日光温室基质培生菜鲜质量无损估算方法

设施栽培中作物鲜质量动态变化作为生长发育的指示性特征,是蔬菜长势无损监测的重要指标之一。水培蔬菜通过离水直接称量实现长势无损监测,但是土培或基质培无法通过直接鲜质量称量实现生长过程的无损测量。本文提出了基于表型特征参数和生长过程环境参数融合的鲜质量估算方法,用于日光温室环境下基质培生菜个体和群体的鲜质量无损估算。首先,监测生菜全生命周期的环境参数,采集第1批次生菜生长过程中的多样本图像和部分样本鲜质量,提取样本图像中不同生长期生菜的形状、颜色、纹理等特征,计算环境信息中的累积辐热积等参数。然后,利用高斯过程回归方法建立表型参数和环境参数与生菜鲜质量的关系模型。最后,采集第2批次生菜群体的样本数据,基于上述模型预测生菜3个生长阶段的个体和群体鲜质量,以验证鲜质量估算模型的泛化能力和可靠性。结果表明,与支持向量机、线性回归、岭回归和神经网络相比,高斯过程模型的决定系数R2为0.9493,相对误差的均值和标准差分别为11.50%和11.21%。模型泛化能力试验中,生菜群体鲜质量比个体鲜质量的预测相对误差的平均值小（3个生长阶段分别相差4.44、5.71、5.89个百分点）,且随着群体数量增加,群体鲜质量预测的相对误差的均值和标准差逐渐减小。本鲜质量估算方法预测的群体鲜质量数据可为基质培绿叶菜类作物的栽培管理决策提供数据支撑。

Non-destructive Estimation Method of Fresh Weight of Substrate Cultured Lettuce in Solar Greenhouse

As an indicative feature of crop growth and development, the dynamic change of fresh weight of crops in facility cultivation is one of the important indicators for non-destructive monitoring of vegetable growth. Hydroponic vegetables can be directly weighed out of water to achieve non-destructive monitoring of growth, but it is difficult to achieve fresh weight by non-destructive measurement in soil or matrix. To solve this problem, the fresh weight estimation method based on the combination of phenotypic parameters and environmental parameters was proposed to estimate the fresh weight of individuals and groups of lettuce in solar greenhouse. Firstly, the environmental parameters of the whole life cycle of lettuce were monitored. The multi sample images and the fresh weight of some samples were collected during the growth of the first batch of lettuce. The shape, color, texture and other characteristics of lettuce in different growth periods were extracted from the sample images, and the accumulated heat product in the environmental information was calculated. Then, the relationship model between the parameters of phenotype and environment and fresh weight of lettuce was established by Gaussian process regression. Finally, the sample data of the second batch of groups of lettuce were collected to predict the fresh weight of individuals and groups of lettuce at three growth stages based on the above model, so as to verify the generalization ability and reliability of the fresh weight estimation model. The results showed that compared with support vector machine, linear regression, ridge regression and neural network, the determination coefficient of Gaussian process model was 0.9493, and the mean of relative error was 11.50%, while the standard deviation of relative error was 11.21%. In the model generalization ability test, the average value of relative error of prediction of fresh weight of groups of lettuce was smaller than that of individuals of lettuce,and the difference of them were 4.44, 5.71 and 5.89 percentage points at the three growth states. The average value and standard deviation of predicted fresh weight of groups of lettuce was gradually decreased with the increase of groups. The fresh weight data of groups predicted by this method can provide data support for the cultivation and management decision of substrate cultivated green leafy vegetables.