Researchers Create TillerPET AI Model for High-Throughput Phenotyping of Rice Tiller Traits

Tiller number and plant compactness are pivotal phenotypes determining panicle density, population structure, and yield formation. However, field measurement of these key traits has long been constrained by severe occlusion, uneven illumination, and the inefficiency of traditional manual assessment. High costs and complex workflows of automated or hardware-dependent imaging systems further impede high-throughput acquisition of these traits.

 

Recently, researchers from the Institute of Genetics and Developmental Biology, Chinese Academy of Sciences, together with Huazhong University of Science and Technology and Yazhou Bay National Laboratory, developed an AI model named TillerPET, which overcomes these longstanding limitations by enabling simultaneous in-situ high-throughput phenotyping of tiller number and compactness from post-harvest rice RGB images. The model demonstrates stable performance across multi-year, multi-location rice RGB datasets.

 

These findings were published online in The Crop Journal (DOI:10.1016/j.cj.2025.09.022) on November 7, 2025.

 

This study leverages a multi-year, multi-location rice RGB image dataset to propose the TillerPET, which adopts a point-query-based transformer architecture and incorporates a depth-aware rice region extraction module to build a lightweight feature extractor. The backbone of the point-query transformer is replaced with the Swin series, simplifying the original encoder design while substantially reducing computational load and simultaneously improving performance. TillerPET achieves an R² of 0.941 for tiller counting and an R² of 0.978 for measuring tiller compactness.

 

The tillering and architectural traits extracted by TillerPET further enable the classification and identification of rice varieties with different genotypes. In addition, the multi-year, multi-site phenotypic data on rice tillering and plant architecture provide valuable data support for rice ideotype breeding.

 

Figure. The technical pipeline of TillerPET, with thorough exposition of the proposed modules. (Image by IGDB)

 

The study was supported by the National Natural Science Foundation of China and the Hubei Provincial Natural Science Foundation.

 

Contact:

Dr. HU Weijuan

Institute of Genetics and Developmental Biology, Chinese Academy of Sciences

Email: wjhu@genetics.ac.cn