Satellite Imagery-Based Mapping Solutions for Autonomous Navigation in Outdoor Settings

This study investigates the use of satellite imagery for robotics applications, specifically focusing on evaluating a hybrid satellite image segmentation method against conventional soft-computing techniques. The research is conducted in Ankara, where a satellite image is processed and segmented into road, building, vegetation/forest, and ground categories using several methods. Supervised methods are trained on sample images from the city district satellite image and compared to manual segmentation conducted in Adobe Photoshop. Results from the Feed-forward Neural Network (FNN) and Probabilistic Neural Network (PNN) are contrasted with the Hybrid method, which demonstrates robust performance in identifying features such as buildings, roads, ground, and forest. The hybrid method achieves a higher Kappa coefficient (0.4060) compared to FNN (0.3908) and PNN (0.3757), indicating superior segmentation accuracy. Challenges persist in distinguishing similar color spectrums, particularly between ground and other classes. Future research directions involve refining color-based feature extraction methods and exploring advanced machine learning techniques to enhance segmentation accuracy, especially in complex urban environments. The application of satellite imagery holds great potential for robotic mapping and navigation, with opportunities to integrate multi-modal data sources and deploy deep learning architectures for more reliable performance. As researchers continue to innovate in satellite image analysis, the prospects for transformative advancements in robotics applications remain compelling.