A Machine Learning Model for the Diagnosis of Coffee Diseases

The growing and marketing of coffee is an important source of economic resources for many countries, especially those with economies dependent on agricultural production, as is the case of Colombia. Although the country has done a lot of research to develop the sector, the truth is that most of its cultivation is carried out by small coffee families without a high degree of technology, and without major resources to access it. The quality of the coffee bean is highly sensitive to diverse diseases related to environmental conditions, fungi, bacteria, and insects, which directly and strongly affect the economic income of the entire production chain. In many cases the diseases are transmitted rapidly, causing great economic losses. A quick and reliable diagnosis would have an immediate effect on reducing losses. In this sense, this research advances the development of an embedded system based on machine learning capable of performing on-site diagnoses by untrained personnel but taking advantage of the know-how of expert coffee growers. Such a system seeks to instrument the visual characteristics of the most common plant diseases on low-cost, robust, and highly reliable hardware. We identified a deep network architecture with high performance in disease categorization and adjusted the hyperparameters of the model to maximize its characterization capacity without incurring overfitting problems. The prototype was evaluated in the laboratory on real plants for recognized disease cases, tests that matched the performance of the model validation dataset.