Ruihan Bai

Introduction: Accurate weed detection from unmanned aerial vehicle (UAV) images remains difficult because weed targets often blend with surrounding vegetation, vary substantially in scale, and exhibit irregular boundaries under cluttered grassland backgrounds. Methods: This study presents a UAV grassland weed detection framework that combines intelligent annotation with task-oriented detector adaptation. For label generation, SAM produces instance-level vegetation masks, SigLIP performs few-shot semantic matching, and a vision-language model audits ambiguous candidates using few-shot weed references, local candidate patches, and full-image context. Candidate outputs are converted into YOLO-format labels for detector training. For detection, a YOLOv11-based model is adapted with four complementary components: EIEStem for shallow boundary preservation, C3k2-EMA for multi-scale feature aggregation, SPPF-LSKA for contextual modeling, and LDConv for adaptive downsampling of irregular weed targets. Results: On 200 reference images, the SAM–SigLIP–VLM workflow achieved 92.4% label precision, a mean IoU of 0.813, and an F1-score of 0.905, while reducing the annotation time per image from 96.8 s to 31.6 s. Under five-fold image-level cross-validation on a UAV grassland weed dataset, the improved detector achieved 0.762 ± 0.004 mAP@0.5 and 0.545 ± 0.005 mAP@0.5:0.95, improving the YOLOv11 baseline by 4.0 and 4.7 percentage points, respectively. Discussion: Additional evaluations cover annotation quality, module ablation, module placement, downsampling mechanisms, detector comparison, and validation on an external public crop–weed dataset. These results show that the proposed framework improves both annotation efficiency and UAV grassland weed detection performance in complex vegetation backgrounds.

Small object detection in remote sensing imagery is challenging because tiny targets often suffer from large scale variation, weak boundaries, and heavy background interference. Although modern one-stage detectors are efficient, repeated down-sampling often removes fine boundary details. As a result, dense tiny objects may become difficult to separate in deep feature maps. Meanwhile, background textures such as roads, roofs, and shadows may look similar to small targets, leading to false detections. To address these problems, we propose EAR-Net, an Edge-Guided Adapter-Enhanced Recursive Fusion Network built on YOLOv11n. EAR-Net improves small object representation from three aspects. First, the Multi-Scale Edge-Guided Feature Injection Module (MSEG) uses Sobel-initialized learnable edge operators to provide explicit boundary guidance. Second, the Adapter-Enhanced C2PSA module introduces lightweight bottleneck adapters into the attention branch to suppress high-frequency background noise while preserving global context. Third, the Recursive Feature Fusion mechanism introduces a recursive “focus-distribute-refocus-fuse” process to reduce cross-scale spatial misalignment. Experiments on NWPU VHR-10, DIOR, RSOD, and HRRSD show that EAR-Net improves detection accuracy and localization precision while maintaining a practical balance between accuracy and efficiency.

Small-object detection in remote sensing remains challenging due to low-resolution, weak cues, and complex backgrounds. Although semantic-context enhancement (SCE) methods leverage spatial priors effectively, their effectiveness is often limited by the stage, which contextual features are integrated. When introduced in the backbone, contextual information is subsequently degraded by the neck’s repeated resampling and fusion operations. Conversely, integrating context modeling mechanisms between the neck and the detection head is also suboptimal, as the features have already undergone multiscale mixing and spatial transformations, resulting in weakened semantic representation. To address this issue, we propose YOLO-feature supplement (FeatSup), a context-preserving detector built upon YOLOv11, which supplements the detection head with preserved semantic features via a dedicated bypass path. Specifically, a global−local spatial aggregation (GLSA) module is integrated in the backbone to capture high-fidelity semantic features before scale mixing, while the scale sequence (ScalSeq) fusion module performs effective multiscale feature fusion and delivers the enriched features directly to the small-object detection branch, thereby avoiding the degradation typically caused by repeated resampling and cross-scale operations in the feature aggregation process. Experiments on six public datasets demonstrate that YOLO-FeatSup achieves superior small-object detection accuracy with reduced computational cost, offering an efficient solution for remote sensing applications.

The rapid development iteration of YOLO models has spurred extensive research into specialized adaptations tailored for remote sensing object detection (RSOD). Typically, these adaptations involve modifying specific YOLO versions to excel in dataset-specific benchmarks. Such efforts primarily focus on enhancing model performance on specific datasets, a trend that, while beneficial for achieving high leaderboard standings, frequently overlooks the critical aspects of generalizability, scalability, and transferability. Consequently, there is a significant gap in the research concerning the development of universally applicable modules that are easy to adapt across various baseline YOLO models—an essential factor for practical deployment and widespread adoption in real-world applications. This study addresses these shortcomings by introducing a novel feature fusion method, Context Enhancement (CE), Weight Adjustment-based Spatial Attention (WA), and Channel Shuffling (CS) — collectively termed CE-WA-CS, specifically designed for RSOD. This research rigorously tests the proposed feature fusion method across various YOLO models, substantiating its versatility and effectiveness in enhancing remote sensing object detection. Notably, in comparative performance evaluations involving several YOLO architectures, integrating our feature fusion method with YOLOv8 emerged as the most balanced in terms of accuracy, computational complexity, and inference speed. This optimal combination has been designated as YOLORemote. YOLORemote not only exemplifies a significant improvement in performance metrics but also demonstrates a practical balance that is crucial for real-time applications.

Object detection in remote sensing imagery exhibits difficulties due to complex backgrounds, diverse object scales, and intricate spatial context relationships. Motivated by the problems mentioned above, this paper introduces AeroDetectNet, a novel lightweight and high-precision object detection network custom-designed for aerial remote sensing scenarios, building upon the YOLOv7-tiny algorithm. It enhances performance through four key improvements: the normalized Wasserstein distance for consistent object size sensitivity, the Involution module for reduced background noise, a self-designed RCS-biformer module for better spatial context interpretation, and a self-designed WF-CoT SPPCSP feature pyramid for improved feature map weighting and context capture. Ablation studies conducted on a hybrid dataset composed of three open-source remote sensing datasets (including NWPU VHR-10 remote sensing images, RSOD remote sensing images, and VisDrone UAV images) have demonstrated the effectiveness of four improvements specifically for small-size object detection. Visualizations through Grad-CAM further demonstrate AeroDetectNet’s capacity to extract and focus on key object features. Upon individual testing across three open-source datasets, AeroDetectNet has successfully demonstrated its ability to identify objects in images with a smaller pixel area. Through experimental comparisons with other related studies, the AeroDetectNet achieved a competitive mAP while maintaining fewer model parameters, highlighting its highly accurate and lightweight properties.

This letter proposes AccuLiteFastNet, a novel object detection model tailored for aerial remote sensing scenarios constructed on the YOLOv8. AccuLiteFastNet is designed to enhance the precision in detecting small-size objects while maintaining a lightweight model framework and supporting fast inference speeds for deployment. Specifically, we use the normalized Gaussian Wasserstein distance (NGWD) as the loss metric to evaluate the similarity between the predicted and ground-truth object bounding boxes, ensuring uniform sensitivity for objects of different sizes. Moreover, we replace the model backbone’s last feature extraction module with the self-designed multiscale FasterNet (MSFN) module to reduce the model complexity while providing an expanded receptive field. Finally, we propose a novel module named context aggregation-C2f (ContA-C2f) specifically designed to evaluate the informativeness of each pixel in the image. Across three open-source remote sensing datasets, AccuLiteFastNet demonstrates a significant improvement over the original YOLOv8, achieving a 4.5% increase in mean average precision (mAP). In comparison to existing object detection models, AccuLiteFastNet strikes an optimal balance with high inference speed (58.14 fps) and good accuracy (87.5% mAP) at lower computational costs (8.3 GFLOPs).

The objective of this study is to investigate the methods for soil liquefaction discrimination. Typically, predicting soil liquefaction potential involves conducting the standard penetration test (SPT), which requires field testing and can be time-consuming and labor-intensive. In contrast, the cone penetration test (CPT) provides a more convenient method and offers detailed and continuous information about soil layers. In this study, the feature matrix based on CPT data is proposed to predict the standard penetration test blow count N. The feature matrix comprises the CPT characteristic parameters at specific depths, such as tip resistance q_c, sleeve resistance f_s, and depth H. To fuse the features on the matrix, the convolutional neural network (CNN) is employed for feature extraction. Additionally, Genetic Algorithm (GA) is utilized to obtain the best combination of convolutional kernels and the number of neurons. The study evaluated the robustness of the proposed model using multiple engineering field data sets. Results demonstrated that the proposed model outperformed conventional methods in predicting N values for various soil categories, including sandy silt, silty sand, and clayey silt. Finally, the proposed model was employed for liquefaction discrimination. The liquefaction discrimination based on the predicted N values was compared with the measured N values, and the results showed that the discrimination results were in 75% agreement. The study has important practical application value for foundation liquefaction engineering. Also, the novel method adopted in this research provides new ideas and methods for research in related fields, which is of great academic significance.