New AI Model Enhances Wheat Spike Counting Accuracy to 95%

A recent study led by researchers from the Chinese Academy of Agricultural Sciences (CAAS) has introduced a new deep learning model that significantly improves the accuracy of wheat spike counting. By optimizing the selection of growth stages and enhancing the YOLOX-P model, the team achieved precision levels exceeding 95% while uncovering novel genetic loci associated with spike number.

Wheat is a vital component of the global food supply, providing essential carbohydrates, protein, and dietary fiber for billions. With the world’s population projected to grow significantly, wheat production must increase by at least 60% by 2050. Traditional breeding methods have primarily focused on enhancing yield through kernel size and quantity, but advancements in spike number—a complex trait influenced by genetics and environmental factors—have lagged.

Manual counting of spikes in field trials is a labor-intensive process that is often subjective and impractical for large-scale applications. In response, the research team harnessed advances in computer vision and deep learning to automate the detection of key plant traits. Despite existing challenges in achieving accuracy under real-world conditions, particularly with dense and overlapping spikes, the new model shows promising results.

The study, published in Plant Phenomics on May 13, 2025, involved the development and testing of spike number detection models using images collected across six developmental stages of the Zhongmai 175 × Lunxuan 987 recombinant inbred line (RIL) population. The team assessed ten models based on four performance indices: precision, recall, mean average precision (mAP), and F1 score.

Results indicated that the precision of spike counting was consistently high across models, ranging from 88.19% to 95.48%. However, recall varied significantly, from 44.60% to 81.23%. The late grain-filling stage consistently yielded the best results, achieving the highest recall rate of 81.23% and mAP values between 85.69% and 89.47%. Early stages were more challenging due to potential obstructions from leaves.

The study highlighted that the quality of manual annotations greatly influenced the accuracy of model predictions. The smallest discrepancies between manual and model counts were observed during the late grain-filling stage. Furthermore, strong correlations were identified between manual and model counts, particularly in models trained in Dezhou and tested in Changping.

To enhance detection capabilities, the researchers developed an improved YOLOX-P algorithm by integrating attention modules and utilizing higher-resolution input images. This modification resulted in a significant performance boost, increasing mAP by 5.30% to 5.99% and raising F1 scores by 0.06 compared to the original YOLOX model. Among the integrated datasets, the CD&DD subset trained with YOLOX-P achieved the highest mAP of 91.81%.

Validation of the models in the Zhongmai 578 × Jimai 22 population confirmed their robustness, with correlation coefficients ranging from 0.73 to 0.82 between automated and manual counts. Genetic analysis further revealed four stable quantitative trait loci (QTLs) linked to spike number: QSN.caas-4A2, QSN.caas-4D, QSN.caas-5B1, and QSN.caas-5B2. Kompetitive allele-specific PCR markers were developed for two of these loci, providing valuable tools for marker-assisted breeding.

The implementation of the YOLOX-P model represents a significant advancement in wheat breeding, combining artificial intelligence with plant genetics. Automated spike counting reduces reliance on labor-intensive field assessments while accelerating the phenotyping process. This not only offers objective, scalable data for breeding programs but also equips breeders with immediate tools to select varieties with higher spike numbers, directly contributing to yield improvements.

This groundbreaking research was funded by the STI2030-Major Projects (Project No. 2023ZD04076), the National Natural Science Foundation of China (No. 32372196), and the Beijing Joint Research Program for Germplasm Innovation and New Variety Breeding (G20220628002), along with additional support from the National Natural Science Fund of China (No. 32250410307).

The findings underscore the potential of integrating modern technology in agriculture to address the pressing challenges of food security and sustainable production practices worldwide.