Monitoring of in-vitro plant cultures using digital image processing and random forests

Digital image techniques can play an important role in the study of plant phenotyping and monitoring, allowing the measurement of plant features without recurring to destructive or exhaustive methods. In in-vitro plant cultures, the nutritive and environmental culture conditions are determinant to the evolution of in-vitro plants. Therefore, to evaluate the performance of each essay, it i s essential to monitor these cultures. In this paper, we present a computer vision and machine learning system for in-vitro plant monitoring. For our study, we tracked the evolution of Nandina Domestica Sunset Boulevard in-vitro containers by weekly acquiring visible and near-infrared images. Our approach pre-processes, registers, and normalize s the acquired images from one container, classifies plant and non-plant pixels using a trained Random Forest classifier and exports plants total area and average NDVI (Normalized Difference Vegetation Index) for each monitoring session. Our plant segmentation achieved 96.9% accuracy, 96,8% sensitivity and 96.9% specificity using our ground truth data.