Ali Ahmed

Routing a medical-delivery UAV to multiple hospitals must jointly account for distance, flight time, payload, and clinical priority. We present a reproducible GIS-based framework for multi-stop UAV emergency medical routing in Khartoum State, Sudan, using a common obstacle-aware 3-D leg model with grid-based A* routing around no-fly zones and payload-dependent climb, cruise, and descent. We show that a linear priority term in a composite tour cost is order-invariant and cannot distinguish visiting sequences. We therefore use a priority-weighted completion-time objective with hard deadlines for Critical hospitals, with criterion weights derived from health professionals using the Analytic Hierarchy Process. On a 14-hospital network, exact Held–Karp and label-setting dynamic programming provide proven optima for benchmarking classical, metaheuristic, and Advantage Actor–Critic sequencing methods (A2C-SP). The revised objective reduces optimal priority-weighted completion time by 14% with only a 0.2% increase in distance. A2C-SP is 14.7% above the optimum across 47 unseen instances while generating routes in approximately 2 ms. Without retraining, it also plans capacity- and range-feasible multi-sortie missions on a 59-hospital city network, achieving a priority-weighted completion time of 110 min versus 137 min for the published toolbox route. At city scale, a single UAV cannot serve more than 5 of 16 Critical hospitals within 15 min, highlighting fleet sizing as a key constraint for meeting such deadlines.

Hyperspectral image (HSI) classification is pivotal for remote sensing applications, including environmental monitoring, precision agriculture, and urban land-use analysis. However, its accuracy is often limited by scarce labeled data, class imbalance, and domain discrepancies between standard RGB and HSI imagery. Although recent deep learning approaches, such as 3D convolutional neural networks (3D-CNNs), transformers, and generative adversarial networks (GANs), show promise, they struggle with spectral fidelity, computational efficiency, and cross-domain adaptation in label-scarce scenarios. To address these challenges, we propose the Transformer–Graph Convolutional Network–Diffusion with Hybrid Domain Adaptation (TGDHTL) framework. This framework integrates domain-adaptive alignment of RGB and HSI data, efficient synthetic data generation, and multi-scale spectral–spatial modeling. Specifically, a lightweight transformer, guided by Maximum Mean Discrepancy (MMD) loss, aligns feature distributions across domains. A class-conditional diffusion model generates high-quality samples for underrepresented classes in only 15 inference steps, reducing labeled data needs by approximately 25% and computational costs by up to 80% compared to traditional 1000-step diffusion models. Additionally, a Multi-Scale Stripe Attention (MSSA) mechanism, combined with a Graph Convolutional Network (GCN), enhances pixel-level spatial coherence. Evaluated on six benchmark datasets including HJ-1A and WHU-OHS, TGDHTL consistently achieves high overall accuracy (e.g., 97.89% on University of Pavia) with just 11.9 GFLOPs, surpassing state-of-the-art methods. This framework provides a scalable, data-efficient solution for HSI classification under domain shifts and resource constraints.

Emotion recognition and sentiment analysis are crucial tasks in natural language processing, enabling machines to understand human emotions and opinions. However, the complex, nuanced relationship between emotions and sentiment in conversation poses significant challenges to accurate emotion recognition, as sentiment cues can be easily misinterpreted. Deploying emotion recognition and sentiment analysis tasks on edge devices poses substantial challenges due to computational resource constraints. We present an adaptive multitask learning approach that jointly leverages resource-constrained Mobile Bidirectional Encoder Representations from Transformers (MobileBERT) and Distilled BERT (DistilBERT) models to optimise emotion recognition and sentiment analysis. Our proposed approach utilises prototypical networks to learn effective representations of emotions and sentiment, while a focal weighted loss function effectively mitigates the class imbalance. We adaptively fine-tune the learning process to balance task importance and resource utilisation, resulting in better performance and efficiency. Our experimental results demonstrate the efficacy of our method, achieving the best results on MELD and IEMOCAP benchmark datasets while keeping a compact model size. Despite limited computational demands, our solution demonstrates that emotion and sentiment analysis can deliver performance comparable to resource-intensive large language models (LLMs). Facilitating various applications in human-computer interaction, affective computing, social media, dialogue conversion, and healthcare.

Recently, there have been efforts to create automated systems for diagnosing engine problems using sound detection. However, most of these methods lack robustness and interpretability, functioning as "black boxes" that make it difficult to understand their decision-making processes. The Learning Classifier System (LCS), a machine learning approach that operates using a set of rules, has demonstrated potential for providing robust, interpretable, and generalizable solutions across different domains. This work aims to develop a new LCS-based system for automatically detecting engine problems, with a focus on making its decision-making process understandable, contributing to explainable artificial intelligence. The system's performance is evaluated using features from the time domain, frequency domain, and time-frequency domain. Its robustness is tested with noisy sound data gathered under various normal and abnormal conditions. Experimental results show that this new approach outperforms conventional state-of-the-art methods by 2.6%−6.0%, achieving a maximum performance accuracy of 98.6%.

Memes, cultural entities that spread predominantly through social media, have recently seen an influx of harmful content used for trolling and bullying. Detecting such harmful memes presents a significant challenge due to the semantic disparity between their textual and visual components. To address this issue, we propose a Dual-Branch Neural Network (DBNN) for bridging the semantic gap in harmful meme detection. The proposed DBNN framework utilizes global visual and textual embeddings to learn discriminative joint representations. By capturing the context between the textual and visual modalities, DBNN bridges the semantic gap, enabling robust detection of harmful content.We evaluate the performance of the proposed DBNN framework on the Harm-C and Harm-P benchmark datasets, employing Accuracy and Macro-Averaged Mean Absolute Error (MMAE) as evaluation metrics. Our experiments demonstrate that DBNN significantly outperforms state-of-the-art approaches, achieving higher accuracy and robustness in detecting harmful memes. This work highlights the potential of DBNN in effectively identifying harmful memes, contributing to a safer online environment.

Machine learning algorithms face important implementation difficulties due to imbalanced learning since the Synthetic Minority Oversampling Technique (SMOTE) helps improve performance through the creation of new minority class examples in feature space before preprocessing. The underlying problem leading to performance deterioration emerges from noise and boundary instances found in the minority class rather than an excessive imbalance ratio. Noise and marginal samples become problematic during oversampling operations since they reduce the performance of classification models. This paper proposes an advanced version of Borderline-SMOTE (BL-SMOTE), called ABL-SMOT approach that utilizes INFFC for noise filtering to optimize data quality in the dataset. This improvement method for minority class sample placement and noise filtering ensures higher classifier performance through ABL-SMOTE. The proposed technique receives assessment through the usage of software defect datasets along with multiple testing classifiers against conventional data sampling procedures. The experimental outcomes show that ABL-SMOTE performs better than alternative approaches in most tested datasets because it is an effective preprocessing tool for imbalanced classification problems.

Hyperspectral (HS) and multispectral (MS) image fusion is the most favorable way to obtain a hyperspectral image that has high resolution in terms of spatial and spectral information. This fusion problem can be accomplished by formulating a mathematical model and solving it either analytically or iteratively. The mathematical solutions class has serious challenges, e.g., computation cost, manually tuning parameters, and the absence of imaging models that laboriously affect the fusion process. With the revolution of deep learning, the recent HS-MS image fusion techniques gained good outcomes by utilizing the power of the convolutional neural network (CNN) for feature extraction. Moreover, extracting intrinsic information, e.g., non-local spatial and global spectral features, is the most critical issue faced by deep learning methods. Therefore, this paper proposes an Extensive Feature-Inferring Deep Network (EFINet) with extensive-scale feature-interacting and global correlation refinement modules to improve the effectiveness of HS-MS image fusion. The proposed network retains the most vital information through the extensive-scale feature-interacting module in various feature scales. Moreover, the global semantic information is achieved by utilizing the global correlation refinement module. The proposed network is validated through rich experiments conducted on two popular datasets, the Houston and Chikusei datasets, and it attains good performance compared to the state-of-the-art HS-MS image fusion techniques.

The expansion of the tourism industry has led to the development of various methods to find optimal tourism paths. However, planning tourism paths in hilly areas remains complex and has specific challenges. Different algorithms have been used to plan tourism paths in flat and hilly terrains, including the traditional Ant Colony Optimization (ACO). Although widely used, this algorithm faces a number of limitations due to its slow implementation and pheromone update rules. This paper introduces a new approach to overcome these limitations. It presents a method for efficiently optimizing tourism paths in hilly areas based on an improved version of the ACO algorithm. The limitations of the traditional ACO and the Genetic Algorithm (GA) are addressed by improving pheromone updating techniques and implementing new initialization parameters. This approach provides a comprehensive and efficient method for planning hiking trails in hilly regions, considering dynamic tourism objectives such as temperature, atmospheric pressure, and health status. The proposed method is implemented to develop tourist routes in the hilly Jebel Marra region in Western Sudan. A comparison is provided between the effectiveness of this approach and the GA and traditional ACO algorithms. The advantage of the proposed approach is illustrated by results showing an optimization time of 0 points and 27 s compared to 0 points and 45 s and 0 points and 40 s for GA and ACO, respectively.

Content-based medical image retrieval (CBMIR), a specialized area within content-based image retrieval (CBIR), involves two main stages: feature extraction and retrieval ranking. The feature extraction stage is particularly crucial for developing an effective retrieval system with high performance. Lately, pre-trained deep convolutional neural networks (CNNs) have become the preferred tools for feature extraction due to their excellent performance and versatility, which includes the ability to be re-trained and adapt through transfer learning. Various pre-trained deep CNN models are employed as feature extraction tools in content-based medical image retrieval systems. Researchers have effectively used many such models either individually or in combined forms by merging feature vectors from several models. In this study, a method using multiple pre-trained deep CNNs for CBMIR is introduced, utilizing two popular models, ResNet-18 and GoogleNet, for extracting features. This method combines the feature vectors from both models in a way that selects the best model for each image based on the highest classification probability during training. The method's effectiveness is assessed using two well-known medical image datasets, Kvasir and PH2. The evaluation results show that the proposed method achieved average precision scores of 94.13% for Kvasir and 55.67% for PH2 at the top 10 cut-offs, surpassing some leading methods in this research area.