Deep Learning for Understanding Satellite Imagery: An Experimental Survey.

Area of Science:

• Geospatial analysis
• Computer vision
• Artificial intelligence

Background:

• Satellite imagery translation into maps is time-consuming and often inaccurate, especially during crises.
• Deep learning and recent datasets enable automated satellite image analysis.
• Automated mapping is crucial for disaster response and conflict monitoring.

Purpose of the Study:

• To introduce a competition and dataset for automated satellite image analysis.
• To develop and evaluate deep learning models for building detection in satellite imagery.
• To advance automated annotation techniques for geospatial data.

Main Methods:

• Development of five segmentation models, including U-Net and Mask R-Convolutional Neuronal Networks.
• Implementation of training adaptations such as boosting algorithms, morphological filters, and Conditional Random Fields.
• Utilization of custom loss functions for improved model performance.

Main Results:

• Models achieved high performance metrics, demonstrating effectiveness in building detection.
• The developed approaches show significant improvements over existing methods.
• Successful application of deep learning for automated satellite image annotation.

Conclusions:

• Deep learning is a feasible and effective approach for automated satellite image annotation.
• The introduced competition and models facilitate further research in automated geospatial analysis.
• AI-powered mapping can significantly improve response times and accuracy in disaster and conflict situations.