Accurate determination of crops growth period plays an important role in field management and agricultural decision-making. The current work mostly extracts the crop normalized difference vegetation index curve from multi-temporal data and identifies the crop phenology based on its trend or special nodes. However, these time-series-based identification methods are difficult to be applied to practically crop monitoring tasks. In this paper, the unmanned aerial vehicle remote sensing platform is used to collect the multi-spectral images of the experimental field and identify the sunflower growth period based on the different population features during its different growth periods. According to the actual field management needs, this study obtains the plot-level sunflower growth period result by analyzing statistically the distribution area of different sunflower periods in a field plot. This study uses the data of 2018 in the study area to build the model and test its performance on the data of 2019. Through comparative experiments, PSPNet can achieve a good balance between accuracy and efficiency in this study. Further, given to time-series relationship between the adjacent growth periods classification, this paper proposes an improved loss function to weight different types of misclassification to optimize model performance. The results show that improved PSPNet with proposed weighted loss function achieves the optimal recognition accuracy of 89.01%, which provides a solution for sunflower growth period recognition based on the single-phase data.

准确测定农作物生长期对田间管理和农业决策具有重要作用。目前的工作主要是从多时相数据中提取作物归一化差异植被指数曲线，并根据其趋势或特殊节点识别作物物候。然而，这些基于时间序列的识别方法很难应用于实际的作物监测任务。本文利用无人机遥感平台采集试验田的多光谱图像，根据向日葵不同生育期的不同种群特征识别其生育期。根据现场实际管理需要，本研究通过对田间样地不同时期向日葵分布面积的统计分析，得出样地级向日葵生育期结果。本研究使用研究区2018年的数据建立模型，并在2019年的数据上测试其性能。通过对比实验，PSPNet在本研究中取得了准确率和效率的良好平衡。此外，考虑到相邻生长期分类之间的时间序列关系，本文提出了一种改进的损失函数来对不同类型的误分类进行加权，以优化模型性能。结果表明，采用加权损失函数的改进PSPNet实现了89.01%的最优识别准确率，为基于单期数据的向日葵生长期识别提供了解决方案。