Hong Zhang

Exposure mismeasurement is commonly seen in causal inference based on observational data. Recent evidence suggests that 35%–61% of medical studies using diagnostic tools report non-negligible exposure mismeasurement. For example, in the Global Enteric Multicenter Study, enzyme-linked immunosorbent assays (ELISA)-based rotavirus detection showed substantial discordance with polymerase chain reaction-based TaqMan Array Card (TAC) testing: 20.1% of ELISA-negative cases tested TAC-positive, whereas 19.5% of ELISA-positive cases tested TAC-negative. The existing methods handling exposure measurement error usually require that the mismeasurement distribution is known or can be estimated from a subset of study subjects with validated exposure statuses. This condition, however, is often violated when the mismeasurement mechanism is complex. In this article, we develop a novel method for estimating average causal effect using samples with exposures subject to misclassification and limited validation subsamples with correctly ascertained exposures. Our method does not require making any modeling assumptions about exposure mismeasurement. We establish some large-sample properties for the proposed method, based on which statistical inference can be easily carried out. The superiority of our proposed estimator is illustrated through extensive simulation studies. Our proposed estimator is shown to achieve a 39%–53% reduction in variance over the estimator relying solely on validation data in our simulation situations. In the application to the Global Enteric Multicenter Study, the proposed method reveals a substantial adjustment of pathogen-specific diarrhea associations that were previously distorted by inaccurate diagnoses.

Recent approaches in visual place recognition (VPR) have leveraged foundation models (FMs) and introduced novel aggregation techniques. However, these methods do not fully exploit key FM concepts, such as the effective use of different training datasets, and they overlook the potential of classical aggregation methods like GeM and NetVLAD. Building on these insights, we revisit classical feature aggregation methods and develop more fundamental VPR models, collectively referred to as SuperPlace. To address the first limitation, we propose a supervised label alignment method that facilitates training across diverse VPR datasets within a unified framework. To tackle the second limitation, we introduce G²M and NVL-FT². G²M is a compact feature aggregation method that utilises two GeMs, where one branch uses GeM pooling and an MLP to learn channel attention weights that calibrate the output of the other GeM. NVL-FT² is a high-dimensional aggregation method (relative to GeM) that proposes a secondary fine-tuning (FT²) strategy for NetVLAD-Linear (NVL). NetVLAD first learns feature vectors in a high-dimensional space and subsequently compresses them into a lower-dimensional space via a single linear layer. Extensive experiments highlight our contributions and demonstrate the superiority of SuperPlace. Specifically, SuperPlace-G²M achieves promising results with 768-dimensional descriptors, less than one-tenth of the dimensions of SALAD and BoQ. Source code is available at https://github.com/BinuxLiu/SuperPlace.

1. Flooding is one of the most frequent and widespread natural disasters worldwide and poses a severe threat to human life and property (Garg et al. 2024; Rentschler, Salhab, and Arga Jafino 2022)....

Construction and demolition waste (C&DW) mapping is crucial for urban environmental management, yet accurate extraction remains challenging due to complex backgrounds, spectral confusion, and irregular boundaries. To address these issues, this study presents the first optical-SAR multimodal semantic segmentation dataset based on GF-3 SAR, Sentinel-1 SAR, and Sentinel-2 optical imagery with a unified spatial resolution of 3 m for fine-scale C&DW extraction and proposes CDW-Net, a dual-branch multimodal fusion framework. The framework incorporates a Cross-Frequency Interaction (CFI) module for cross-modal feature fusion and a Multi-Source Boundary Attention Module (MS-BAM) for improved boundary representation. Experiments in Beijing show that CDW-Net achieves an IoU of 0.8470 and an F1-score of 0.9171, outperforming state-of-the-art unimodal and multimodal methods. Transfer experiments in Guangzhou achieved an IoU of 0.8247 and an F1-score of 0.9040 on manually interpreted validation samples, demonstrating good cross-regional transferability. The proposed dataset and framework provide valuable support for fine-scale C&DW monitoring in complex urban environments.

Loop closure detection is important for simultaneous localization and mapping (SLAM), which associates current observations with historical keyframes, achieving drift correction and global relocalization. However, a falsely detected loop can be fatal, and this is especially difficult in repetitive environments where appearance-based features fail due to the high similarity. Therefore, verifying a loop closure is a critical step to avoid false-positive detections. Existing works in loop closure verification predominantly focus on learning invariant appearance features, neglecting the prior knowledge of the robot’s spatial-temporal motion cue, i.e., trajectory. In this article, we propose ROVER, a loop closure verification method that leverages the historical trajectory as a prior constraint to reject false loops in challenging repetitive environments. For each loop candidate, it is first used to estimate the robot trajectory with pose-graph optimization. This trajectory is then submitted to a scoring scheme that assesses its compliance with the trajectory without the loop, which we refer to as the trajectory prior constraint (TPC), to determine if the loop candidate should be accepted. Benchmark comparisons and real-world experiments demonstrate the effectiveness of the proposed method. Furthermore, we integrate ROVER into state-of-the-art SLAM systems to verify its robustness and efficiency. Our source code and self-collected dataset will be made available online at https://rover-lcv.github.io/ upon publication of this article.

Knowledge-guided decision making for multi-robot exploration remains limited by the weak operational coupling between explicit geometric knowledge and continuous policy learning. In most hybrid frameworks, topological knowledge is used only as auxiliary guidance, which often leads to unstable mode switching, redundant exploration of topologically invalid regions, and dependence on computationally heavy perception modules. To address this problem, we propose PRIME, a priorguided framework that operationalizes geometric priors as explicit constraints on continuous policy optimization. By combining dynamic Voronoi partitioning with multidimensional utility evaluation,PRIME prunes structurally invalid exploration regions before action generation and couples macrolevel topological knowledge with micro-level adaptive control in a unified decision process. This prior-constrained formulation also enables a lightweight multilayer perceptron policy with compact multimodal state encoding, supporting responsive obstacle avoidance without convolutional feature extraction. Extensive simulation and real-world experiments show that PRIME converges in about 500 episodes with roughly 4 minutes of wall-clock training in simulation. In unseen complex environments, it reduces exploration time by 16.8% relative to the geometric baseline CURE and improves average map coverage from 94.0% to 97.6%. PRIME further achieves a 94.5% zero-shot generalization success rate with minimal edge computational overhead. These results indicate that explicit geometric knowledge can be embedded as operative decision constraints to improve efficiency,generalization, and deployability in multi-robot exploration.