Automatic Target Detection from Satellite Imagery Using Machine Learning.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Object detection in satellite imagery is crucial for applications like precision agriculture, urban planning, and defense.
• Challenges include low pixel resolution, small object detection, class variations, and dense backgrounds.

Purpose of the Study:

• To compare the performance of deep learning object detection algorithms in satellite imagery.
• To analyze accuracy and speed of various CNN-based frameworks.

Main Methods:

• Developed a dataset of satellite imagery for object detection.
• Evaluated convolutional neural network (CNN) frameworks: Faster RCNN, YOLO (you only look once), SSD (single-shot detector), and SIMRDWN (satellite imagery multiscale rapid detection with windowed networks).

Main Results:

• SIMRDWN achieved 97% accuracy on high-resolution images; Faster RCNN achieved 95.31% on standard resolution.
• YOLOv3 reached 94.20% accuracy, while SSD achieved 84.61%.
• YOLO demonstrated superior speed (170-190 ms) compared to SIMRDWN (5-103 ms), though SIMRDWN is faster in some contexts, but fails in real-time surveillance.

Conclusions:

• Deep learning algorithms offer varying trade-offs between accuracy and speed for satellite imagery object detection.
• YOLO is the preferred choice for real-time applications due to its speed, while SIMRDWN excels in specific high-accuracy scenarios.