DART
Open Access DepositedDomain Adaptive Robust Transfer Learning for Object Detection in Satellite Imagery
Detecting and identifying objects of interest in satellite images has wide rangingapplications in fields like national security, environmental monitoring, and urban planning. It is compounded by the challenging qualities of satellite images like birds-eye perspective, scale variations, and complex backgrounds. These characteristics render models that have been trained on natural images incapable of accurately transferring their knowledge to identify objects in satellite images. This praxis presents a framework, Domain-Adaptive Robust Transfer learning(DART), that improves object detection accuracy. DART builds on the "pre-train, fine- tune" paradigm of transfer learning by introducing domain-aware data augmentations and a domain adaptation strategy that utilizes dual regularization with distillation to preserve the knowledge of the pre-trained base models, thereby improving object detection accuracy in satellite images. The methodology used in this research systematically compares ConvolutionalNeural Network (CNN) and Vision Transformer (ViT) architectures on the large-scale DIOR benchmark satellite image dataset. Empirical results demonstrate the superiority of ViT-based backbones over CNN-based backbones with an 11-percentage point improvement in mean Average Precision (mAP) with baseline fine-tuning. Domain- aware data augmentation and knowledge-preserving domain adaptation added another four percentage points of gain, leading to a 16.14% relative improvement over conventional transfer learning methods. This study concludes that successful satellite imagery object detection requiresboth an architecture capable of modeling global context and a fine-tuning methodology that addresses domain shift while preserving learned knowledge.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
Chakravarthy_gwu_0075A_17639.pdf | 2025-12-12 | Open Access |
|