基于LSTM的温室番茄蒸腾量预测模型研究

Prediction Model of Transpiration of Greenhouse Tomato Based on LSTM

摘要:

作物蒸腾量是指导作物灌溉关键参数之一，实时获取作物蒸腾量，实现按需灌溉是节约用水的有效途径。然而，温室内小气候效应显著，作物蒸腾与环境因子间关系较为复杂，且各环境因子之间相互关联并呈非线性变化。本文以番茄作为研究对象，使用称量法测量作物实时蒸腾量，通过布设传感器实时获取温室小气候数据，包括空气温度(Air temperature, AT)、相对湿度(Relative humidity, RH)、光照强度(Light intensity, LI)作为模型的小气候环境输入，冠层相对叶面积指数(Relative leaf area index，RLAI)作为模型的作物生长输入，在此基础上，提出了基于长短期记忆网络(Long short term memory, LSTM)的番茄蒸腾量预测模型。利用该模型对番茄蒸腾量进行预测，并与非线性自回归(Nonlinear autoregressive with exogeneous inputs, NARX)神经网络、Elman神经网络、循环神经网络(Recurrent neural network, RNN)模型进行了对比。试验结果表明，LSTM预测模型决定系数(Determination coefficient, R2)与平均绝对误差(Mean absolute error, MAE)分别为0.9925和4.53g，与NARX神经网络、Elman神经网络、RNN方法进行对比，其决定系数分别提高了8.97%、1.18%和0.82%，其平均绝对误差分别降低了8.16、6.23、0.52g。本研究所提的预测模型具有较高的预测精度及泛化性能，研究成果可为温室作物需水规律及需水量研究提供参考。

Abstract:

Crop transpiration is one of the key parameters to guide crop irrigation. It is an effective way to save water to obtain crop transpiration in real time and realize irrigation on demand. However, the microclimate effect in greenhouse is significant, the relationship between crop transpiration and environmental factors is complex, and each environmental factor is interrelated and presents nonlinear change. Taking tomato as the research object, the weighing method was used to measure the real-time transpiration of crop. Greenhouse microclimate data could be obtained in real time through the installation of sensors, including air temperature (AT), relative humidity (RH) and light intensity (LI) as the microclimate environment input of the model, and canopy relative leaf area index (RLAI) as the crop growth input of the model. On this basis, a prediction model of tomato transpiration by long short term memory (LSTM) network was proposed. The model was used to predict the transpiration of tomato, and compared with the nonlinear autoregressive with exgeneous inputs (NARX) neural network, Elman neural network and recurrent neural network (RNN). The results showed that the determination coefficient (R2) and mean absolute error (MAE) of the LSTM prediction model were 0.9925 and 4.53g,respectively. Compared with NARX, Elman neural network and RNN, their R2 were increased by 8.97%, 1.18% and 0.82%, respectively, and their MAE were decreased by 8.16g, 6.23g and 0.52g, respectively. The prediction model proposed had high prediction accuracy and generalization performance, and the research results could provide reference for the study on the regularity and water demand of greenhouse crops.