Work

The RP-PhenoUAS: Autonomous Uncrewed Aerial System for In-Field Rice Panicle Segmentation Based on RP-DeepLabV3+

Year: 2026

Type: article

Abstract: Rice panicle (RP) is a critical component determining yield, and in-field acquisition of panicle phenotypic characteristics can provide reliable data support for yield estimation, variety improvement, and precision field management. This article presents a custom-developed uncrewed aerial system, the RP phenotyping uncrewed aircraft system (RP-PhenoUAS), for in-field RP segmentation, integrating uncrewed aerial vehicle (UAV)-based image acquisition, lightweight edge intelligence, and an Internet of Things (IoT) communication framework. A lightweight segmentation model, RP-DeepLabV3+, was developed and deployed on both the RP phenotyping UAV (RP-PhenoUAV) and the handheld terminal. Experimental results show that RP-DeepLabV3+ achieves 93.41% mean Intersection over Union (mIoU) and 75.62% Boundary Intersection over Union (BIoU) with only 2.05M parameters and 41.69 GFLOPs. Compared with DeepLabV3+, it improves mIoU and BIoU by 4.91% and 11.18%, respectively, while reducing parameters and GFLOPs by over 93%. For edge deployment, the INT8-quantized RP-DeepLabV3+ running on Jetson Orin NX attains 91.24% mIoU and 74.16% BIoU, with an edge-side latency of 211 ms per full-resolution image. In addition, an IoT communication framework for robust field connectivity was developed and validated. Overall, the RP-PhenoUAS is a promising tool for RP segmentation and shows significant applicability in rice fields.

Source: IEEE Internet of Things Journal

DOI: 10.1109/jiot.2026.3726612