Enhancing rice phenology identification by synergistic learning canopy optical signals and plant height dynamics

Accurate monitoring of rice phenological transitions plays a pivotal role in enhancing breeding efficiency and optimizing agronomic practices. Current spectral-based approaches frequently encounter limitations in detecting subtle growth stage boundaries within large-scale breeding programs, particularly due to visually imperceptible canopy variations during critical transitional phases. To address this issue, this study introduces a deep learning framework named GrowAI that synergistically combines dynamic plant architecture parameters with hyperspectral canopy signatures for robust phenological identification. Through a two-year breeding experiment, we established a time-series multispectral image dataset covering complete growth cycles. Our methodology innovatively integrates three-dimensional plant height dynamics with canopy optical properties through multimodal fusion architecture. Experimental results demonstrated GrowAI's superior performance, achieving classification accuracies of 0.937 (OA) and 0.927 (F1-score), representing average improvements of 6.9 % and 7.0 % respectively over conventional full-spectrum deep learning approaches. Notably, the framework exhibited exceptional temporal generalizability with cross-year validation accuracy reaching 0.977. Moreover, by accurately tracking the phenological stages of different rice genotypes in the breeding trials, the GrowAI framework can help breeders identify climate-resilient cultivars that have the most suitable phenological characteristics.