基于PLSR和BPNN方法的番茄光合速率预测比较

CO₂作为温室作物光合作用的重要原料,不同环境因子交互作用的植株叶片对CO₂浓度需求具有较大差异。为寻求CO₂浓度合理增施量,该文基于偏最小二乘法和BP神经网络方法对不同生长阶段番茄作物进行光合速率预测,进而探讨作物生长过程中可通用的光合速率预测方法。试验以无线传感器网络系统实时监测环境信息(CO₂浓度,光照强度,空气温度及相对湿度),以LI-6400XT光合速率仪获取作物单叶净光合速率。剔除样本奇异点后,对样本值进行统一归一化。以CO₂浓度、光照强度、空气温度及相对湿度为模型输入变量,以光合速率为输出量,利用偏最小二乘法和BP神经网络方法分别建立番茄幼苗期,开花期及结果期的光合速率预测模型。模型验证结果表明,偏最小二乘法在番茄各生长阶段的决定系数分别为0.74,0.88和0.85,最大相对误差为15.01%；而BP神经网络在各阶段具有较高的预测精度,其决定系数分别为0.94,0.96和0.97,最大相对误差为9.56%。因此,基于BP神经网络模型预测了特定环境下的CO₂浓度饱和点,为温室CO₂增施提供依据。

Comparison of photosynthesis prediction methods with BPNN and PLSR in different growth stages of tomato

To add the appropriate amount of CO₂ based on plant requirements, single-leaf photosynthesis was studied under interaction influence of other environmental factors.Two types of methods were employed to predict the photosynthetic rate in different growth stage of tomato, namely, a back-propagation neural network(BPNN) and partial least squares regression(PLSR).A series of experiments on predicting the photosynthetic rate of tomato plants was performed in a greenhouse based on environmental information.A wireless sensor network system was developed to monitor environmental information automatically, including CO₂ concentration, photosynthetically active radiation, air temperature, and relative humidity.An LI-6400XT photosynthetic rate instrument was used to obtain the net photosynthetic rate of a functional leaf.Before building the prediction models, input data should be preprocessed to improve accuracy.The preprocessing procedure comprised two steps.First, singular points were deleted by analyzing the variable box plot.The data were then normalized.After comprehensively considering CO₂ concentration in the greenhouse, the prediction models of tomato single-leaf photosynthesis were established via PLSR and BPNN during different growing stages.The prediction results were compared by evaluation indices.The results showed that the correlation coefficients between the simulated and observed data sets were 0.94, 0.96, and 0.97 in the three growing stages using BPNN, and 0.74, 0.88, and 0.85 using PLSR.The results proved that the BPNN model exhibited higher prediction accuracy than the PLSR model and could be used to control CO₂ air fertilizer precisely in a greenhouse.