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

Lima Bean Yield Prediction Using Robotic Multi-View Imaging and Deep Learning

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

Citation:  2026 ASABE Annual International Meeting  2600633.(doi:10.13031/aim.202600633)
Authors:   Ashish Reddy Mulaka, Emmalea Garver Ernest, Yin Bao
Keywords:   3D Gaussian Splatting, 3D-PodTracker, heat-resilient, instance matching.

Abstract. Lima bean (Phaseolus lunatus) production in the Delmarva region is threatened by high night temperatures, necessitating the rapid development of heat-resilient cultivars. Developing these varieties requires accurate yield evaluation; however, current measurement process is labor-intensive and time-consuming. Because pod count is a primary yield component, utilizing an automated counting pipeline as a non-destructive proxy would significantly accelerate the screening process. While 2D Multi-Object Tracking (MOT) offers automated counting, complex canopy architectures trigger severe occlusion and re-identification failures, resulting in systemic overcounting. Furthermore, 3D phenotyping techniques rely heavily on density-based clustering, which inherently fails to separate physically touching pods. To address these limitations, we propose the 3D-PodTracker, a pod counting pipeline utilizing 3D Gaussian Splatting (3DGS). By back-projecting 2D instance masks (YOLOv12L and SAM2) into high-fidelity 3DGS point clouds, the system utilizes an instance matching module that explicitly isolates individual pods without relying on volumetric clustering. The 3D-PodTracker vastly outperformed the state-of-the-art 2D and 3D baselines. The pipeline reduced the counting Root Mean Squared Error (RMSE) from 83.76 to 11.51. Despite high algorithmic accuracy, the predicted 3D pod count demonstrated a weak Pearson correlation (r=0.350) with the physically harvested yield, highlighting the inherent challenges of extreme in-field canopy occlusion. Ultimately, while this framework establishes a highly reliable method for algorithmic pod counting, further research focuses on a subset of harvested plants for reduced occlusion, to establish proxy for yield and accelerate climate-resilient crop development.

(Download PDF)    (Export to EndNotes)