Consistent and efficient sorghum analysis via UAV imagery with a unified multitask evaluation framework
TL;DR
This study proposes a unified multi-task evaluation framework for sorghum detection and segmentation in UAV imagery, and compares several recent CNN-based and Transformer-based models, including YOLOv8, YOLO11, RT-DETR- L, Swin Transformer, and MobileNetV3.
Abstract
Unmanned Aerial Vehicles (UAVs) have become widely used in agricultural monitoring and precision agriculture. In real field conditions, however, UAV-based analysis is often affected by dense planting patterns, background interference, and unclear boundaries, which make reliable computer vision analysis more difficult. To address this problem, this study proposes a unified multi-task evaluation framework for sorghum detection and segmentation in UAV imagery. Using this framework, we compare several recent CNN-based and Transformer-based models, including YOLOv8, YOLO11, RT-DETR- L, Swin Transformer, and MobileNetV3. The results show that RT-DETR-L achieved the best detection performance, with an mAP@0.5 of 0.915, while YOLO11s-seg achieved the best segmentation result, with an mIoU of 0.923. In addition to standard benchmarking, this study introduces two extra evaluation metrics: Region Consistency Score (RCS), which measures region containment between detection and segmentation outputs, and Unified Performance Evaluation (UPE), which combines accuracy and efficiency into a single deployment-oriented measure. Because detection and segmentation rely on different task-specific metrics, UPE is used here as a cross-task comparison from a deployment perspective rather than as a strict within-task ranking. Under this setting, the unified multi-task framework shows strong cross-task region containment and good deployment-oriented performance, providing a more practical basis for agricultural model selection.