Damage assessment of soybean and redroot amaranth plants in greenhouse through biomass estimation and deep learning-based symptom classification

Abstract

Greenhouse plant assessment is key part in the process of developing and testing new herbicides as it serves to analyze the response of the species to those different products and doses in a controlled way. With that purpose, trials are carried out in greenhouse where the damage in the treated plants is daily assessed. This assessment of every pot is often performed in comparison with an untreated reference pot, also named as control pot. This assessment is currently done pot by pot through a time-consuming process which consists of visual assessments done by experts in the field. Digital tools to reduce time and to endow the experts with more objective and repetitive methods for establishing the damage in the plants are required.A novel solution based on image processing and deep learning techniques is proposed to estimate the damage in the plants in different growing stages in the greenhouse. Different damage types and in different stages are produced in plants and images of them are acquired to create a dataset. The available annotation is the damage estimation value provided by the experts. The proposed methodology tries to emulate the way the experts estimate the damage over the plants through a two-step procedure. First, the biomass reduction of the assessed plant compared to the corresponding control plant is calculated, and secondly, the possible disease symptoms in the plant are detected. The first part is done using classical image processing techniques and the second part relies on a deep learning based multi-label classification model for symptom classification.The algorithm has been tested over two species: Glycine max (soybean) and Amaranthus retroflexus (redroot amaranth). An R2 of 0.87 and 0.89 respectively is obtained for the damage estimation. The method improves the performance of the current manual process in terms of efficiency and objectivity.