DEVELOPMENT OF A MACHINE VISION SYSTEM FOR STRAWBERRY POWDERY MILDEW DISEASE DETECTION

Abstract

Strawberry powdery mildew (Sphaerotheca macularis) has been a devastating foliar disease of both nursery and fruit production strawberry crops causing significant yield loss up to 70%. Traditionally, visual observations are made to monitor strawberry powdery mildew disease each week by human experts which is a laborious and time consuming endeavour. The main objective of this study was to detect powdery mildew disease in real-time field condition by developing an image processing-based machine vision system. The machine vision system consisted of a graphical user interface based powdery mildew detection program, two µeye cameras, a real-time kinematic global positioning system, and a ruggedized laptop computer. A colour co-occurrence matrices based image texture analysis program was generated by using C# programming language. Factors affecting machine vision parameters were optimised by evaluating performance upon 12,000 collected images from healthy and diseases affected strawberry leaves. Results suggested artificial cloud lighting system improved real-time powdery mildew detection accuracy as compared to natural lighting conditions. Results also suggested that feature model formed by green ratio, hue, saturation and intensity images including 23 features, image acquisition speed 1.5 km h-1 and camera working depth of 300 mm were outperformed for disease detection. Results of the classifier selection revealed that artificial neural network performed better than support vector machines and k-nearest neighbor based supervised learning classifiers to detect powdery mildew disease in strawberry leaves. Real-time performance of developed vision system was tested in 36 randomly selected rows in three commercial strawberry fields. Visual observations were compared with developed automatic detection system. Real-time test evaluation results demonstrated that the machine vision system was capable to detect powdery mildew disease with mean absolute error of 4.00, 3.42 and 2.83 per row and root mean square error of 4.12, 3.71 and 3.00 per row. The overall study reported the developed strawberry powdery mildew detection system can be used to help strawberry growers by reducing expenses associated with field scouts for manual monitoring.