Aiming at the problem of low accuracy in object detection caused by complex background, large scale changes, and dense small targets in remote sensing images, an improved YOLOv8n detection algorithm, MGL-YOLO, is proposed. Firstly, design MSConv to reduce model parameters and reconstruct C2f module based on MSConv to improve multi-scale feature extraction capability; secondly, based on GLSA and GSConv modules, the BiFPN structure is improved to simplify the neck network and enhance its feature processing capability. Design a Light head model at the head to further reduce weight and enhance the ability to extract small target features; finally, the NWD loss function is introduced to replace the original loss function and enhance the localization accuracy for small targets. Verified on the DIORR, DOTAv1.0, and VEDAI datasets, the experimental results show that the MGL-YOLO model has high accuracy on the DIOR-R dataset mAP@0.5 Improved by 1.7% and 1.3% compared to the benchmark model, 1.0% and 1.1% on the DOTAv1.0 dataset, and 2.4% and 2.1% on the VEDAI dataset. The parameter count has been reduced by 47%, and the computational complexity has been reduced by 32%. Compared with other small object detection algorithms, it has also achieved good detection performance.