented to analyze the correlation between Photosynthetically Active Radiation (PAR) and physical parameters. Key findings indicate: 1) a singular weighting method is not applicable at different places within a reservoir across time, depth, and location, and 2) PAR, modeled with cost-effective parameters, consistently reflects the impact of light-dependent biological processes on overall water quality.

Breast cancer is a leading public health concern that demands improved strategies for early diagnosis and prognosis. In this study, we propose P3DE, a novel computational framework that combines deep ensemble learning with the Parameter-less Population Pyramid (P3) metaheuristic for the identification of breast cancer biomarkers from gene expression data. P3DE integrates multiple autoencoders with diverse activation functions and dynamically computes ensemble weights based on reconstruction performance, enabling robust and adaptive feature selection without manual parameter tuning. Applied to the GSE42568 dataset, P3DE achieved outstanding classification performance (97.22

Structural optimization is crucial in engineering design, aiming to determine the optimal configuration and sizing of structural elements while satisfying critical constraints. Although swarm intelligence optimization algorithms are among the most promising approaches for this purpose, they often suffer from a loss of population diversity, which significantly compromises their performance, particularly as problem dimensionality increases. This paper presents the Modified Social Spider Algorithm (MSSA), an enhanced swarm-intelligence algorithm for the scalable optimization of truss structures. MSSA employs a modified random walk search strategy with a memory mechanism to enhance both exploration and exploitation, while preserving population diversity. The performance of MSSA is evaluated using global optimization benchmark problems and six well-known truss optimization problems with 10, 18, 25, 72, 200, and 942 bars, covering a wide range of structural scales. The results are compared with those of several advanced optimization algorithms, and the nonparametric Friedman rank test is used to analyze performance. The results demonstrate that MSSA is highly competitive, particularly for large-scale truss structures, outperforming most compared algorithms in the reported experiments in terms of solution quality and robustness. Furthermore, MSSA effectively addresses complex, nonlinear optimization problems, producing efficient truss designs that achieve superior solution quality while satisfying all structural constraints.

The seismic assessment of monopile-supported offshore wind turbines (OWTs) under multi-hazard loading is challenging, as traditional simulations are computationally prohibitive for large-scale probabilistic studies. This study introduces a data-driven framework to develop efficient surrogate models for predicting seismic responses and Probabilistic Seismic Demand Models (PSDM). A comparative analysis of 16 machine learning and deep learning algorithms reveals that advanced deep learning models, particularly the Kolmogorov-Arnold Network (KAN) and Multi-Layer Perceptron (MLP), significantly outperform traditional techniques. The KAN model achieved the highest accuracy, producing a robust surrogate capable of nearinstant predictions. This drastically reduces the computational cost of seismic risk assessments, making large-scale probabilistic analyses feasible.

Breast cancer remains a global health challenge, and early detection is crucial for improving survival rates. However, traditional biomarker detection methods in machine learning face challenges such as high false positives, large gene datasets, and limited sample sizes. These challenges are exacerbated by the limitations of traditional differential evolution techniques, which struggle with scalability and effectiveness in high-dimensional, complex problems. This study introduces Ensemble-Based Logical Binary QANA (LBQANA_En), an improved differential evolution variant tailored for large-scale global optimization in gene expression analysis. By integrating six gene expression datasets, LBQANA_En overcomes the constraints of small sample sizes and the complexities inherent in gene expression data. Inspired by the quantum navigation of migratory birds and leveraging logical operators like XOR and OR, LBQANA_En outperforms other binary QANA variants in biomarker detection. The algorithm successfully identifies key biomarkers, including LPL, LEP, CD36, CDC20, TOP2A, and EZH2, significantly improving breast cancer detection accuracy and reducing false positives, achieving an impressive F1 score of 98.958%. These biomarkers provide critical insights into important pathways, such as AMPK and PPAR signaling, setting a new benchmark in computational biology and bioinformatics research, and paving the way for advancements in diagnostic techniques and medical science.

Segmenting peptic ulcer and bleeding lesions from endoscopic video frames is critical for timely clinical intervention but remains challenging due to the subtle intensity contrast between diseased tissue and healthy mucosa. This work presents HyDAM (Hybrid Dynamic Aggregation Model), a lightweight framework that combines flexible multi-source feature extraction with adaptive dynamic fusion. The pipeline comprises two stages: a replaceable feature superposition module that aggregates local texture, global color statistics, and edge cues via learnable attention weights, followed by a lightweight refinement network that focuses on automatically identified target regions. Extensive experiments compare HyDAM against ten state-of-the-art methods. On the Kvasir-SEG dataset, HyDAM achieves 99.45

The identification of unknown parameters for proton exchange memberane fuel cells (PEMFCs) using nature-inspired optimization algorithms has emerged as a significant field of research in recent years. In the present study, a novel approach is presented, namely the hybrid Gray Particle Cuckoo (GPC) algorithm based on the hybrid properties of the grey wolf optimizer (GWO), particle swarm optimization (PSO), and cuckoo search (CS) to address the identification problem associated with PEMFCs. The effectiveness of the proposed GPC algorithm is evaluated on four commercially available PEMFCs (BCS500-W, Ballard Mark V, Temasek, as well as NedStack PS6). The fitness function has been expressed as the sum of the squared errors (SSE) that occurred between the estimated voltage and the data that corresponded to it. To further validate the model of the PEMFC, it is contrasted with other complex algorithms. The GPC algorithm showed the lowest SSE across all cases, resulting in SSE values of 0.011699, 0.813912, 2.267687, and 0.123276775 for the BCS500-W, Ballard Mark V, NedStack PS6 and Temasek PEMFC stack, respectively. Also, the PEMFC stacks are evaluated using different partial temperature and pressure conditions. In addition to real-world challenges, the GPC algorithm has been assessed on 100-digit CEC 2019 benchmarks and contrasted to other MH algorithms. Furthermore, both the parametric and non-parametric statistical tests are conducted to evaluate the efficacy of the GPC algorithm. The results in terms of mean square error (MSE), individual absolute error (IAE), mean bias error (MBE), mean absolute error (MAE), and root-mean-square error (RMSE) demonstrate that the GPC algorithm is the optimal choice contrasted to other algorithms due to its better solution quality and faster convergence time.

Neural contextual bandits are vulnerable to adversarial attacks, where subtle perturbations to rewards, actions, or contexts induce suboptimal decisions. We introduce AdvBandit, a black-box adaptive attack that formulates context poisoning as a continuous-armed bandit problem, enabling the attacker to jointly learn and exploit the victim's evolving policy. The attacker requires no access to the victim's internal parameters, reward function, or gradient information; instead, it constructs a surrogate model using a maximum-entropy inverse reinforcement learning module from observed context-action pairs and optimizes perturbations against this surrogate using projected gradient descent. An upper confidence bound-aware Gaussian process guides arm selection. An attack-budget control mechanism is also introduced to limit detection risk and overhead. We provide theoretical guarantees, including sublinear attacker regret and lower bounds on victim regret linear in the number of attacks. Experiments on three real-world datasets (Yelp, MovieLens, and Disin) against various victim contextual bandits demonstrate that our attack model achieves higher cumulative victim regret than state-of-the-art baselines.

Robot-assisted surgery has advanced significantly, offering benefits such as reduced blood loss, smaller incisions, faster recovery, and decreased physical burden on surgeons. However, external surgical tasks such as continuous wound suturing, tissue cutting, and laser-assisted tissue soldering remain challenging because they require precise, repetitive point-to-point tool maneuvering in visually unstructured environments. These tasks increase surgeons’ cognitive load, especially when performed through tele-operated robotic systems that rely heavily on manual control. Existing inverse kinematics (IK) methods often struggle to deliver fast, stable, and accurate convergence, limiting autonomous execution of such procedures. To address these challenges, this paper introduces a vision-guided point-to-path maneuvering framework driven by an Improved Levenberg Marquardt Inverse Kinematics (ILMIK) model. The proposed architecture integrates geometric forward kinematics, a finite-difference Jacobian approximation, and an optimization-driven ILMIK solver that improves convergence stability and accuracy while maintaining competitive computation time. Color-coded markers provide closed-loop vision feedback for real-time detection of tool and target positions. The ILMIK formulation employs simplified task-space and joint-space weighting matrices, enabling high convergence precision while maintaining low computation time and reducing iteration count than numerical baselines. Experimental validation on an artificial skin setup shows that ILMIK achieves the competitive accuracy, superior convergence in terms of iteration count, and robust performance across stitching trajectories, while maintaining low computation time comparable to the fastest baseline. Compared with state-of-the-art numerical and learning-based IK methods, ILMIK provides the best overall trade-off between accuracy, stability, and computational efficiency. The proposed framework demonstrates strong potential for supporting repetitive external surgical tasks including suturing, tissue-plane tracing, and laser-assisted bonding, thereby reducing surgeon workload and moving toward greater procedural autonomy.

Accurate land use and land cover (LULC) classification remains challenging when visually and spectrally similar classes coexist, particularly in RGB-based remote sensing imagery. While numerous machine learning (ML) and deep learning (DL) models report high overall accuracy, limited attention has been given to the spatial structure and persistence of misclassification errors. In order to gain more insight into model-dependent and systematic error patterns, this study introduces a misclassification-driven evaluation perspective through a controlled cross-architecture comparison of custom Convolutional Neural Networks (CNNs), AlexNet, and ResNet50 models applied to the EuroSAT dataset under a fixed RGB-only setting. This approach reduces input-related variation and allows a controlled comparison of model behavior across architectures. We demonstrate that models with similar overall accuracy can exhibit different class-wise error distributions. Across the evaluated architectures, misclassification patterns were stable for specific confusable class pairs (notably Annual Crop vs Permanent Crop), indicating that RGB-only inputs impose a dominant limitation that persists across model families. These results show that improving the reliability of land-use and land-cover classification requires more than changing models; it also requires richer feature representation and greater data diversity. The paper identifies several shortcomings that must guide the next generation of multispectral and domain-adaptive research and provides practical suggestions regarding the improvement of accuracy and robustness of LULC classification systems.