The Use of Deep Learning Methods for Object Height Estimation in High Resolution Satellite Images.

Area of Science:

• Remote Sensing
• Photogrammetry
• Computer Vision

Background:

• Extracting detailed information from single high-resolution satellite images is challenging due to spectral heterogeneity.
• Traditional image processing methods are insufficient for accurate feature extraction from complex urban landscapes.

Purpose of the Study:

• To develop a universal, fully automated algorithm for estimating object heights using high-resolution optical satellite data.
• To overcome limitations of classical methods by employing deep learning for shadow analysis.

Main Methods:

• Utilized deep learning algorithms for automated object and shadow detection in satellite and aerial imagery.
• Calculated object heights by analyzing shadow lengths and incorporating metadata like sun elevation and satellite azimuth angles.
• Validated the algorithm's performance against LiDAR data in Warsaw, Poland.

Main Results:

• Achieved a global accuracy of ±4.66 m in height estimation across several hundred thousand objects.
• Demonstrated the algorithm's capability to accurately measure heights of both typical (buildings) and atypical (wind turbines) objects.
• Proposed a set of algorithms for iterative analysis of object-shadow relationships.

Conclusions:

• Deep learning-based shadow analysis is effective for automated height estimation from single satellite images.
• The developed method offers a versatile solution for urban landscape analysis and object inventory.
• The algorithm shows significant potential for applications requiring precise 3D information from remote sensing data.