Transfer Learning in Ensemble Convolutional Neural Networks Toward the Detection of Microscopic Fusarium Oxysporum

Abstract

The Panama disease, also known as the Fusarium wilt, is a deadly disease known to affect banana plants all over the world. Caused by a fungal pathogen known as Fusarium oxysporum f. sp. cubense (Foc), the disease has been a constant threat to banana producers considering that it cannot be eradicated once it has infected the soil. A new strain that emerged in 1989 called Tropical Race 4 (TR4) is now threatening the Cavendish cultivar, the most popular banana variety being grown today. Furthermore, symptoms of the disease are not visible until late stages of the infection. While there are methods that accurately determine the presence of Foc in a soil sample, these are costly and inaccessible to most banana producers. Thus, we propose the use of convolutional neural networks in the automatic detection of Foc TR4 in soil samples with the use of microscopy. This study utilized a dataset containing microscopy images of various fungal species captured using three distinct microscopy techniques: brightfield, darkfield, and fluorescent. Transfer learning has shown to make significant improvements to the performance of the models in classifying microscopic fungi. The best performing individual model was trained exclusively on brightfield images and has achieved an accuracy score of 93.92% while the best ensemble model was able to achieve an accuracy of 97.55%. Furthermore, tests on a subset meant to simulate realistic appearances of Foc have shown that the final model is viable for actual field use.