Summer Research Project Grant Report
Cross-disciplinary Domains: Agricultural Science and Computer Science
Research Project Objective
This project fine-tunes SAM2.1, a state-of-the-art instance segmentation model, to automatically identify and quantify individual weed instances across morphologically diverse monocotyledon and dicotyledon species prevalent in Australian cropping systems. Using synthetic multi-species canvases generated from two complementary datasets, the model is evaluated against out-of-the-box baseline performance to quantify the benefit of domain-specific adaptation. The model developed here will directly contribute to weed assessment pipelines within the “Next-gen multi-layer analytics to improve profit from integrated weed management” (UWA2307-005RTX) nationwide project and establish the groundwork for a low-cost, automated phenotyping software for Australian weeds.
