Abstract:
Applications like urban planning, environmental monitoring, disaster management,
and security require detecting small objects in satellite images. This work aims to
tackle the issue of low visual information and high background noise that challenges
small object detection. Most previous work uses conventional object detectors, which
encounter challenges, especially under long tail distribution and having high overlap
between different categories. This limitation was addressed by developing a comprehensive dataset (consisting of satellite images) where objects are depicted at different
resolutions and sizes. The data augmentation methods used were very elaborate to
increase the diversity and generalizability of the dataset. Current state-of-the-art deep
learning based object detection models, such as, You Only Look Once (YOLO) object
detection models. i.e. YOLOv5, YOLO6, YOLO7 and YOLO8 were chosen for training and evaluation. These models leverage sophisticated techniques such as feature
pyramid networks and anchor boxes, aiding in more accurate detection. The methodology was built on dynamic data fusion and transfer learning, which made the models
adapt to smaller objects much better. To assess the models, a range of metrics were
used, such as, confusion matrix, precision-recall curve, F1 score, and mean average
precision (mAP). The results show that deep learning models can identify small objects from satellite images with complexity. Moreover, with a mAP@0.5 of 0.94 and
mAP@0.5-0.95 of 0.64, YOLOv8 was found to be the best performing model. That
said, there are still challenges concerning both model generalization and coping with
high intra-class variance. This study offers important guidance for using deep learning techniques for small object detection and points out some possible future works in
improving automated monitoring systems and remote sensing accuracy.
