Related Experiment Video
Updated: Jul 23, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
581
State-of-the-Art Deep Learning Methods for Objects Detection in Remote Sensing Satellite Images
Adekanmi Adeyinka Adegun1, Jean Vincent Fonou Dombeu2, Serestina Viriri1
1School of Mathematics, Statistics and Computer Science, University of KwaZulu-Natal, Durban 4041, South Africa.
Sensors (Basel, Switzerland)
|July 14, 2023
Summary
Deep learning object detection in satellite images faces challenges. This study evaluated YOLO and R-CNN models on diverse landscapes, achieving over 90% accuracy for vegetation and swimming pools with YOLOv8 showing the fastest detection speed.
Area of Science:
- Remote sensing and geospatial analysis.
- Computer vision and deep learning applications.
- Environmental and urban monitoring.
Background:
- Object detection in satellite imagery is vital for disaster prevention, resource management, and urban planning.
- Deep learning models struggle with heterogeneous landscapes, similar object classes, and diverse object appearances in remote sensing data.
- Acquiring representative training data for complex remote sensing scenarios remains a significant challenge.
Purpose of the Study:
- To evaluate the performance of deep learning-based object detection algorithms on remotely sensed satellite images.
- To address challenges in object detection posed by landscape heterogeneity and data limitations.
- To compare various object detection models, including R-CNN and YOLO architectures, using a newly created dataset.
Main Methods:
- Utilized multi-object detection deep learning algorithms combined with a transfer learning approach.
- Developed and employed a new dataset of diverse features from Google Earth Engine in southern KwaZulu-Natal, South Africa.
- Experimentally evaluated five object detection methods based on R-CNN and YOLO architectures on the new dataset and two public datasets (Visdron, PASCAL VOC2007).
Main Results:
- Comprehensive performance evaluation and analysis of recent deep learning object detection methods were conducted.
- Highest detection accuracy exceeding 90% was achieved for vegetation and swimming pool instances.
- The YOLOv8 model demonstrated the fastest detection speed, recorded at 0.2 milliseconds.
Conclusions:
- Deep learning object detection models show promise for analyzing high-resolution remote sensing imagery.
- The study highlights the effectiveness of YOLOv8 in terms of both accuracy and speed for specific object classes.
- Transfer learning and tailored datasets are crucial for improving object detection performance in complex remote sensing applications.
