YOLOv8-OmniDet: Enhancing Object Detection in Remote Sensing with Improved Localization and Multi-Scale Feature Extraction

Object detection in satellite imagery presents unique challenges due to class imbalance, complex backgrounds, varying object scales, dense and sparse object distributions, and resolution limitations.To address these, we propose YOLOv8-OmniDet, an optimised extension of YOLOv8, specifically designed for remote sensing applications.The term YOLOv8-OmniDet is used to denote our customized variant of YOLOv8, designed for omnidirectional multi-scale detection in remote sensing imagery.The model incorporates a custom backbone for improved feature extraction, a feature pyramid network (FPN) -based neck for enhanced multi-scale detection, and a custom detection head optimized for precise classification and bounding box regression.Trained on 7,000 annotated satellite images and evaluated on a separate benchmark of 2,000 images spanning 10 object classes, YOLOv8-OmniDet achieves a mean average precision (mAP) of 86.5% with an inference time of 26.2 milliseconds per image.On an independent test set of 1,000 images, the model attained a precision of 70.17% and sensitivity of 71.23%, outperforming the baseline YOLOv8 in class-wise accuracy and small-object localization.These results confirm the model's effectiveness in handling the diverse characteristics of satellite imagery and highlight its suitability for tasks such as urban monitoring, land use classification, and environmental surveillance using deep learning-based object detection.