Where are the fences? Large-scale fence detection using deep learning and multimodal aerial imagery

Abstract

Fences play a crucial yet often overlooked role in land use management, biodiversity preservation, and ecological connectivity. However, their fine-scale linear nature poses significant challenges for automated detection using traditional remote sensing approaches. In this study, we propose a deep learning-based method for large-scale fence detection using freely available multimodal remote sensing data. We leverage high-resolution orthophotographs combined with Digital Surface Models (DSM) to enhance the fences identification across diverse landscapes. This work makes two major contributions: the development and open release of a dedicated dataset for fence semantic segmentation, and a comprehensive ablation study evaluating multiple deep learning configurations on multimodal RGB and DSM imagery. Our findings indicate that fusing DSM with RGB data leads to improved segmentation accuracy, particularly in complex and vegetated areas. Additionally, the use of Binary Cross-Entropy (BCE) loss provides marginal performance gains over other loss functions, reinforcing its effectiveness for fine-scale object detection. However, these improvements remain relatively small when considering the significant computational cost associated with processing LiDAR-derived elevation data. Our results suggest that while DSM data can enhance fence detection, its use should be carefully evaluated based on the study area’s characteristics and available resources. In many cases, high-resolution orthophotographs alone provide a viable and scalable alternative for detecting fences at a national scale. We systematically evaluate the impact of different experimental parameters, including sampling strategies, data normalization techniques, and loss functions, highlighting the importance of methodological choices in optimizing model performance. Future work should explore the classification of LiDAR point clouds or high-resolution drone imagery to further enhance fence detection capabilities while optimizing computational efficiency. The code and the dataset are freely available on Zenodo (https://zenodo.org/records/13902550) and Github (https://github.com/r-wenger/MultiFranceFences).