Abstract: The precise quantification of flowering traits in mung bean (Vigna radiata L.) plants holds significant importance for crop phenotyping and breeding research, with the quantitative characterization of flowering synchrony being particularly crucial as it directly influences population trait uniformity and varietal maturity consistency, thereby determining the selection of breeding materials and the evaluation of varietal adaptability. Accurate segmentation of flowers serves as the foundation for quantitative analysis. However, challenges such as the complex inflorescence structure of mung bean plants, the minute size of floral organs, their similar coloration to leaves and stems, coupled with intense lighting variations, extensive occlusions, and overlapping phenomena in field environments resulting in low contrast with the background, make it difficult for traditional semantic segmentation models to effectively address these issues. In recent years, although deep learning-based methods have made remarkable progress in image segmentation, they still face challenges such as insufficient feature extraction, loss of spatial information, and inadequate model generalization when processing small targets in complex agricultural scenarios. To address these challenges, this study utilizes high-resolution images of mung bean plants during the flowering and pod-setting stage acquired by an unmanned aerial vehicle (UAV) remote sensing platform and proposes an IDCA-UNet model that integrates a dual channel attention (DCA) mechanism and instance normalization (IN). Based on the classic U-Net architecture, the model employs the DCA module to dynamically aggregate features from global average pooling and max pooling, achieving adaptive feature enhancement in the channel dimension. This enables the model to focus on features with high discriminative power for flowers, effectively improving sensitivity and recognition capability for tiny flowers. Simultaneously, the introduction of Instance Normalization layers to replace some Batch Normalization layers mitigates the issue of unstable statistical estimation during small-batch training, reduces internal covariate shift, and enhances the model's robustness and generalization capability under complex and variable field conditions. Experimental results demonstrate that IDCA-UNet significantly outperforms various models including ResNet-UNet, YOLOv11, and DeepLabV3+ in the mung bean flower segmentation task, achieving mIoU, mAP, and F1 scores of 88.95%, 93.47%, and 93.79%, respectively. These represent improvements of 1.62, 1.48, and 0.72 percentage points compared to the VGG16-UNet benchmark model with Focal loss. Based on the high-precision segmentation results, this study further quantitatively analyzed the dynamic growth patterns of mung bean plants during the flowering and pod-setting stage. Through daily flower count statistics and fitting, it was found that the flowering dynamics exhibit a distinct "Λ"-shaped growth curve with a concentrated peak flowering period, indicating strong flowering synchrony within the population. This provides important evidence for assessing varietal maturity consistency and suitability for mechanical harvesting. This research not only addresses the specific challenge of mung bean flower segmentation but also provides a practical cross-platform application solution for organ-level phenotyping analysis in crops. In the future, integrating multi-source remote sensing data and temporal analysis could further elucidate the genetic and environmental regulatory mechanisms of flowering, thereby offering robust theoretical and technical support for accurate assessment of crop growth status, scientific formulation of management decisions, and high-yield, high-efficiency breeding in precision agriculture.