MBHNet: a deep learning framework for UAV-satellite image matching in wheat field phenotyping

Key message By integrating multi-scale semantics and stable plot boundaries, MBHNet enables robust UAV–satellite alignment, supporting multi-source observation integration and high-throughput wheat field phenotyping. Abstract Precise image matching is an important condition for wheat field phenotyping and multi-source observation integration. However, conventional matching methods designed for urban scenes often fail in agricultural scenarios due to multi-view variations, temporal appearance variations, high texture similarity, and significant resolution discrepancies between UAV and satellite images. In this paper, we propose MBHNet, a deep learning framework for UAV-satellite image matching in wheat fields by integrating multi-scale semantic and boundary geometric features. First, a multi-scale semantic feature extraction (MSFE) strategy based on Vision Transformer (ViT) and dual-branch cross-attention is designed to capture comprehensive information across various levels. To mitigate the impact of temporal appearance variations and canopy similarity, a bidirectional feature aggregation-based boundary geometric feature extraction (BGF) block is designed to identify stable plot boundaries as reliable references. Furthermore, a heatmap segmentation-based feature matching (HFM) module is employed to ensure accurate registration and category-guided regional partitioning. Experimental results demonstrate that MBHNet significantly outperforms existing models such as FSRA and University-1652. This framework provides a spatial basis for integrating high-throughput field phenotypes in wheat research.