Hydrographic research projects are often described as costly, data-intensive, high-stakes operations. This holds especially true for sub-bottom object detection where decisions regarding route planning, dredging or hazard assessment are often reliant upon sub-bottom profiler (SBP) data. Today, this data is commonly processed and interpreted manually. Whilst deep learning (DL)-based object detectors are increasingly being utilised to automate processing and interpretation, preparation of raw SEG-Y data for optimal exploitation by various models remains underexplored. Because proprietary subsurface datasets are scarce, to say that training DL models from scratch is impractical would be an understatement. Consequently, researchers must rely on transfer learning and models pre-trained on large datasets of out-of-domain general digital images. This reliance underscores the critical importance of the transformation techniques used to translate raw acoustic-response data to digital image pixels. This paper explores and evaluates several SEG-Y-to-digital image transformation techniques, i.e., scaling or normalisation techniques on real-world data in conjunction with two well-known object detection frameworks. Image normalisation is positioned not as just another preprocessing step, but as a critical experimental component of the SBP detection pipeline. Evaluations were conducted on a single-system SBP dataset, representing a specific acoustic frequency range and environmental boundary condition. It was demonstrated that contrast enhancement and outlier clipping increase downstream model convergence, yielding a more robust target recovery performance despite severe background clutter.
Software-Defined Networking (SDN) transforms traditional network design by decoupling the control and data planes, enabling centralized programmability, scalability, and dynamic reconfiguration. This paper presents a QoS-driven performance analysis of two popular SDN controllers, Ryu and ONOS, combined with Dijkstra and Multipath routing algorithms in scalable Fat-Tree topologies ( k = 4–10). Experiments were conducted in Mininet (OpenFlow 1.3) using iPerf with TCP and UDP traffic, each repeated 100 independent runs for statistical confidence. Key QoS metrics including throughput, jitter, round-trip time (RTT), packet loss ratio (PLR), CPU utilization, flow setup time, and scalability were evaluated. The results show that ONOS-Multipath achieved the highest average TCP throughput (up to 3.58 Gbps), with the lowest jitter, minimal PLR, and maximum host scalability (up to 250 hosts at k = 10). Compared with earlier POX- and Floodlight-based benchmarks, the ONOS-Multipath setup improves throughput by approximately 20 percent and reduces packet loss by up to 80 percent, confirming that the controller architecture critically shapes routing efficiency and that ONOS + Multipath provides the most robust and scalable SDN configuration for large-scale deployments.
Rice blast is one of the most destructive fungal diseases affecting rice cultivation, which causes large yield losses and creates a major risk to the world's food security. In addition to early detection, accurate severity-level assessment is crucial for effective disease management. A hybrid deep learning method based on ConvNeXt and Vision Transformer (ViT) models is proposed for assessing rice blast disease severity in the present work. A new well-organised dataset comprising 1,440 rice leaf images annotated into seven severity levels ranging from 1–10% to 61–70% based on the lesion area coverage is utilised in the model. The proposed hybrid ConvNeXt–ViT model has two parallel feature extraction branches, where ConvNeXt captures spatial and textural patterns, while ViT learns global contextual representations. Complementary feature representations from both branches are fused through a unified classification head to enable fine-grained severity prediction. A series of tests was carried out to evaluate the effect of different fusion strategies, pre-trained weights, patch sizes, and batch sizes on model performance. The proposed approach demonstrates consistent training behaviour and generalisation across different configurations. Comparative evaluation against state-of-the-art CNN models such as ResNet, EfficientNet, VGG, MobileNet, and Xception, as well as a transformer-based baseline shows that the proposed approach achieves superior performance. The ConvNeXt–ViT model achieves a classification accuracy of 92.63%, demonstrating its potential for automated rice blast severity assessment in agriculture applications.
The rapid expansion of the Internet of Things (IoT) has led to a proliferation of services offering identical functionalities with heterogeneous capabilities and quality levels. Efficiently composing these IoT services to satisfy user requirements while maintaining system performance under strict Quality of Service constraints, such as energy consumption, remains a challenging combinatorial optimization problem. Metaheuristic algorithms have proven effective in addressing such large-scale optimization tasks. This paper first proposes an energy consumption model for IoT services. It then proposes two novel discrete versions of the Slime Mould Algorithm (SMA) specifically designed for IoT service composition. The first, called Discrete Slime Mould Algorithm (DSMA), adapts SMA to the discrete domain using adaptive fitness-based weights, probabilistic gene copying from the best solution, and an adaptive mutation mechanism. The second, called Chaotic Discrete Slime Mould Algorithm (CDSMA), further integrates Chebyshev chaotic maps for population initialization and weight updating to improve population diversity and accelerate convergence. Experimental results show that CDSMA consistently outperforms DSMA and other classical metaheuristics. CDSMA achieves an optimality rate of 96.33% with a small population size and 99.92% with a moderate population size. Furthermore, extensive experiments over 30 independent runs confirm that both proposed algorithms significantly outperform classical metaheuristics. CDSMA achieves 96.46% (±0.65) optimality, DSMA achieves 93.75% (±0.75), significantly outperforming SMA (84.74%), GWO (82.54%), HHO (82.87%), and WOA (80.15%) with statistical significance (t-test, p < 10-20).
The number of individuals who live with limb loss has increased worldwide. Individuals with limb loss face significant daily challenges in independent mobility and functional activities. A primary technical barrier in modern active prostheses is the effective management of Ground Reaction Force (GRF), which introduces unpredictable disturbances during the gait cycle and complicates control due to nonlinear dynamics. This research addresses the need for a control framework that can simultaneously compensate for unknown GRF-related disturbances and manage nonlinear prosthetic dynamics. The study proposes an adaptive control framework for a 2-Degree-of-Freedom (DOF) lower-limb prosthesis, integrating a computed-torque-based adaptive controller with a generalized momentum observer for GRF estimation. To enhance performance, Particle Swarm Optimization (PSO) was employed to tune the controller and observer gains, replacing traditional trial-and-error methods. Simulations were conducted in MATLAB/SIMULINK using realistic disturbances derived from the GAITREC GRF dataset. Results demonstrate that PSO-based tuning significantly improved tracking accuracy, reducing Root Mean Square (RMS) tracking errors by 47% at the knee and 94% at the ankle compared to manual tuning. Additionally, GRF estimation errors were reduced by 57% and 48% for the knee and ankle, respectively. Comparative analysis shows that the proposed PSO approach provides superior global search capabilities and convergence stability over previous bio-inspired methods such as the Bat Algorithm(BA). While the framework offers a robust, data-independent solution for various locomotion scenarios, it is currently limited by its simplified 2-DOF model and lack of real-time validation. Future research should focus on implementing the controller in physical hardware to verify its performance and safety in real-world environments.
