AgriSafeNet: Leveraging RGB-TIR UAV Remote Sensing Imagery for Transmission Line Detection in Smart Agriculture

Abstract

In modern smart agriculture, the uncrewed aerial vehicles (UAVs) have emerged as indispensable low-altitude remote sensing platforms, particularly for precision irrigation and crop protection. While they enable the efficient acquisition of high-resolution spatial information and significantly enhance agricultural productivity, UAVs operating in complex environments face severe collision risks with power infrastructure, which poses a substantial threat to flight safety. Consequently, robust transmission line detection (TLD) is a critical prerequisite for ensuring safe autonomous UAV navigation. Despite the importance of this task, the existing detection methods often exhibit limitations in modeling fine-grained structural characteristics and fail to effectively integrate cross-modal contextual information. To address these challenges, this article proposes AgriSafeNet, a novel RGB-TIR TLD framework designed to bolster feature representation and multimodal fusion. Specifically, AgriSafeNet orchestrates efficient shallow-level cross-modal interaction through a multiscale feature decoupling and fusion module (MFDFM). Furthermore, it refines deep semantic integration via a spatial compression attention fusion module (SCAFM) and incorporates a context-guided feature extraction module (CGFEM) to capture intricate structural geometric details. This hierarchical design significantly improves the discriminability of transmission line (TL) features and the robustness of the model against environmental interference. Extensive experiments conducted on multiple benchmark datasets demonstrate that AgriSafeNet achieves the state-of-the-art performance, consistently outperforming the existing methods in complex agricultural scenarios.