This study evaluated the performance of two portable spectral devices (visible: Vis; near-infrared: NIR) and one benchtop combined Vis/NIR instrument for predicting beef biochemical traits related to meat quality. A total of 497 muscle samples were collected from 55 fattening farms across five European countries. Reference values were obtained exclusively from pooled samples. For the benchtop device, two sampling strategies were compared: 1) direct scanning of pooled samples (n = 99) used for reference analyses, and 2) scanning individual samples followed by spectral averaging to match pooled references. Partial least squares regression models were developed using several preprocessing methods. For total and heme iron, predictive performance was also evaluated by separately analyzing the visible region of the benchtop device. Results showed that the benchtop Vis/NIR instrument clearly outperformed the portable devices, providing higher prediction accuracies for all traits across the full spectral range. Compared with portable instruments, the benchtop device achieved superior performance for proximate composition (R-2 = 0.51-0.85 vs 0.10-0.22), major fatty acids (R-2 = 0.50-0.68 vs 0.01-0.31), and total and heme iron (R-2 = 0.79-0.93 vs 0.46-0.91). Moreover, scanning individual samples and averaging spectra yielded better prediction models than direct scanning of pooled samples, with good accuracies for moisture, protein, fat content, major saturated and monounsaturated fatty acids, and iron. Conversely, poor prediction (R-2 < 0.40) was observed for polyunsaturated fatty acids, TBARS, carnosine, anserine, zinc, selenium and B-group vitamins. Overall, Vis/NIR spectroscopy showed strong potential for beef quality assessment, although further optimization is required for portable devices.
Convolutional neural networks (CNNs) are widely used for hyperspectral imaging due to their ability to learn joint spatial-spectral representations. However, conventional 2D and 3D architectures are often based on striding or pooling operations that implicitly perform subsampling, discarding spatial-spectral features, and potentially introducing aliasing artifacts. This effect might hinder the accuracy and reliability of hyperspectral semantic segmentation procedures, in which fine-grained spatial and spectral signals are critical for accurately defining boundaries. To address this limitation, this work introduces 3D-SCube, a spatial-spectral rearrangement block designed for 3D convolutional architectures. The proposed block replaces strided downsampling operations by reorganizing its intermediate feature maps into multiple non-overlapping spatial-spectral sub-cubes, thereby preserving all feature-map values while reducing spatial-spectral resolution through channel rearrangement. 3D-SCube is parameter-free, architecture-agnostic, and can be seamlessly integrated into existing 3D CNNs without modifying their overall structure. Experimental results on the AeroRIT urban dataset and the Shenzhou UAV agricultural dataset show that replacing standard strided convolutions in a 3D ResNet-6 backbone with 3D-SCube consistently improves dense segmentation, especially under limited capacity and scarce supervision. On AeroRIT, 3D-SCube improves mIoU from 0.651 for the 2D ResNet-6 baseline and 0.769 for the standard 3D ResNet-6 baseline to 0.780, while reaching an OA of 0.921. Under severe label scarcity, the proposed model maintains approximately 0.70 mIoU in the 10% training setting, whereas the 2D baseline drops to approximately 0.42 mIoU. The block is parameter-free and introduces moderate deployment overhead, while remaining competitive with representative Transformer- and Mamba/SSM-based dense segmentation baselines.
A novel cBN-hardened B-12(C,Si,B)(3)-SiC composite was successfully fabricated by reactive spark plasma sintering (SPS), and its unlubricated sliding wear performance was evaluated. Powder mixtures with varying B4C:Si:cBN ratios (vol%) were consolidated under different SPS conditions, and the resulting materials were characterised microstructurally and mechanically to identify suitable processing conditions for obtaining a dense (>98%) fine-grained (<1 mu m) cBN-hardened (similar to 33.4 GPa) B-12(C,Si,B)(3)-SiC composite-namely, 72B(4)C:18Si:10cBN (vol%) SPSed at 1500 degrees C, with a heating rate of 200 degrees C/min, a dwell time of 15 min, and an applied pressure of 75 MPa. This novel composite was then tested tribologically against diamond at a 40 N load under unlubricated sliding conditions, demonstrating a threefold improvement in wear resistance (similar to(2.7 +/- 0.3)& centerdot;10(7)(N & centerdot;m)/mm(3)) relative to that of its cBN-free reference counterpart-with very mild abrasion and no evidence of plastic grooving or grain pull-out. This enhanced performance results from the cBN third phase increasing the overall hardness of the composite and promoting the formation of a more protective and lubricating oxide tribolayer. Incorporating cBN into ceramic microstructures could thus be an effective strategy to enhance hardness and tribological performance, provided that sintering conditions are sufficiently mild to prevent its transformation into hBN.
Accurate speed and range estimation is critical in modern radar and sonar systems operating under challenging conditions such as high target velocities, low Signal-to-Noise Ratios (SNRs), and Doppler-induced distortions. Traditional linear frequency-modulated (LFM) signals suffer from range-Doppler coupling and reduced correlation performance when subjected to substantial Doppler shifts. This paper presents a comprehensive analytical and simulation-based study of Dual Hyperbolic Frequency-Modulated (DHFM) signals, which combine upward and downward frequency sweeps to enable simultaneous and decoupled estimation of target speed and range. We derive closed-form expressions for the matched filter response of DHFM waveforms under Doppler scaling and develop explicit formulas for speed and range extraction from the time separation of correlation peaks. Through extensive software simulations of an airborne sonar system, we systematically evaluate the impact of three key parameters: the sweep rate (k), Signal-to-Noise Ratio (SNR), and propagation-induced attenuation. Our results demonstrate that speed estimation errors drop below 0.05 m/s for k > 5000 across target speeds of up to 100 m/s, while range estimation remains below 0.2 m accuracy even at SNR levels as low as-15 dB. Distance attenuation analysis under ISO 9613-1 atmospheric absorption models shows that both estimators remain robust up to 200 m range with moderate k values. The findings establish DHFM signals as a superior alternative to conventional LFM waveforms for applications requiring high-accuracy parameter estimation in Doppler-sensitive environments, with particular relevance to airborne and underwater sonar systems.
Therapeutic resistance in cancer is a critical problem that often requires simultaneously targeting multiple genes or pathways. Given the high costs of drug development, the exploration of existing approved drugs is gaining popularity in cancer research. Within this context, druggable gene networks, defined as sets of functionally related and frequently altered genes that are targeted by approved drugs, represent a promising framework for identifying potential therapeutic options. This search for druggable gene networks can be formulated as a multi-objective optimization problem where gene druggability, patient mutation coverage, and network correlation need to be simultaneously maximized. To address this challenge, we propose an Objective Space Partitioning Algorithm for Druggable GEne NEtwork Detection (OSPA-DGENED), which employs a set of reference points to balance convergence and diversity during the evolutionary process. Furthermore, our approach incorporates problem-specific operators that leverage biological information to produce high-quality solutions. We evaluated OSPA-DGENED’s performance against six state-of-the-art methods across 27 biological scenarios derived from real cancer data. OSPA-DGENED achieved statistically significant improvements over all compared methods in every experimental scenario. These results, along with the biological analysis of the solutions, highlight the effectiveness of the proposed approach in detecting high-quality and biologically relevant druggable gene networks.
