针对河湖“四乱”基层巡查移动终端资源有限,传统深度学习检测模型参数量大、计算复杂度高难以部署的问题,提出基于YOLOv8n的轻量型遥感目标检测算法FBL-YOLO。该算法通过三级优化实现精度与效率的平衡:主干网络采用 FasterNet单元构建FasterC2f轻量型特征提取模块,颈部网络融合MAFPN与BiFPN优势设计BIMAFPN强化多尺度特征融合能力,检测头引入基于组归一化与共享卷积的LSCD轻量化结构精简参数规模。试验结果显示,FBL-YOLO在河湖“四乱”自建数据集上的参数量Params与浮点计算量FLOPs较基线模型分别降低50%和33.3%,mAP50提升2.8%,FPS达93.8帧/s;在NWPU VHR-10公共数据集上泛化性良好,可满足移动设备实时检测需求。

[Objective] Remote sensing images of river and lake scenes are characterized by dense small objects, complex backgrounds, and diverse target scales, which require models with both precise feature extraction capability and lightweight efficiency. This study develops a lightweight remote sensing object detection algorithm based on YOLOv8n, named FBL-YOLO, to achieve a balance between detection accuracy and model efficiency while meeting the real-time detection requirements of mobile devices for the supervision of river and lake “four chaos”. [Methods] FBL-YOLO achieved three-level architectural optimization based on the YOLOv8n framework. (1) Backbone optimization: The bottleneck structure in the C2f module of YOLOv8n was replaced with an improved FasterNet unit to construct the FasterC2f lightweight feature extraction module. (2) Neck optimization: A BIMAFPN structure was designed by integrating the advantages of multi-branch auxiliary feature pyramid network (MAFPN) and bidirectional feature pyramid network (BiFPN). (3) Head optimization: A lightweight shared convolutional detection head (LSCD) was introduced, which employed shared convolution mechanisms to capture multi-scale image information uniformly. Experimental evaluations were performed on both a self-built river and lake “four chaos” dataset (3 311 images and 6 958 suspected issue patches, covering three categories: illegal construction, occupation, and stacking) and the public NWPU VHR-10 dataset. [Results] (1) Ablation experiments: Compared with the baseline YOLOv8n, FBL-YOLO integrating all three modules achieved a 50% reduction in parameter count (from 3.01 MB to 1.5 MB) and a 33.3% reduction in FLOPs (from 8.1 GB to 5.4 GB), with mAP50 increasing by 2.8% (from 56.1% to 58.9%). (2) Category-specific performance: FBL-YOLO showed significant improvements in detecting small and complex targets. The mAP50 for “illegal stacking” (small and scattered targets) increased by 8.4 percentage points (from 24.5% to 32.9%), and that for “illegal occupation” improved by 1.6 percentage points (from 69.7% to 71.3%), while that for “illegal construction” decreased slightly by 1.5 percentage points. (3) Comparison with state-of-the-art algorithms: FBL-YOLO outperformed other lightweight models (e.g., YOLOv5n, YOLOv9t, YOLOv10n, NanoDet-Plus, PP-PicoDet) in overall performance, achieving the highest mAP50 (58.9%) with the smallest parameter count (1.5 MB). (4) Generalization performance: On the NWPU VHR-10 dataset, FBL-YOLO achieved an overall mAP50 of 90.2%, which was 5.6 percentage points higher than that of YOLOv8n. Significant improvements were observed in categories such as storage tanks and ground track fields, verifying its strong cross-scene adaptability. [Conclusion] The proposed FBL-YOLO algorithm effectively balances detection accuracy and model compactness through three-level optimization of the backbone, neck, and head. It reduces the parameter size and computational complexity by 50% and 33.3%, respectively, while improving mAP50 by 2.8% on the self-built river and lake “four chaos” dataset, particularly enhancing the detection performance for small targets such as “illegal stacking”. Its excellent performance on the NWPU VHR-10 dataset further confirms its generalization capability. FBL-YOLO meets the real-time deployment needs of mobile terminals in grassroots inspections, providing technical support for automated supervision of river and lake “four chaos”. Future research should focus on improving robustness under extreme weather conditions, exploring multimodal data fusion, and realizing real-time model updates at the edge to support comprehensive watershed management.