Urban growth simulation is crucial for sustainable spatial planning and policy-making. Cellular automata (CA) are among the most widely used approaches, and existing models can be categorized into raster-based (RCA) and vector-based (VCA) paradigms, in which structural differences hinder integration and transferability of available models. Here, a universal graph-based cellular automata (UniGCA) framework is proposed to equally support these two distinct paradigms and address the challenges in spatial structure design and spatial interaction modeling. Building on the UniGCA framework, the Partitioned Graph U-Net Cellular Automata (PGUN-CA) model is developed as a practical implementation. PGUN-CA employs optimized graph storage and multilevel graph partitioning to overcome memory and computational constraints in large-scale urban simulation, while integrating Graph U-Net with attention mechanisms to capture complex spatial interaction. Case studies in Shanghai (raster data) and the Pine Catchment, Australia (vector data), demonstrate that PGUN-CA achieves higher accuracy than baseline models, with figure of merit (FoM) improvements of 6.96% and 9.41%. These findings highlight the generality and scalability of the UniGCA framework, which will provide a practical and flexible solution for large-scale urban growth simulation across heterogeneous data types and solid support for urban planning and sustainable development.

Directly training on quantized model via gradient scale correction for edge device.

Real-time multi-object tracking (MOT) in sports videos faces persistent challenges arising from dense player occlusion, rapid nonlinear motion, and the stringent computational constraints of edge devices. This paper presents an integrated framework that co-designs a lightweight detector, an adaptive tracking strategy, and an edge deployment pipeline to address these challenges jointly. We propose Sports-LiteDet, a compact detection network featuring a coordinate-attention-enhanced backbone and a bidirectional multi-scale feature fusion neck, which achieves 71.8% mAP@0.5 on sports test data at only 3.2 GFLOPs. For tracking, we introduce a motion-consistency association module coupled with adaptive keyframe scheduling and an asynchronous detection-tracking pipeline, attaining 68.3% MOTA and 71.6% IDF1 while nearly doubling throughput over synchronous baselines. A sequential compression strategy combining structured pruning and INT8 quantization reduces the model to 5.8MB, enabling 68.0 FPS inference on NVIDIA Jetson Orin Nano at under 10W power consumption. Extensive experiments on public sports benchmarks and self-collected sequences demonstrate that the proposed framework achieves competitive tracking accuracy with substantially lower latency compared to existing methods, validating its viability for field-deployable intelligent sports analytics.

ABSTRACT While mobile battery energy storage systems (MBESSs) are typically used to improve the stability of power systems, their ability to move also creates good opportunities for businesses to earn money through energy arbitrage. This profit depends heavily on decisions about timing and location, and is affected by uncertain conditions like fluctuating electricity prices and traffic. However, finding the best real‐time control strategy that considers long‐term profit and these uncertainties requires significant computing power. To tackle this issue, this paper presents a deep reinforcement learning framework for MBESSs designed to get the most profit from market arbitrage. Within this framework, we introduce the Knowledge‐Assisted Deep Deterministic Policy Gradient (KA‐DDPG) algorithm to learn the best policy more efficiently. The core novelty of KA‐DDPG lies in its probabilistic hybrid action selection mechanism that unifies the agent's learned policy, offline expert criteria, and random exploration to manage the complex hybrid action space. Additionally, a two‐phase guidance strategy is implemented to transition from offline‐based to real‐time‐based criteria actions, ensuring both learning acceleration and policy robustness under computational constraints. Our rigorous statistical evaluations demonstrate that the proposed KA‐DDPG approach leads to a 3%–7% improvement in average profits over the state‐of‐the‐art Soft Actor‐Critic baseline. Furthermore, it achieves exceptional policy stability, exhibiting a variance reduction of over 60% compared to standard DRL baselines and over 92% compared to the deterministic closed‐loop MPC. The KA‐DDPG algorithm also substantially speeds up the learning phase, validating its efficacy for real‐time MBESS control under high uncertainty.

Predicting spatial transcriptomics from histology images offers cost-effective insights but faces privacy barriers preventing cross-institutional data sharing. To address this, we present privateST, a homomorphic-encryption-optimized framework that substantiates the feasibility of secure spatial transcriptomics prediction. To accommodate the computational constraints of homomorphic encryption, we downsampled the input images using bilinear interpolation. To improve prediction accuracy for the 100 target genes, we incorporated predictions for an auxiliary set of 150 highly expressed genes. This approach enables the model to learn from a more diverse genomic context, thereby refining the feature representations for the high-priority target genes. To adapt the architecture for homomorphic encryption, we replaced Max-Pooling and ReLU with Average-Pooling and polynomial approximation, respectively. These modifications eliminate non-linear comparison operations, significantly reducing the multiplicative depth. We also implemented multiplexed packing to optimize the efficiency of encrypted data processing. Remarkably, our results demonstrate that despite the reduced input resolution, this proposed approach achieves accuracy comparable to the original ResNet-18 configurations without downsampling.

Chemical space visualization supports exploration of high-dimensional molecular data by revealing patterns of similarity, diversity, and structure-property relationships. As chemical libraries expand from thousands to billions of compounds, practical visualization increasingly depends on pipelines that balance chemical meaning with computational and memory constraints. In this survey, we review 56 studies published between 2000 and 2025 and synthesize the end-to-end workflow of chemical space visualization across five core stages: dataset choice, molecular featurization, dimensionality reduction (DR), clustering, and evaluation. We quantify usage trends over time and relate method selection to dataset scale. Across the literature, fingerprints remain the dominant representation for large libraries due to their scalability, while recent studies increasingly incorporate fragments, SMILES-based encodings, and learned embeddings when richer signals are needed. DR practice shows a shift from PCA-centric baselines to neighborhood-preserving methods such as t-SNE and UMAP, with graph layout approaches like TMAP enabling visualization at extreme scale. Clustering shows the weakest convergence to a single standard: hierarchical methods, K-means, and SOM remain frequent choices, complemented by scalable summarization and domain-driven strategies (e.g., BIRCH/BitBIRCH and scaffold-based partitioning) when all-pairs similarity becomes prohibitive. Finally, we propose dataset-size-aware pipeline archetypes and identify open challenges, including inconsistent structure-aware evaluation, limited reproducibility reporting, and the need for scalable, chemically grounded methods for ultra-large libraries.

Parallelized hybrid ensemble machine learning framework for scalable and accurate rainfall prediction

HAB-Net: A heterogeneous augmentation based network for weld defect detection

Holographic displays are highly anticipated as a three-dimensional (3D) display technology that satisfies the essential requirements of human visual perception. However, the generation of computer-generated holograms (CGHs) is hindered by the significant challenges of high computational complexity and long processing times. In this study, we propose a method to reformulate the point-cloud method into a matrix multiplication framework. By representing the redundancy of object coordinates in the separable convolution method and the computational domain constraints of the wavefront recording plane method as sparse matrices, we eliminate redundant calculations and achieve efficient hologram generation. In addition, we further accelerate matrix-based hologram calculations by leveraging the matrix computation capabilities of modern graphics processing units.

Ultrafast dynamic vision remains a major challenge for autonomous systems like drones, particularly considering their strict power and computational constraints. Here, we present an artificial spherical compound eye (ASCE) that enables nanosecond-scale, event-driven motion detection across a panoramic field of view (294°). Each pixel is self-powered and responds exclusively to dynamic light changes, enabling real-time sensing with minimal processing. An 8 ns response time is achieved under femtosecond laser excitation, alongside exceptional stability exceeding 10⁸ switching cycles. In addition, the ASCE features a simple structure with in situ grown nanowires and dielectric layer. The one-step fabrication process allows for conformal pixel integration on spherical and other three-dimensional surfaces. We demonstrate its versatility across multi-scale applications, including drone-based laser alarming and real-time motion tracking. These results establish ASCE as a promising platform for energy-efficient, high-speed vision in autonomous robotics and intelligent surveillance.

Accurate and transparent classification of breast cancer histopathology remains a major challenge due to morphological variability, class imbalance, and computational constraints in whole-slide image analysis. Convolutional neural networks (CNNs) capture local tissue features but tend to ignore more global context cues; on the other hand, Vision Transformers are data-hungry and sensitive to staining variations. We provide a systematic, controlled comparison, and propose a hierarchical Swin Transformer framework designed to leverage both local and global representations via adaptive channel recalibration and attention-based feature aggregation on RoI images. Class-balanced upsampling helps further improve robustness against uneven distribution of samples. Evaluations on the BRACS dataset demonstrate performance gains of 7-10 % in the accuracy and F1 score compared to strong CNN and ViT baselines. We assessed multiple explainability techniques to maintain clinical transparency and found that the model highlights tissue regions that are diagnostically meaningful. The proposed framework strikes a good balance between predictive performance and interpretability for computer-aided breast cancer diagnosis.

Object detection in UAV imagery faces severe challenges arising from drastic scale variations, pervasive background noise, and the dominance of tiny objects, further compounded by the computational constraints of edge devices. Existing lightweight detectors frequently suffer from feature degradation, in which low-frequency background information tends to submerge the high-frequency details of small objects. To overcome this limitation, we propose S3-Det, a novel Spectro-Spatial Synergistic Detector. Specifically, we design the Spectro-Spatial Synergistic Network (S3Net) as the backbone architecture. Subsequently, we design an Implicit Recursive Feature Aggregator to enhance feature semantics without explicitly increasing network depth. By leveraging implicit refinement units with shared weights, this module recursively refines the feature pyramid. This mechanism achieves deep representational capacity with the parameter efficiency of a shallow network. Finally, we propose a decoupled detection head incorporating large-kernel context regression and SIoU loss to effectively alleviate the misalignment between classification and localization for tiny objects. Extensive experiments demonstrate that S3-Det achieves an exceptional trade-off between accuracy and latency, surpassing state-of-the-art lightweight detectors to establish a new benchmark for real-time aerial surveillance.

Real-time accurate evaluation of heart rate status during physical activity is a critical requirement in physiological analysis, performance enhancement, and early warning systems for cardiovascular diseases. In this work, a lightweight hybrid machine learning model for classifying the status of heart rate during physical activities into normal and warning levels using a combination strategy of Multilayer Perceptron, Naïve Bayes, and K-Nearest Neighbors algorithms in a weighted voting-based ensemble model is presented. The model analyzes a set of physiological and environmental parameters such as temperature, humidity, speed, incline, and activity time. Before embarking on model development, a thorough statistical analysis of available data was performed, which involved descriptive statistics, density analysis, normality tests using Shapiro-Wilk tests, and both parametric and non-parametric hypothesis tests to establish the discriminative power of all input variables in spite of non-normality. Every individual base classifier model was tested for performance using a fixed split of 70/15/15 for training, validation, and testing, and their respective strengths were tapped using a weighted voting mechanism based on relief factor analysis. The simulation outcome shows that the proposed ensemble classifier performs better than standalone classifiers in achieving an accuracy of up to 96.67%, an F1-score of 96.66%, and an MCC of 0.9354, which is a manifestation of excellent classification balance and statistical significance. The proposed system was also tested on an embedded system developed using an Arduino platform for real-time processing, and it achieved an accuracy of 90.83% and an MCC of 0.8167. The performance degradation is due to sensor noises and limited numerical precision in an embedded system. The proposed framework contributes to improving real-time heart rate monitoring by addressing computational constraints and enabling efficient deployment on embedded platforms.

A quality-gated hybrid Viola–Jones pipeline for efficient face detection under computational constraints

Abstract The demand for accurate on-device pattern recognition in edge applications intensifies, yet existing approaches struggle to reconcile accuracy with computational constraints. To address this critical challenge, a resource-aware hierarchical network based on multi-spectral fusion and interpretable modules, namely the Hierarchical Parallel Pseudo-Image Enhanced Fusion Network (HPPI-Net), is proposed to enable real-time, on-device Human Activity Recognition (HAR) tasks. Deployed on ARM Cortex-M4 MCU for low-power real-time inference, HPPI-Net achieves 96.70% accuracy while utilizing only 22.3 KiB of RAM and 439.5 KiB of ROM after optimization. HPPI-Net employs a two-layer architecture: the first layer extracts preliminary features using Fast Fourier Transform (FFT) spectrograms, while the second layer selectively activates either a dedicated module for stationary activity recognition or a parallel LSTM-MobileNet network (PLMN) for dynamic states. PLMN fuses FFT, Wavelet, and Gabor spectrograms through three parallel LSTM encoders and refines the concatenated features with Efficient Channel Attention (ECA) and Depthwise Separable Convolution (DSC), thereby offering channel-level interpretability while substantially reducing multiply–accumulate operations. Compared to MobileNetV3, HPPI-Net notably increases accuracy by 1.22% and significantly reduces RAM usage by 71.2% and ROM usage by 42.1%. These results demonstrate that HPPI-Net realizes a favorable accuracy-efficiency trade-off and provides explainable predictions, establishing a practical solution for wearable, industrial, and smart-home HAR on memory-constrained edge platforms.

Small streams exhibit rapid and nonlinear flood responses due to steep slopes, short flow paths, and limited storage capacity, making real-time flood prediction difficult under both computational and data constraints. This study presents a measurement-based flood prediction framework for real-time estimation of flood discharge and depth in small-stream basins. Conventional approaches, such as physically based hydrodynamic models, require detailed boundary conditions and high computational cost, while data-driven models often lack physical interpretability. The proposed framework integrates high-frequency monitoring data from the Small-Stream Smart Monitoring System, short-term rainfall nowcasting from the MAPLE system, and nonlinear regression-based hydraulic relationships within a unified operational structure. Rainfall–discharge nomographs and depth–discharge rating curves were developed using a four-parameter logistic regression model based on long-term observations from 12 small streams in Korea. Additional comparisons with alternative regression forms confirmed the suitability of the 4PL model for representing nonlinear hydrological responses. Forecast rainfall was used to estimate discharge, which was subsequently converted to flood depth through calibrated rating curves. For ungauged reaches, depth–discharge relationships were derived using HEC–RAS-based scenario simulations and the Manning equation to enable spatially continuous prediction along stream networks. Model performance was evaluated using independent validation events, showing mean prediction accuracies of approximately 89% for discharge and 90% for flood depth. The framework reduces computational demand by relying on pre-established relationships while maintaining physically interpretable structures. The results indicate that the proposed approach can support real-time flood prediction in small streams under conditions like those examined in this study, although its applicability to other regions requires site-specific calibration and further validation.

Reliable autonomous navigation for Tianwen-3-class Mars rotorcraft must satisfy both sampling-level accuracy and hard real-time execution under severe onboard computational constraints. To address this challenge, we develop MarsBird-VII, a mission-constrained stereo visual–inertial navigation system that combines a computation-aware vision front-end with a Parity-Window sliding-window optimization back-end. The front-end decouples high-rate tracking from feature replenishment to bound perception latency, while the back-end alternates updates over interleaved state subsets and preserves full-window coupling through unified marginalization. Unlike simply reducing the sliding-window size, the proposed strategy reduces the per-update optimization cost without shrinking the geometric observation horizon, thereby improving the accuracy–runtime trade-off for embedded avionics. Earth-analog flight experiments demonstrate strong navigation performance under mission-relevant conditions. In full-sequence evaluation, the proposed system achieves an SE(3)-aligned translation APE of 0.31 m RMSE/0.47 m Max and further reaches 0.06 m RMSE/0.15 m Max on a nominal stable segment. Runtime profiling over 5000+ update cycles shows that the Parity-Window back-end keeps the maximum optimization latency below 58.32 ms, satisfying the 66.7 ms hard real-time deadline while maintaining accuracy close to full-window optimization. These results show that the proposed system provides a practical balance of accuracy, robustness, and deterministic real-time performance for Tianwen-3-class Mars rotorcraft navigation.

Autonomous odor source localization remains a challenging problem for aerial robots due to turbulent airflow, sparse and delayed sensory signals, and strict payload and computation constraints. While prior unmanned aerial vehicle (UAV)-based olfaction systems have demonstrated gas distribution mapping or reactive plume tracing, they rely on predefined coverage patterns, external infrastructure, or extensive sensing and coordination. In this work, we present a complete, open-source UAV system for online odor source localization using a minimal sensor suite. The system integrates custom olfaction hardware, onboard sensing, and a learning-based navigation policy that we train in simulation and deploy on a real quadrotor. Through our minimal framework, the UAV is able to navigate directly toward an odor source without constructing an explicit gas distribution map or relying on external positioning systems. We incorporate vision as an optional complementary modality to accelerate navigation under certain conditions. We validate the proposed system through real-world flight experiments in a large indoor environment using an ethanol source, demonstrating consistent source-finding behavior under realistic airflow conditions. The primary contribution of this work is a reproducible system and methodological framework for UAV-based olfactory navigation and source finding under minimal sensing assumptions. We elaborate on our hardware design and open-source our UAV firmware, simulation code, olfaction–vision dataset, and circuit board to the community.

The real-time state estimation and safety early warning are the two most important components of the distribution network management of power systems. The conventional strategies cannot satisfy the current grid requirements of efficiency, accuracy, and real-time response. In order to overcome this, this paper utilizes an edge computing-based multi-objective ant colony distributed algorithm to carry out distributed state estimation in distribution networks, which eliminates the computational time and accuracy constraints of conventional state estimation strategies. To ensure safety early warning a fault information matrix is defined to determine faulty sections. Faults on line segments are defined based on predefined threshold conditions, allowing thorough safety monitoring of the distribution network. The combination of the suggested models of state estimation and safety detection forms an IoT-based real-time state estimation and safety early warning system of distribution networks. It continuously estimates states, identifies faults, and sends safety warnings. In its application to check a three-feeder armored buried cable system in one facility, the system correctly determined that the first feeder was the faulty one, which is how it really was. It shows the promising potential of the system applications.

TS-PET: A Novel Framework for Fine-Tuning Pretrained Time-Series Models.