A Comparative Analysis of Convolutional Neural Network Architectures for Coffee Leaf Rust Detection

Abstract

Coffee plays a cultural part in the lives of people. Eight out of ten adults in the Philippines drink an average of 2.5 cups of coffee per day, and nine out of ten households have coffee in their pantries. In 2021, it is reported that coffee will be an approximately PHP 3.0 billion (USD 5 million) industry in the country. As such, much effort is placed into sustaining its economic production. Botanical studies reveal that some factors severely affect the quality of coffee plants. One of them is coffee leaf rust which is caused by the fungus known asHemileia vastatrix and affects many coffee-growing regions.If not detected early enough it can greatly reduce coffee yield or worse wipe out an entire plantation affecting the farmers. This study proposes a methodology for detecting coffee leaf rust disease using computer vision and employs multiple pre-trained architectures of Convolutional Neural Networks (CNNs). Specifically, this study used twenty-five convolutional neural networks, and the best-performing models are ResNet101V2, InceptionV3, ResNet50V2, Xception, and DenseNet169. They have a training accuracy of 92.62%, 96.90%, 93.45%, 98.45%, and 96.79% respectively. Meanwhile, their validation accuracy is 91.67%, 90%, 94.44%, 93.89%, and 95.56% respectively. Out of the top five CNNs, ResNet101V2 achieved the highest test accuracy with 95.56% and it also excels in other evaluation metrics such as Precision, Recall, and F1 – score. Although this study used a variety of CNNs, it is also recommended to use more types of algorithms as well as increase the number of epochs. Future studies can also consider the detection of other coffee diseases and a larger dataset for a wider scope. Overall, the result of this study is a step closer to achieving improved coffee production and the livelihoods of farmers.