考虑全生长周期的吊蔓西瓜表型识别方法研究

Phenotypic Identification Method for Whole Growth Cycle of Hanging Watermelon

摘要:

针对当前作物表型研究大多关注单一生长阶段表型特征，无法准确监测植物全生长周期长势等问题，以吊蔓西瓜为研究对象，提出了结合多种深度学习方法及机器视觉技术的吊蔓西瓜全生长周期关键表型参数高精度识别方法。其中，在幼苗期，分别建立了基于Leaf SAM的叶面积计算模型和基于Xception的叶片数计算模型，试验结果表明，叶面积、叶片数计算模型决定系数分别为0.96、0.98，均方根误差分别为2.98 cm2、0.14；在伸蔓期，分别建立了基于YOLO v5和双目视觉原理的株高计算模型和基于OpenCV的茎粗计算模型，试验结果表明，株高、茎粗计算模型决定系数分别为0.94、0.92，均方根误差分别为4.18 cm、0.17 mm；在坐果期和成熟期，构建了基于UNet的果实投影面积计算模型，试验结果表明，果实面积投影模型决定系数和均方根误差分别为0.99、9.85 cm2。上述结果表明模型计算值与人工测量值间的线性关系显著，综合误差较低，能够有效计算吊蔓西瓜全生长周期的关键表型参数。

Abstract:

In response to the current focus on single growth stage phenotype characteristics in most phenotype research, which makes it difficult to accurately monitor plant growth throughout whole growth cycle, a high-accuracy identification method for key phenotypic parameters of hanging watermelon throughout whole growth cycle was proposed by combining multiple deep learning methods and machine vision technologies. In the seedling stage, leaf area calculation model based on Leaf SAM and leaf count model based on Xception were established, and the experiments results showed that the R2 of the leaf area and leaf count models were 0.96 and 0.98, and the root mean square error（RMSE）was 2.98 cm2 and 0.14, respectively. During the elongation period, plant height measurement model based on YOLO v5 and binocular vision principles, as well as stem thickness calculation model based on OpenCV, were established separately, and the experiment results showed that the R2 of the plant height and stem thickness measurement model were 0.94 and 0.92, and the RMSE was 4.18 cm and 0.17 mm, respectively. In the fruiting and ripening stages, a fruit projection area calculation model based on UNet was established, and the experiment results showed that the R2 of the fruit projection model and the RMSE were 0.99 and 9.85 cm2, respectively. The above results showed that the linear relationship between the calculated and manually measured values was significant, and the comprehensive error was low, which can effectively calculate the key phenotypic parameters throughout the whole growth cycle of hanging watermelon.