Hyperspectral-physiological based predictive model for transpiration in greenhouses under CO2 enrichment

Abstract

Accurate prediction of transpiration in protected agricultural settings with CO2 enrichment is of high importance for the deployment of precision irrigation. Several physical and mechanistic based models are deployed for the estimation of transpiration. However, these models may not accurately explain physiological interactions for greenhouse settings under CO2 enrichment. Thus, it is necessary to build site-specific and microclimate-targeted models that can accurately estimate transpiration and irrigation water requirements. In this work, a data-driven predictive model is proposed for the estimation of short-term transpiration (mmol/m2.s) for cucumber crops grown in greenhouses with CO2 enrichment. Three machine learning models were investigated for transpiration modelling and prediction: deep neural networks (DNN), extreme gradient boosting (XGBoost), and support vector machine regression (SVR). These predictive models assimilate microclimate, physiological and hyperspectral features with high temporal and spatial resolutions. The results demonstrated the inclusion of hyperspectral-based vegetation indices significantly increased the performance of the three machine learning models in predicting transpiration. The XGBoost model outperformed the DNN and SVR models with higher R2 of 0.62 – 7.53%, lower RMSE and MAE values of 9.74 – 46.19 mmol/m2.s and 13.74 – 39.08 mmol/m2.s respectively with microclimate, physiological and hyperspectral based vegetation index inputs. Moreover, the time series prediction with different time steps improved the XGBoost model, achieving an R2 of up to 99.0%. The XGBoost proved to be the most efficient in predicting transpiration for cucumbers grown in greenhouses under CO2 enriched environments.