Improving the Performance of Vegetable Leaf Wetness Duration Models in Greenhouses Using Decision Tree Learning

Hui Wang,Ming Li,Jorge Sanchez-Molina,Francisco Rodríguez Díaz

doi:10.3390/w11010158

Abstract

Leaf wetness duration (LWD) is a key driving variable for peat and disease control in greenhouse management, and depends upon irrigation, rainfall, and dewfall. However, LWD measurement is often replaced by its estimation from other meteorological variables, with associated uncertainty due to the modelling approach used and its calibration. This study uses the decision learning tree method (DLT) for calibrating four LWD models—RH threshold model (RHM), the dew parameterization model (DPM), the classification and regression tree model (CART) and the neural network model (NNM)—whose performances in reproducing measured data are assessed using a large dataset. The relative importance of input variables in contributing to LWD estimation is also computed for the models tested. The LWD models were evaluated at two different greenhouse locations: in a Chinese (CN) greenhouse over three planting seasons (April 2014–October 2015) and in a Spanish (ES) greenhouse over four planting seasons (April 2016–February 2018). Based on multi-evaluation indicators, the models were given a ranking for their assessment capabilities during calibration (in the Spanish greenhouse from April 2016 to December 2016 and in the Chinese greenhouse from April 2014 to November 2014). The models were then evaluated on an independent set of data, and the obtained areas under the receiver operating characteristic curve (AUC) of the LWD models were over 0.73. Therein, the best LWD model in this case was the NNM, with positive predict values (PPVs) of 0.82 (SP) and 0.90 (CN), and mean absolute errors (MAEs) of 1.85 h (SP) and 1.30 h (CN). Consequently, the DLT can decrease LWD estimation error by calibrating the model threshold and choosing black box model input variables.

Highlights

Leaf wetness duration (LWD) is the time of visible water presence on plant surfaces, i.e., the leaves, stems, flowers and fruits [1]
The classification tree function in MATLAB R2018a was used to calibrate the models, and their performances were assessed by the statistical indexes: positive predictive value parameters of the leaf wetness models, and their performances were assessed by the statistical (PPV), negative predictive value (NPV), area under the receiver operating characteristic curve (AUC)
The calibration of leaf wetness model parameters is a main source of variation in LWD estimation in different locations, especially for empirical models which are based on local climatic data and input conditions

Summary

Introduction

Leaf wetness duration (LWD) is the time of visible water presence on plant surfaces, i.e., the leaves, stems, flowers and fruits [1] It is caused by over irrigation, rainfall or condensation, and acts as a catalyst for disease onset as it favors fungal infections, causing a high possibility of heavy yield losses. Further attempts have been made for estimation of LWD using neural network models (NNMs), which can be considered black box models These are self-adaption models trained with reference data, and are widely used to various prediction problems, including temperature [17], leaf area index [18], gray leaf spot on maize [19] and flooding [20]. Before using leaf wetness models operationally, they should be compared to analyze their performance, and to help farms in greenhouse management, e.g., to avoid over irrigation, untimely ventilation and disease infection

Objectives

Methods

Results

Discussion

Conclusion