Lightweight ResNet-U-Net Framework for Onboard Agricultural Boundary Delineation from Drone Imagery

Automated analysis of high-resolution Unmanned Aerial Vehicle (UAV) imagery is crucial for applications like precision agriculture. However, deploying deep learning models directly onto resource-constrained drone platforms for real-time processing remains a significant hurdle. This paper addresses this challenge by developing and evaluating an efficient pipeline for agricultural field boundary delineation. We employ a robust ResNet-based U-Net architecture to effectively handle complex visual features and field boundary delineation in challenging winter UAV imagery covering four different types of farmlands. While achieving an accuracy of 82%, our primary contribution lies in demonstrating the successful optimization of this segmented model into lightweight variants suitable for edge deployment, paving the way for future onboard processing. Initial tests on representative aerial data confirm that the optimized lightweight model successfully delineates field boundaries with an accuracy of 81% and a rapid inference time of approximately 1ms, underscoring its potential for edge execution. The optimized model represents a critical step towards autonomous, real-time boundary delineation directly on UAVs or mobile edge devices, paving the way for more responsive and intelligent aerial monitoring systems in agriculture and beyond.