Pei-Jun Lee

The rapid electrification of maritime transport demands reliable emission monitoring systems to guarantee data integrity, regulatory compliance, and environmental sustainability. This study introduces a digital twin-based Internet of Things (IoT) framework that integrates hybrid machine learning models, including Support Vector Regression (SVR), Random Forest (RF), and Deep Neural Network (DNN), with the Isolation Forest (IF) anomaly detection algorithm for real-time ship emission assessment and cyberattack resilience. Through extensive experimentation, the proposed models achieved training and testing mean squared error (MSE) values of 0.0352–0.0387 and R² scores above 0.96, confirming high prediction accuracy in modeling complex emission behaviors. Under false data injection and adversarial perturbation scenarios, the hybrid approaches, including RF-IF, SVR-IF, and DNN-IF demonstrated superior anomaly detection accuracy ranging from 92.44% to 95.64%, outperforming state-of-the-art thresholding, isolation-based methods, Variational Autoencoder (VAE), and Long Short-Term Memory Autoencoder (LSTM-AE). The DNN-IF model exhibited the highest detection precision and resilience against cyber threats. These findings highlight the novel contribution of combining predictive modeling with anomaly detection in a digital twin environment, offering a secure and intelligent monitoring solution for next-generation low-emission marine vessels.

The fusion of low-resolution hyperspectral images and high-resolution multispectral image is a crucial approach to derive high-resolution hyperspectral images. This paper proposes a prior-guided dual-branch network that integrates $\boldsymbol{S}^{\mathbf{2}}$-attention Transformer to model both local and global dependencies while maintaining spectral consistency. SpatialSpectral Total Variation (SSTV) loss is introduced to improve reconstruction fidelity. Experimental results demonstrate that the proposed model achieves a 1.805 dB gain in PSNR and a 0.295 reduction in SAM, confirming its superior spatial detail reconstruction and spectral consistency.

Ensuring data security is a critical requirement in satellite-to-ground station communication systems. This paper proposes a novel quantum-inspired image encryption framework wherein images are represented via the Bloch sphere model to facilitate quantum operations. Furthermore, two distinct chaotic maps are integrated to perform pixel-level confusion and permutation, thereby significantly enhancing algorithmic complexity and security. Experimental results show that the proposed method requires only 0.6981 seconds to encrypt large-scale images while generating highly random cipher images. In the differential attack analysis, the algorithm achieves an NPCR (Number of Pixels Change Rate) of 0.9903 and a UACI (Unified Average Changing Intensity) of 0.3269, demonstrating excellent resistance against differential attacks. Furthermore, the decryption process enables lossless reconstruction of the original image, confirming the strong potential of the proposed encryption framework for secure satellite image transmission.

Cloud coverage significantly degrades the quality and usability of remote sensing imagery, while also leading to unnecessary bandwidth consumption and power expenditure in satellite payloads. To address this challenge, a high-efficiency hardware–software co-design framework for cloud removal in remote sensing applications is presented. At the algorithmic level, a two-stage model compression strategy is designed, combining structured channel pruning and adaptive unstructured pruning, which reduces model parameters by more than 98% while preserving segmentation accuracy across multisource datasets. At the hardware level, a low-voltage sparse matrix accelerator enhanced with a super balanced path mechanism is designed, enabling stable operation at 0.65 V and achieving 43.2% higher energy efficiency and 11.5% better area efficiency compared to prior designs. Extensive experiments on GF-1 WFV, Sentinel-2 CloudSEN12, and Lilium-1 imagery validate the generalization capability and robustness of the proposed approach. The joint optimization of algorithm and hardware not only delivers state-of-the-art efficiency but also provides a practical pathway for reducing noninformative data transmission and enhancing the sustainability of future satellite missions.

This paper proposes an optimized cloud segmentation accelerator to enable low data rate transmission in satellite payload applications. The system employs an advanced cloud detection and segmentation framework to distinguish cloud-covered and cloud-free regions in high resolution satellite imagery accurately. A quantized U-Net architecture has been developed with a focus on efficient resource utilization, tailored specifically for FPGA implementation. This design is built upon the H.265 video encoder, significantly reducing bandwidth usage while ensuring high-quality image output. Experimental results highlight impressive segmentation accuracy, achieving a mean Intersection over Union (mIoU) of 85.64%. The architecture is also notable for its low resource consumption, with only 0.27M parameters and 5.2 GFLOPs, along with minimal power requirements of just 0.16W. These attributes make this solution an ideal choice for edge deployment on satellite platforms.

In this study, a human action recognition approach with dynamical frame processing is proposed to fulfill the need for action recognition in a real-time manner. A novel architecture with a modified convolutional long short-term memory (ModConvLSTM) with pose heatmaps as input features is proposed to achieve human action recognition without a fixed frame number. The effects of ModConvLSTM are verified at different depths. In our proposed architecture, we replace global average pooling (GAP) with Bidirectional Encoder Representations from Transformers (BERT) to address the limitations of temporal processing in a two-dimensional convolutional neural network (2D-CNN). By incorporating a training-stage mask and leveraging BERT’s attention mechanism, the model gains improved contextual understanding. This enhancement enables our network to achieve accuracies of 91.46% on the NTU60 dataset and 83.06% on NTU120, marking an improvement of 1.63% and 3.22%, respectively, with only a slight increase of 0.1 GFLOPs in computational cost. Unlike other real-time models that rely on large-scale datasets such as Kinetics for extensive pretraining, our model achieves competitive performance while being trained directly on NTU datasets. A key contribution of our work is its ability to perform dynamical frame recognition, efficiently updating the ModConvLSTM cell and computing BERT every four new frames, achieving real-time inference in just 14.2ms on a CPU. This study enables continuous action recognition without requiring predefined action boundaries, making our model well-suited for real-world applications where actions occur naturally without explicit segmentation.

Effective landslide change detection from remote sensing images is essential for disaster prevention and risk management. This paper introduces a lightweight neural network architecture that integrates a Feature Enhancement Block (FEB) with the SimAM attention module to improve change detection accuracy in complex terrains. Additionally, the use of Focal Tversky Loss helps optimize the model's focus on genuine changes. The proposed lightweight network architecture significantly accelerates processing on GPGPUs, enabling real-time detection capabilities. Experimental results demonstrate that our model achieves an 88% reduction in parameters while improving Intersection over Union (IoU) by 8.5% and Recall by 8.37% compared to the baseline, showcasing its efficiency and accuracy.