Convolutional neural network and thermal images for hydric stress prediction

Abstract

Sustainability of agricultural production is directly related to the management of the availability of water to the crop during its development period, directly interfering with the final production. In order to provide improvement in this management, in recent years, a technology that has been highlighted is the use of thermal images, allowing the visual analysis of the temperature of a plant. Despite the information provided by this type of image in the diagnosis, the analysis is not intuitive and requires knowledge about the physical-chemical processes of the plant. In the advent of the new information provided by the thermal images and the complexity of its interpretation, we searched for neural network models capable of assisting in the determination of water deficit in a plant. By the reason that it is a problem of image classification, it was determined that the convolutional neural models would best suit the task. Thus, the reference convolutional models of image classification were selected, and the transfer learning technique will be implemented in these models in order to carry out their training for the classification of the level of water deficit in a plant. The aim of this project is to be able to produce a trained model that can perform this task in a non-destructive way and that allows the replacement of existing classical methods without the need for large infrastructures. (AU)