Instance segmentation of remote sensing imagery (RSI) is vital for applications like geographic information system (GIS) updates and urban planning. Due to RSI’s diversity (e.g., scale variations and complex object shapes), instance segmentation models require extensive labeled data, which is costly and labor-intensive. To mitigate this dependence on extensive labeled data, we propose a reinforcement learning (RL)-driven iteratively-refined instance segmentation model for RSI by RL (RL-ISegNet) framework, which uses policy gradient optimization to enhance the learning of hard samples in limited labeled data, thereby improving data utilization efficiency under limited samples. Specifically, our RL-ISegNet defines a sequential optimization process for hard samples, using cumulative rewards from each subtask (region proposal, detection, and segmentation) to provide future feedback guidance for optimizing network parameters. Meanwhile, considering the conflicts between subtasks, we propose a metric to quantify the positive or negative impacts between subtasks as a loss constraint, thereby increasing the sampling probability of trajectories with positive impacts. Additionally, we developed a Fisher information-based method to freeze low uncertainty parameters, reducing the training time. The experimental results show that our RL-ISegNet improves data utilization efficiency by +12.2% and achieves improvements of +7.1% in instance segmentation metrics with 4% of the training dataset.

Full-waveform hyperspectral LiDAR (HSL) generates comprehensive hyperspectral waveforms for scenes to reveal the shape and spectral heterogeneity of multiple natural targets. Nevertheless, current waveform processing methods are primarily designed for single-wavelength LiDAR systems, resulting in a shortage of methods tailored for full-waveform HSL data processing and in a restriction to further quantitative applications for HSL. This study is designed to extract targets’ physical and spectral characteristics by integrating spectral-dimension features into the HSL waveform processing. The core idea of the method involves a rigorous processing technique consisting of parameter initialization, parameter optimization, and re-optimization over calculating the median (M) after ranking central locations of natural target echoes (Rclonte). The medians in the re-optimization step serve as the reference parameter sets for supplementing the hidden or weak components at some wavelengths for HSL. Two groups of datasets, the simulated and measured datasets, were utilized to evaluate the component detection ability of the proposed Rclonte-M method. The results suggest that the Rclonte-M method demonstrates excellent component detection performance on both simulated and measured data, outperforming the multispectral waveform decomposition (MSWD) method. The HSL system designed by us owns an overall ranging error of about 7 cm for adjacent components, with the relative neighbor distance error (RNDE) limited to 0.160. Besides, spectra retrieval results from HSL easily distinguish the natural targets along the laser path. This study enriches the full-waveform HSL data processing algorithm library and could be considered in other full-waveform HSL systems and the simulated airborne or space-borne HSL waveforms. Codes are freely available on https://github.com/Jie-Bai/Rclonte-M-TGRS.

Location-awareness has become a fundamental requirement for multiple emerging applications with the rapid development of wireless technologies. The high-accuracy ranging enabled by ultra-wide bandwidth (UWB) signals is often deteriorated by clocks imperfections and non-line-of-sight (NLOS) propagation. Existing supervised learning methods for NLOS identification and mitigation are time-consuming, labor-intensive, and cost-inefficient due to the need for training data acquisition and label assignment. This paper presents an indoor localization system that enables NLOS mitigation based on self-training. The system provides a general information fusion framework that integrates map, inertial sensors, and UWB measurements, where the weak labels for UWB measurements are produced and iteratively refined by multi-sensory information fusion for self-training. In addition, the system utilizes the maximum likelihood ranging estimator that considers the impact of clock drift. The effectiveness of the proposed system is demonstrated via extensive experimentation in multiple real-world environments, e.g., the proposed methods reduce the NLOS ranging error by 80% and result in a 90th localization error percentile of 0.5 meters in a complex indoor environment.

As a novel remote sensor, hyperspectral LiDAR is faced with the incidence angle effect, which restricts its quantitative applications. However, the current radiometric correction algorithms have some limitations, concentrating on: 1) the mathematically polynomial fitting; 2) adjacent wavelength ratio such as ratio vegetation index; and 3) perfect Lambertian assumption and using the Lambert cosine law to correct the effect. In this study, a practical and proper correction algorithm is proposed to overcome these limitations. First, to better characterize the complex reflection characteristics of the object surface, a combination of the Lambert law and Beckmann law is applied to represent the object surface. Then, it considers the impact of both wavelength and incidence angle on describing the surface roughness factor and diffuse fraction. Finally, a modified and physically based radiometric correction algorithm is generated. It provides the detailed correction equations for intensity and reflectance data recorded by hyperspectral LiDAR. To obtain its parameters and verify its reliability, leaves (ten samples per species, 30 in total) were stochastically collected from three broadleaf trees to experiment. The results showed that the algorithm achieved good performances by comparing the intensity and reflectance changes before and after removing the leaf incidence angle effect. Since it is physically based, the algorithm could promisingly be a fundamental solution to eliminate the incidence angle effect for hyperspectral LiDAR.

We present a novel self quality evaluation metric SQE for parameters optimization in the challenging yet critical multi-object tracking task. Current evaluation metrics all require annotated ground truth, thus will fail in the test environment and realistic circumstances prohibiting further optimization after training. By contrast, our metric reflects the internal characteristics of trajectory hypotheses and measures tracking performance without ground truth. We demonstrate that trajectories with different qualities exhibit different single or multiple peaks over feature distance distribution, inspiring us to design a simple yet effective method to assess the quality of trajectories using a two-class Gaussian mixture model. Experiments mainly on MOT16 Challenge data sets verify the effectiveness of our method in both correlating with existing metrics and enabling parameters self-optimization to achieve better performance. We believe that our conclusions and method are inspiring for future multi-object tracking in practice.