A team led by Prof. JIANG Ni from the Institute of Genetics and Developmental Biology, Chinese Academy of Sciences (CAS), has developed a lightweight deep learning model that enables precise assessment of freeze injury in wheat.
The study was published in Plant Phenomics (https://doi.org/10.1016/j.plaphe.2025.100061), which introduces FreezeNet, a model designed to address the challenges of field-based freeze injury evaluation in winter wheat.
Winter wheat is highly susceptible to low-temperature stress, which reduces yields and complicates breeding for freeze tolerance. Traditional manual evaluations are subjective and inconsistent, while UAV-mounted multispectral imaging often struggles with background noise such as exposed soil and withered stems. This has created a need for accurate, robust, and field-adaptable assessment tools.
FreezeNet combines shallow texture features with deep spatial semantics through a dual-attention mechanism, achieving high segmentation accuracy (Dice 89.95%, IoU 81.85%) while reducing parameter size by 97% and computational cost to 1/42 of compared with UNet. This efficiency makes it suitable for mobile deployment in agricultural fields.
Applied to field phenotyping, FreezeNet significantly improved the evaluation of freeze injury indices such as the yellow vegetation fraction (YVF). Genetic mapping further identified 11 quantitative trait loci (QTLs) linked to freeze tolerance, eight of which were associated with key CBF genes. Incorporating favorable alleles reduced YVF from 0.88 to 0.48, underscoring the potential to accelerate breeding for cold tolerance.
The research establishes a foundation for integrating AI-based phenotyping with molecular breeding, offering new technical pathways and genetic markers for improving freeze tolerance in wheat.
This work was supported by the National Key Research and Development Program of China and the Biological Breeding-National Science and Technology Major Project.
Comparison between manual freeze injury grading and model-extracted image features. From left to right, the five columns correspond to the five levels determined by manual evaluation (Level 1–5, with severity increasing progressively); from top to bottom, each row shows: the original input image, plant segmentation results, yellow leaves regions, and green leaves regions. (Image by IGDB)
Contact:
Institute of Genetics and Developmental Biology, Chinese Academy of Sciences
Email: njiang@genetics.ac.cn
CAS
中文
.png)
Comparison between manual freeze injury grading and model-extracted image features. From left to right, the five columns correspond to the five levels determined by manual evaluation (Level 1–5, with severity increasing progressively); from top to bottom, each row shows: the original input image, plant segmentation results, yellow leaves regions, and green leaves regions. (Image by IGDB)Contact:Prof. JIANG NiInstitute of Genetics and Developmental Biology, Chinese Academy of SciencesEmail: njiang@genetics.ac.cn