A method for detecting small objects in remote sensing images based on feature enhancement and scale-adaptive optimization in complex scenarios
- 1Army Engineering University of PLA, Nanjing, China
- 2State Key Laboratory of Advanced Communication Technology, Nanjing, China
- 3Chinese People’s Liberation Army Unit, Beijing, China
Abstract
To address the challenges of small object detection in remote sensing images, such as susceptibility to background noise and difficulties in feature extraction, this paper proposes a method for detecting small objects in remote sensing images based on feature enhancement and scale-adaptive optimization under complex scenarios. First, we constructed a bidirectional complementary feature fusion network, employing a bidirectional feature fusion mechanism to enhance the interaction between shallow-level details and deep-level semantic features, thereby enabling more effective multi-scale feature fusion. Second, we developed a Multi-scale Local-Global Attention (MLGA) module that uses multi-scale pooling to capture both local details and global context, and employs dynamic weights for adaptive feature fusion, effectively suppressing background noise and highlighting the features of small objects. In addition, we designed a scale-adaptive bounding box regression loss function (SAWIoU) that addresses the scale bias issue for small objects by introducing a scale scaling factor, thereby ensuring balanced optimization across objects of different scales and further improving localization accuracy. Finally, experimental results on two public remote sensing datasets DIOR and RSOD show that the proposed method achieved mAP50 scores of 0.885 and 0.971, respectively, significantly outperforming current mainstream detection models, with particularly notable improvements in small-object detection accuracy. Additionally, ablation studies further validate the effectiveness and robustness of this method in detecting small objects in remote sensing under complex backgrounds.
Related
- Optimization of zoned irrigations to balance groundwater sustainability and crop yield under future irrigation source replacement scenarios
- An Artificial Intelligence Integrated Geospatial Decision-Support Framework for Scenario-Based Climate-Adaptive Land-Use Optimization and Environmental Risk Reduction
- Adaptive Inversion Optimization of the Static Factor of Safety for Newmark-Based Earthquake-Induced Landslide Susceptibility Assessment
- Conditional-Risk Optimization and KKT Geometry of a Regime- and Horizon-Adaptive Probabilistic Ensemble for Local Weather Forecasting
- Targeting integrated daylighting comfort in deep office by multi-objective optimization using partial vertical deployment of adaptive corrugated shadings
BibTeX
@article{wang2026method,
title = {{A method for detecting small objects in remote sensing images based on feature enhancement and scale-adaptive optimization in complex scenarios}},
author = {Xiaoxiao Wang and Xia Zou and Meng Sun and Chong Jia and Yongqiang Xie and Xiongwei Zhang},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0350873},
url = {https://doi.org/10.1371/journal.pone.0350873},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A method for detecting small objects in remote sensing images based on feature enhancement and scale-adaptive optimization in complex scenarios
AU - Xiaoxiao Wang
AU - Xia Zou
AU - Meng Sun
AU - Chong Jia
AU - Yongqiang Xie
AU - Xiongwei Zhang
PY - 2026
DA - 2026/10/27
JO - PLOS One
DO - 10.1371/journal.pone.0350873
UR - https://doi.org/10.1371/journal.pone.0350873
AB - To address the challenges of small object detection in remote sensing images, such as susceptibility to background noise and difficulties in feature extraction, this paper proposes a method for detecting small objects in remote sensing images based on feature enhancement and scale-adaptive optimization under complex scenarios. First, we constructed a bidirectional complementary feature fusion network, employing a bidirectional feature fusion mechanism to enhance the interaction between shallow-level details and deep-level semantic features, thereby enabling more effective multi-scale feature fusion. Second, we developed a Multi-scale Local-Global Attention (MLGA) module that uses multi-scale pooling to capture both local details and global context, and employs dynamic weights for adaptive feature fusion, effectively suppressing background noise and highlighting the features of small objects. In addition, we designed a scale-adaptive bounding box regression loss function (SAWIoU) that addresses the scale bias issue for small objects by introducing a scale scaling factor, thereby ensuring balanced optimization across objects of different scales and further improving localization accuracy. Finally, experimental results on two public remote sensing datasets DIOR and RSOD show that the proposed method achieved mAP50 scores of 0.885 and 0.971, respectively, significantly outperforming current mainstream detection models, with particularly notable improvements in small-object detection accuracy. Additionally, ablation studies further validate the effectiveness and robustness of this method in detecting small objects in remote sensing under complex backgrounds.
SN - 1932-6203
ER -