Martinez F. , ROMAINE, JAMES BRIAN, Johnson P. , Cardona A.
No
Proc. IEEE Int. Conf. Interdiscip. Approaches Technol. Manag. Soc. Innov., IATMSI
Conference Paper
Científica
0
0
01/01/2026
2-s2.0-105036981367
Sustained vegetable yields depend on systematic monitoring of crop development, especially in leafy species, where the leaf contour serves as an indicator of its growth status. Irregular and variable contours increase computational demand and energy consumption, which is problematic when inference must run on remotely deployed devices with tight resource budgets, particularly battery life. In this context, a custom performance equation ? is applied to evaluate performance during both training and inference, combining inference time and contour area error. This makes the approach directly applicable to agriculture. The metric associates performance with two weighted factors: computational cost and accuracy. The weightings depend on the application; for this case, 65% and 35% were used, respectively, giving greater importance to inference time. Inference results show 82.9% for YOLOv11, and 76.63% for Mask R-CNN, highlighting the balance between efficiency and precision in agricultural edge detection deployments. Data were acquired in the field using an IoT vision node built from inexpensive, low-power components. © 2026 IEEE.
Budget control; Energy utilization; Green computing; Image segmentation; Smart agriculture; Computational demands; Contour segmentation; Crop development; Detectron2; Energy-consumption; Performance; Resource budget; Segmentation; Segmentation models; YOLOv11; Agriculture; Computer vision