Abstract In recent years, with the rapid growth in the scale of datasets and neural network models, there has been a significant imbalance between the memory requirements during model training and the memory resources available on training devices. Existing memory optimization techniques like recomputation, memory swapping, and their adaptive combinations do not fully consider the structural information of the model and overlook the impact of application costs and timing on training efficiency. Addressing this issue, this paper proposes a memory optimization method called MemOpt, which uses dual-agent reinforcement learning to dynamically search for appropriate memory optimization strategies and execution timing based on model structure and device information. It can optimize memory without compromising accuracy while minimizing additional overhead. Experimental results show that the MemOpt method significantly increases the maximum batch size for model training by up to 8.7% and training throughput by up to 42.3% compared with baseline methods. In the future, this method may find better applications in large-scale neural networks.

Aiming to enhance the docking efficiency and safety of aerial recovery (AR), this paper proposes a bilateral collaborative docking strategy with spatiotemporal constraints. To improve efficiency, the strategy simultaneously controls the unmanned aerial vehicle (UAV) and the drogue to move toward the desired docking position. To ensure safety, the bilateral recovery systems are strictly constrained within a safe docking space and are required to complete docking within a prescribed time. An appointed-time prescribed performance control (ATPPC) is employed to describe these spatiotemporal requirements. Unknown system dynamics estimator (USDE) observers are established for each subsystem to reconstruct the unknown nonlinear dynamics. Based on the USDE–ATPPC framework, a collaborative docking controller is constructed to ensure efficient and safe docking without violating the constraints. Simulation results show that the propose strategy improves docking efficiency by about 43% while ensuring safety and accuracy.

The applications of unmanned swarms have become increasingly widespread, gradually transforming production processes and daily life. Task allocation, the top-level design for unmanned swarm missions, is pivotal to maximizing the efficiency of the entire swarm. However, traditional optimization methods and intelligent algorithms, including Reinforcement Learning (RL), often struggle to adapt to the complex and unpredictable situations in these tasks. To address this challenge, we propose a novel Multi-Agent Proximal Policy Optimization (MAPPO) algorithm combined with the population-based learning and Gaussian Mixture Model (GMM)-based adjustment mechanisms (PG-MAPPO). In PG-MAPPO, the population-based learning mechanism is integrated to enable agents with diverse exploration preferences to uncover optimal collaboration patterns among Unmanned Aerial Vehicles (UAVs), thereby enhancing cooperative efficiency. The GMM-based adjustment mechanism dynamically adjusts UAV formations for each agent, significantly improving the swarm’s flexibility and adaptability in rapidly changing environments. To demonstrate the effectiveness of PG-MAPPO, a maritime rescue simulation containing multiple complex and dynamic scenarios is conducted. Experimental results show that our algorithm achieves higher rescue success rate with faster convergence and greater stability than state-of-the-art Multi-Agent Reinforcement Learning (MARL) methods in all scenarios. Notably, the PG-MAPPO algorithm improves the rescue success rate by 31.6% compared to the best-performing baseline under challenging conditions.

Pedestrian dead reckoning (PDR) and global navigation satellite system are two popular solutions for pedestrian navigation with a smartphone. pedestrian dead reckoning (PDR) estimates the user’s position by analyzing their walking gaits, including step length and heading angle. However, PDR position errors can accumulate over time due to measurement noise. In contrast, GNSS generates position information by processing radio signals. However, these signals can be affected by blockage and interference. GNSS and PDR are often integrated using a Kalman filter (KF) to provide a more reliable solution. While current integration methods rely on position and velocity measurements, pseudo-range measurements for PDR and GNSS integration still need to be explored. To improve the accuracy of pedestrian position estimation in urban areas, we propose a walking-gaits-aided smartphone GNSS approach. This approach involves employing a factor graph optimization (FGO)-based GNSS/ PDR tight integration method. The FGO- GNSS/ PDR tight integration considers the pseudo-range measurements from each satellite, pedestrian position, and step length to optimize the position estimation. We introduce a fuzzy adaptive FGO (A-FGO) to enhance the accuracy further to suppress pseudo-range outliers. We conducted two experiments using a Samsung Galaxy A40 and Huawei Mate 40 Pro smartphones to evaluate the accuracy of the proposed methods. Our experimental results demonstrate that the proposed methods effectively improve the PDR/ GNSS position accuracy.

Existing acoustic-based positioning solutions face difficulties achieving precise ranging and positioning, especially in dynamic situations, due to Doppler frequency shift (DFS). In this article, we present a solution that achieves precise ToF/distance measurements between the kinematic receiver and a stationary transmitter with chirp-based Doppler shift compensation (DSC). In the solution, specific chirp signals with an upchirp and downchirp branch are transmitted by the stationary transmitter. The kinematic receiver receives and detects these signals, accordingly corrects the measurements with the proposed DSC method, and estimates the real-time velocity based on a corresponding model. After obtaining the compensated ToF/distance measurements and real-time velocities of the kinematic receiver, the initial and subsequent locations of the kinematic receiver can be precisely determined with the extended Kalman filter (EKF) and Rauch-Tung-Striebel smoother (RTS). To verify the performance of our solution, experiments in ranging and positioning were conducted in an indoor open space. The results show that the developed DSC is able to achieve an average ranging accuracy of 0.1 m for the kinematic receiver with a motion velocity of larger than 1.5 m/s in line-of-sight (LOS) situations and achieves an average positioning accuracy of 0.46 m for the kinematic receiver with motion velocity up to approximately 2 m/s. Therefore, the developed approach is sufficient for realizing acoustic-based positioning in both static and dynamic situations.

Hyperspectral LIDAR (HSL) is an innovative active remote sensing technology that allows for the simultaneous collection of spectral and spatial information. In this study, we primarily focus on the radiation correction method of the incident angle and distance effects for the backscatter intensity of HSL. We have developed a comprehensive radiometric correction model that addresses these effects. Additionally, we have applied the correction model to point cloud classification using the random forest method. Comparing the accuracy of point cloud classification before and after correction, we observed a 9.6% improvement in overall accuracy (OA) and a 10.8% improvement in the kappa coefficient. These results indicate that the radiometric correction model significantly enhances the classification accuracy.

Convolutional neural networks (CNN) are capable of extracting features with high accuracy, which is dominant in visual-based classification. Previous researches demonstrate that CNN can extract essential features of the target in the plant feature extraction and classification. Hyperspectral LIDAR (HSL) is a novel active remote sensing technology that can simultaneously collect spectral and spatial information. This paper proposed a novel classification method named VI-CNN for hyperspectral LiDAR, which combines the spectral features with the vegetable index(VI). As far as we know, we are the first to apply CNN to HSL data classification. The VI -CNN is divided into two parts. Firstly, spectral CNN focuses on intra-spectral correlations; secondly, the vegetation indices supplement the biological parameters. The evaluation shows that the concatenation has stronger identification and robustness than standalone methods. The experimental results demonstrate that the VI-CNN significantly improves the classification accuracy against other traditional machine-learning methods.

Position and location constitute critical context for Internet of Things (IoT) devices. Global navigation satellite systems (GNSSs) are the primary apparatus providing precise position and location information for IoT devices in outdoor environments. However, in dense urban areas, non-line-of-sight (NLOS) signals will induce large errors in GNSS pseudorange measurements due to the additional signal transmission paths. The vector tracking (VT) technique utilizing a Kalman filter (KF) to estimate navigation solutions has been investigated in NLOS detection, and its advantages have been demonstrated. However, the estimation of NLOS-induced bias has not been thoroughly investigated in the VT framework. In this article, we focus on the estimation and correction of NLOS-induced errors within the VT framework. First, graph optimization (GO) instead of a KF is incorporated with VT to optimize the estimation of navigation solutions. The NLOS-induced bias is then added to the VT state vector as the variable for real-time estimation. Compared with the KF-VT method, in GO-VT, the state transformation and the measurement model are regarded as constraints to optimize the state vector estimation. Hence, the GO-VT framework is more flexible than the KF approach in dealing with state vector changes. An iterative process is conducted to solve for the optimization results; a multiple-correlator scheme is employed in GO-VT to provide the initial values of the NLOS-induced bias. Three collected GPS L1 data sets (static and dynamic) are used to evaluate the proposed method. The statistical results support the conclusion that GO-VT with state augmentation achieves superior position estimation in urban areas.

Airborne hyperspectral images are used for crop identification with a high classification accuracy because of their high spectral resolution, spatial resolution, and signal-to-noise ratio (SNR). However, the tradeoffs between the three core parameters of a hyperspectral imager (SNR, spatial resolution, and spectral resolution) should be considered for designing an efficient imaging system. Only a few reported studies on the analysis of the impact of SNR on identification accuracy are available. Further, the tradeoffs and mutual interactions among these parameters are rarely considered. In this empirical study, our aim was to understand the relationship among the core parameters and their effects on crop identification accuracy by analyzing the tradeoffs and mutual interactions among these parameters. We analyzed the hyperspectral images of a typical plain agricultural area in Xiongan, China, acquired by the newly developed sensor airborne multimodular imaging spectrometer (AMMIS). The fundamental images were transformed to form datasets with different ranges of spectral resolution, spatial resolution, and SNR using data reconstruction methods. We adopted the classification and regression tree (CART), random forest (RF), and k-nearest neighbor (kNN) classifiers, and observed the overall accuracy (OA) across the degraded hyperspectral datasets. The experimental results indicated that the OA decreased with a decreasing SNR. As the spectral resolution became coarser, the OA first increased, plateaued, and then decreased. However, the OA increased with decreasing spatial resolution. This study was performed with the goal of bridging the knowledge gap between the back-end hyperspectral sensor designing and its front-end applications.

During the mining operation, it is a critical task in coal mines to significantly improve the safety by precision coal mining sorting and rock classification from different layers. It implies that a technique for rapidly and accurately classifying coal/rock in-site needs to be investigated and established, which is of significance for improving the coal mining efficiency and safety. In this letter, a 91-channel hyperspectral LiDAR (HSL) using an acousto-optic tunable filter (AOTF) as the spectroscopic device is designed, which operates based on the wide-spectrum emission laser source with a 5-nm spectral resolution to tackle this issue. The spectra of four-type coal/rock specimens collected by HSL are used to classify with three multi-label classifiers: naive Bayes (NB), logistic regression (LR), and support vector machine (SVM). Furthermore, we discuss and explore whether Gaussian fitting (GF) method and calibration with the reference whiteboard (RB) can enhance the classification accuracy. The experimental results show that the GF technique not only improves the accuracy of range measurement but also optimizes the classification performance using the spectra collected by the HSL. In addition, calibration with RB can improve classification accuracy as well. In addition, we also discuss methods to improve the calibration-free classification accuracy preliminarily.

In this letter, a hyperspectral light detection and ranging (HSL) with 10-nm spectral resolution was designed and tested using a supercontinuum laser source. The major difference between the prototyped HSL and similar instruments was that a liquid crystal tunable filter (LCTF) was installed before the avalanche photodiode detector and utilized as a spectroscopic device. The design allowed continuous wavelength selection of the backscattered echoes in the time dimension. Moreover, for general accuracy evaluation of range measurement and spectral measurement, laboratory experiments for vegetation red edge detection were performed using the prototyped HSL to assess its feasibility on agriculture application. Yellow and green leaves from aloe and dracaena plants were measured by the LCTF-HSL for detecting the corresponding “red edge” position. Spectral profiles measured by an SVC-HR-1024 spectrometer which is designed by SVC company were used as a reference to evaluate the measurements of HSL. The comparison results showed that the red edge positions extracted from the two individual measurements were similar, thus indicating that the LCTF-based high-resolution HSL was effective for this application.

Recently, a major effort has been made to develop methods or tools for rock characterization and mineral content mapping. Light detection and ranging (LiDAR) is an efficient active remote sensing technique for collecting geometry information about rock surfaces. However, traditional LiDAR sensors work with a single-wavelength laser source, and it is unfeasible to obtain spectral information using one LiDAR sensor. The combination of hyperspectral imaging and LiDAR techniques is an emerging method for acquiring spatial and spectral information simultaneously that allows remote mapping of high-resolution mineral content and distributions and identifies subtle chemical variations. Unfortunately, spatial and spectral data registration, which introduces additional complicated data processing, is an inevitable and essential issue for this method. In this letter, first, we investigate the feasibility of ore classification applications with hyperspectral LiDAR (HSL). HSL consists of 17 spectral channels covering the visible–shortwave infrared (SWIR) spectral range. Spatial and spectral information about seven different ore samples is obtained under a controlled laboratory environment using HSL. The standard deviation of the distance measurements is less than 1.1 cm for different spectral channels, and the classification accuracy can reach 100% if all 17 spectral measurements are used. To optimize the system design with lower cost and system complexity, a spectral band selection criterion is built based on the feature contribution degree (FCD), which is calculated using the normalized variance of the reflectance values for different ore samples at each wavelength. Two different strategies of FCD selection are tested to generate vectors: ascending sequences and descending sequences. Feature vectors with descending sequences have better classification accuracy. In addition, the results show that the classification accuracy can reach 100% with the feature vector of the seven largest FCD values compared to 59.57% for the feature vector with the seven smallest FCD values. Moreover, we find that the channels with high FCD values are primarily centered in SWIR bands. This result could be a reference for optimizing the hardware design of HSL for ore classification or mineral identification.