ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Auto-Harvest: An End-to-End Automated Cyberinfrastructure for Precision Agriculture Workflows

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600653.(doi:10.13031/aim.202600653)
Authors:   Harikesh Byrandurga Gopinath, Parth Gupta, Vedant Patil, Christopher Dean, Anirudh Potlapally, Hari Subramoni, Scott Shearer, Ramarao Venkatesh, John Fulton, Christopher Stewart
Keywords:   Precision agriculture, Spot spraying, UAV weed detection, High-performance computing, Workflow automation.

Abstract. Machine learning (ML) and Computer vision (CV) are enabling new levels of automation in precision agriculture, especially for critical field operations such as weed detection and spot spraying. This paper presents AUTO-HARVEST, an end-to-end cyberinfrastructure framework that connects UAV mission planning, near-edge data validation and batching, asynchronous HPC job orchestration, perception, and geospatial prescription generation into a repeatable workflow for precision spraying. Geo-referenced imagery collected by a UAV is staged on a near-edge device, validated, transferred in batches to an HPC environment, and processed through a two-stage perception pipeline using OWLv2 for open-vocabulary detection and DINOv3 for exemplar-based classification. The resulting object-level outputs are converted into geospatial prescription layers and shapefiles that can support downstream visualization and variable-rate application. The framework is designed to reduce operational complexity by enabling minimal-interaction execution across distributed compute resources. In the demonstrated workflow, detection optimization reduced average inference runtime from 0.86 s per image to 0.19 s per image, while classification achieved 95.34% overall accuracy and micro-F1 on 343 crop instances across three classes. Although the current study does not yet quantify final spray-placement accuracy in field coordinates, it demonstrates the feasibility of integrating AI, HPC, and geospatial processing into a practical workflow for precision weed management.

(Download PDF)    (Export to EndNotes)