Tackling Environmental Variability: Few Shot Segmentation for Domain-Adaptive Weed Segmentation in Agricultural Robotics

Nico Catalano, Monica Leone, Matteo Matteucci

Abstract: This paper investigates the application of Few Shot Segmentation techniques in addressing weed management challenges within agricultural robotics. Traditional methods, such as indiscriminate herbicide spraying, pose environmental and economic concerns, emphasize the necessity of developing automated precision procedures. Semantic Segmentation models can offer significant advancements in weed management within agricultural robotics but can be negatively impacted by domain shifts, often requiring adaptation, hindering the development of the field. In this context, we propose a Few Shot Segmentation-based approach for precise weed segmentation in corn and bean cultivations. Our model, adapted from PFENet, exhibits robust performance across varying environmental conditions, mitigating domain shift issues. Evaluation on the ROSE dataset demonstrates the model’s reliability and generalizability, highlighting the potential of Few Shot Segmentation in advancing agricultural robotics technologies.