Early detection and prediction of powdery mildew of salad cucumber under polyhouse cultivation

Abstract

The study entitled “Early detection and prediction of powdery mildew of salad cucumber under polyhouse cultivation” was conducted at the Department of Plant Pathology, College of Agriculture, Vellanikkara during the period from 2022-2023. The main objective of the study was to develop a mobile application for early detection of powdery mildew of salad cucumber and a prediction model for the disease based on meteorological parameters during the particular crop period.Symptoms are visible manifestations of disease endured by a plant and imaging techniques are used to detect a disease by identifying the symptoms at an early stage. Early detection of diseases is essential for minimizing economic loss and also managing the disease using biocontrol measures. Moreover, disease spread is faster in protected structures owing to reasons such as intensive monoculture and controlled conditions prevailing. Considering these, salad cucumber was chosen for this project as it is a popular vegetable crop in polyhouses in our state and powdery mildew is a disease which spreads very fast under favorable conditions.Hence, in the present project, visible imaging technique was used in combination with electronics and artificial intelligence tools to detect the presence of the disease at a very early stage of symptom development. The crop was sown in the polyhouse in February 2023 and the first incidence of disease was observed 47days after sowing. The images of healthy and infected leaves were taken during the crop period in a customized manner which were then processed and classified for enhancing the accuracy. As the next step, a disease detection model was developed using two different machine learning algorithms viz. DenseNet and Xception. The image pre-processing was also done through two techniques viz.ellipse and colour channel filtering. These two techniques were used to highlight the infected area in a particular leaf. The images were then divided in 70:20:10 ratio for train, validation and test datasets respectively. DenseNet model was used in the first part of mobile application where it assesses whether the leaf is healthy or not. So both healthy and infected leaf images were used to train the model to identify an infected leaf from healthy leaf. The DenseNet model achieved an accuracy of 90 percent in the detection process. Xception model was used in the second part of the application to assess the severity level of different infected leaves. A total of 2700 images taken during the study were grouped into seven classes from healthy to maximum severity level. The machine learning model was trained to detect the severity level and the model showed different level of accuracy for different severity levels of powdery mildew. The model could detect the presence of powdery mildew at the first infection stage itself in which only 10 percent of leaf area is infected with a test accuracy value of 82.05percent. Later these two models were jointly used for the development of a mobile application. The final application was developed in such a way that it detects the presence of powdery mildew at the earliest visible stage, assess the severity level and suggest management accordingly. Validation of the mobile application was also done using 120 images belonging to different severity levels and achieved an accuracy varying from 80-100 percent. Pearson’s chisquare test was done to check whether the accuracy of disease detection is significant rather than a random chance detection and found to be significant. In the second part of the study, the relationship between meteorological parameters that prevailed inside the polyhouse and severity of powdery mildew was assessed and a prediction model was developed for the disease. Accordingly, atmospheric temperature, relative humidity (RH) and light intensity were recorded daily at two time periods: at 7.30am and 2.30pm starting from 40 days after sowing.The temperature inside the polyhouse during the crop period varied from 25-28°C in the morning hours and 35-37°C in the afternoon hours. Throughout the period of powdery mildew infection, the average temperature inside thepolyhouse was nearly 30±1°C. The average RH inside the polyhouse varied from 63 to 74.5 percent.The light intensity varied widely, which ranged between 4800-6600lux. Percent disease severity (PDS) was calculated for 100 randomly scored leaves on daily basis. Correlation and regression analyses of meteorological data with PDS were performed where temperature showed a significant negative correlation and RH,a significant positive correlation with PDS. While, there was no significant correlation between light intensity and PDS. So based on the data on temperature, RH and PDSrecorded during the experiment, a prediction model of disease was developed using MATLAB software. An equation for the prediction of PDS also developed using multiple polynomial regression model. On validation the prediction model recorded an accuracy of 81 percent. Even though the android mobile application for early detection of powdery mildewis designed for salad cucumber, as the machine learning model uses colour filtering and ellipsing for image recognitionthis application can detect powdery mildew on any crop at first instance of its appearance. Farmers can easily detect the presence of the disease at its earliest stage of symptom expression. The algorithm used for this study can be further modified and utilized for other diseases also