Machine vision monitoring of plant nutrition

Amots Hetzroni, Purdue University

Abstract

Intensive agriculture in complex controlled environments such as greenhouses or life-support systems requires supervision and control by an intelligent monitor that is capable of identifying and diagnosing stress symptoms in plants as early as possible. This monitoring system can be used in long term space missions as well as in tissue culture plant propagation, greenhouses and fields as part of a health assessment system. Although contemporary technology can provide ambient conditions and essential plant needs with high precision, stress symptoms manifested by the plants should be observed and analyzed by an expert. Color machine vision can capture some of these symptoms, and together with a diagnosis system can evaluate the plants' status. This thesis examines the use of machine vision for monitoring plant health by evaluating visual features measured from plant images. A model of plant health evaluation from digital images is described and tested with lettuce plants subjected to nutrient deficiencies (iron, zinc, and nitrogen). A color calibration procedure was developed and evaluated using simulation. It is capable of rendering images captured under different illumination conditions, to look similar to an eye or a sensor. Different approaches (supervised and automatic) to image segmentation for extracting plant image from the picture are discussed and demonstrated. Accuracy of supervised segmentation was found to be superior to automatic segmentation. Neural networks and statistical approaches to health discrimination of plants were examined. A method of augmenting independent information to the process of statistical classification is analyzed and evaluated. A multivariate classification system successfully classified 83% of the plants to the correct treatment groups. The time before the system was able to correctly classify the plants varies among experiments and treatments. Most of the plants were correctly classified between day 11 to 20 since germination. This work provides the basic tools that are necessary for the implementation of monitoring plant health based on color imaging. This includes system calibration and an approach for data analysis and decision support. The method of color calibration that was developed and evaluated is applicable to any application that involves color imaging using RGB video. The classification approach, the method of augmenting and using statistical auxiliary data, and the programs developed are not limited to evaluation of nutrient deficiency of lettuce plants from three channel TIFF images, but are applicable as generic multivariate classification methods and programs.

Degree

Ph.D.

Advisors

Miles, Purdue University.

Subject Area

Agricultural engineering|Electrical engineering

COinS