Deep learning for agriculture

Abstract

The global population is estimated to reach 8 billion by 2023 [1]. To feed such an immense population in a sustainable way, while also enabling farmers to make a living, requires the modernization of production methods in agriculture. In recent years there has been a lot of excitement in academic research and industry about the application of modern computer technology to farming, making farming one of the favorites for investors. According to Forbes Magazine, the agricultural technology gold rush began in 2013, with Monsanto's purchase of the agricultural data company, The Climate Corporation, for $930 million [16]. The total investment in 2017 topped $1.5 billion, setting a new record. All economic indicators point to a huge increase in technology and in particular software used in the agriculture fields. The need for advanced technology in agriculture is clear. The technology is being developed and ready, but what is still lacking is the number of professionals that have both skills-agriculture and technological knowledge-in particular in advanced computing methods like computer vision and deep learning. The main goal of this paper is to report on our approach to close the gap between domain experts in agriculture and computer scientists by developing a practical, hands-on activity in the form of a workshop or tutorial specifically targeted at agricultural engineers and practitioners interested in applying computer vision techniques to solve agricultural problems. The tutorial consists of specific examples like detecting and counting bees, segmentation of fruit trees and automatic fruit classification. The examples for the tutorials are chosen because of their simplicity of implementation and because they are also easily expandable into more complex projects. For example, the segmentation tutorial can be used to estimate pruning weight which is useful to determine the baseline vigor levels of the fruit trees. It is one of the best ways to quantify areas of the orchards that are not uniform in terms of expected yield, which will enable the use of precision agriculture. The benefits of precision agriculture are multifold leading to reductions in cost and increases in production by targeting specific areas. We explain and develop the tutorials for agricultural engineers assuming no previous knowledge in computer programming or computer vision. To test the validity of our approach, we conducted two workshops as part of a symposium on biotechnology for groups of about twenty-five participants (faculty and students) from different disciplines. One of the workshops was on Deep Learning for Plant Disease Detection, consisting of introductory lectures, and a hands-on tutorial. Our main emphasis for this workshop was the use of Jupyter Notebooks delivered through Google's Colab cloud solution as a vehicle to expose participants with no programming background to practical Deep Learning applications. At the end of the workshops we collected feedback from the participants through a survey; a summary is presented in the results sections. The other workshop used a similar approach, but with a topic from biology, categorizing frog embryos.