Real-time deep learning segmentation for agricultural applications: Insights for precision weed management

Recent advances in site-specific weed management (SSWM) have led to a notable shift toward segmentation methods, as opposed to bounding box detection. While detection models offered basic localization, segmentation approaches additionally quantify the spatial area of infestation, enabling more precise mapping of weed presence. This study aims to provide a systematic review of the recent use of segmentation models for real-time performance and their applicability in the weed management pipeline. A comprehensive literature review was conducted covering state-of-the-art segmentation architectures, sensing modalities, optimization strategies, and deployment platforms relevant to real-time SSWM. The key findings are summarized as follows a) Fewer studies deployed segmentation models on an edge device b) Emergence of Transformers over CNNs and the modification of YOLO-seg models for real-time applications c) Pruning and quantization using Tensor RT for redefining the structure of segmentation models to have a speed-accuracy trade-off d) Studies preferred YOLO over Mask R-CNN for UAV based instance segmentation e) Emergence of using DL models like SAM, Depth Anything v2, and Cycle GAN to solve dataset constraints f) panoptic segmentation remains largely unexplored in SSWM. Despite significant algorithmic progress, challenges related to edge deployment, system-level integration, and robustness under variable field conditions remain open. This review provides critical insights and future directions to support researchers, deep learning practitioners, weed scientists, and technology extension specialists in advancing segmentation-driven SSWM systems toward scalable field deployment.