Cross-Domain Few-Shot  and Zero-Shot Road Detection from Satellite to UAV Imagery

Description

Acquiring high-quality annotated data for UAV-based road detection is often costly and time-consuming, particularly in remote or disaster-affected areas. This research leverages large-scale satellite embeddings (e.g., Google AlphaEarth Foundations or AnySat) to enable cross-domain transfer learning from satellite imagery to UAV imagery, allowing the recognition of narrow roads and complex terrain features under few-shot or zero-shot conditions.

By aligning satellite embeddings with UAV image features, the project aims to reduce reliance on UAV annotations while maintaining high detection accuracy across diverse environments and road types. The proposed approach also explores multi-modal and multi-scale feature fusion, combining optical, radar, and LiDAR data to enhance robustness in real-world scenarios.

Objectives

Cross-Domain Feature Alignment and Model Design

• Align satellite embeddings with UAV image features to create a unified feature space.
• Design a lightweight adapter network for efficient cross-domain knowledge transfer.
• Explore multi-scale and multi-modal feature fusion to improve detection accuracy on high-resolution UAV imagery.

Few-Shot and Zero-Shot Learning

• Few-Shot Learning: Achieve accurate road detection with only a limited set of UAV-labeled samples.
• Zero-Shot Learning: Leverage satellite embeddings to directly predict road features without UAV annotations, using prototype matching or embedding-based inference.

Performance Evaluation and Analysis

• Evaluate model accuracy under varying environmental conditions, road types, and UAV viewing angles.
• Compare performance with traditional UAV-based detection methods to quantify cross-domain transfer effectiveness and few-shot robustness.
• Conduct ablation studies on feature alignment strategies, adapter network designs, and multi-modal fusion approaches.

Innovation

• First integration of large-scale satellite embeddings with UAV imagery for few-shot and zero-shot road detection.
• Demonstrates cross-platform transfer learning capability across resolutions, modalities, and sensor types.
• Introduces multi-modal and multi-scale embedding fusion for robust detection in complex terrains.

Expected Contributions

• Provides a low-cost, annotation-efficient solution for UAV-based road detection.
• Advances research in cross-domain transfer learning, multi-modal fusion, and self-supervised embedding adaptation.
• Results can lead to publications in remote sensing, computer vision, and AI for Earth venues (e.g., CVPR, ICCV, NeurIPS Earth AI).

Requirements and Technical Skills

• Proficiency in Python and experience with deep learning frameworks (TensorFlow or PyTorch).
• Familiarity with deep learning models, transfer learning, few-shot and zero-shot learning, and domain adaptation.
• Basic experience in remote sensing, UAV image processing, or multi-modal geospatial data analysis.

The applicant will use publicly available datasets and receive technical support from HUN-REN SZTAKI.

External Thesis Supervisor from SZTAKI:  Prof. Tamás Szirányi,  sziranyi.tamas@sztaki.hun-ren.hu