A new deep-learning network can simultaneously identify and count female and male litchi flowers from images, a task that previously required time-consuming manual observation. Named FQGR-Net (Flower Quantification and Gender Recognition Network), the three-branch model produces separate density maps for each flower sex in a single inference pass. On a self-constructed dataset, it achieved average mean absolute error (MAE) of 8.498 and root mean squared error (RMSE) of 13.209 across categories, with coefficients of determination (R²) of 0.930 for female flowers and 0.971 for male flowers. A portable analyzer built around the model reached over 80% counting accuracy in orchard trials.
Litchi is a major tropical and subtropical crop in China, covering approximately 0.523 million hectares. The industry faces a pronounced alternate bearing pattern, with yields fluctuating sharply between "on-years" and "off-years" — 2024 production was estimated at 1.781 million tons, a 45.94% decrease from the previous year. Unstable floral initiation is the primary cause, especially in mid- to late-maturing cultivars. Because a high ratio of female to male flowers generally correlates with improved fruit set, rapid and accurate quantification of flower sex ratios during flowering is essential for targeted panicle management. Traditional manual counting is labor-intensive and prone to subjective errors, while existing automated methods mainly handle single-class detection or require detached floral samples, failing to support non-destructive monitoring in natural growth states. Because of these challenges, there is a need for in-depth research on multi-class, in-situ litchi flower classification and counting.
Researchers from the College of Engineering at South China Agricultural University (Guangzhou, China), the Guangdong Laboratory for Lingnan Modern Agriculture, and the National Key Laboratory of Agricultural Equipment Technology published (DOI: 10.1016/j.plaphe.2026.100217) the study in Plant Phenomics on 21 September 2026. The work presents FQGR-Net, a three-branch neural network for simultaneous classification and counting of female and male litchi flowers, along with a portable phenotype analyzer for orchard deployment.
FQGR-Net builds on a VGG16 backbone with three parallel branches: a foreground segmentation branch guided by a Frequency-Aware Dilated Module (FADM), a refined density regression branch using LiteIncepDWConv (LIDWC) and DySampleFusion, and a coarse density estimation branch with a Multi-Scale Fusion Module (MSFM). The FADM captures intra-class multi-scale features and suppresses background noise, while the MSFM aggregates hierarchical features from different network depths to address scale variations. The team also extended traditional point-level annotation to multi-class labeling, marking male flowers at stamen filament convergence centers and female flowers at stigma junctions, which improves occlusion resistance. The lightweight convolution and dynamic upsampling achieve a better trade-off between accuracy and model size, reducing parameters by 28.746% compared with standard convolutions and making edge deployment feasible. Ablation studies showed that combining MSFM and FADM improved mean absolute error (MAE)_A and root mean squared error (RMSE)_A by 35.616% and 35.663% over the baseline. The handheld analyzer integrates an NVIDIA Jetson Orin NX Super computing platform, an 8-megapixel IMX415 CMOS (complementary metal-oxide-semiconductor) sensor, and a 5.0-inch touchscreen, running models via ONNX Runtime (Open Neural Network Exchange Runtime) with CUDA (Compute Unified Device Architecture) acceleration. Field trials on 50 samples across multiple cultivars achieved 85.81% and 87.50% accuracy for female and male flower identification, respectively.
The authors said the work addresses a critical gap in automated floral census systems. They explained that accurate, non-destructive estimation of female-to-male flower ratios gives growers and breeders a practical tool for precision flower thinning, variable-rate management, and selection of parents with high fruit-set potential. They added that extending point-level annotation to multi-class counting improves deployment flexibility and could be adapted to other flowering crops with similar dense, small-target morphology, supporting broader applications in precision agriculture and yield estimation.
The technology could guide variable-rate flower-thinning machinery, adjusting thinning intensity to local flower density and reducing mechanical damage to remaining flowers. For horticultural research on plant growth regulators and floral sex differentiation, FQGR-Net enables rapid, reliable quantification of treated panicles, saving time and labor while reducing counting errors. The authors noted that combining panicle phenotypes with genome sequencing could help map quantitative trait loci (QTLs) related to flower number and female flower ratio, supporting breeding of offspring with high fruit-set rates and stress resistance. Future work will expand training data from overexposed and low-light conditions, incorporate 3D flower panicle reconstruction, and investigate multi-class counting in intercropped systems.
