Evaluating Impacts between Laboratory and Field-Collected Datasets for Plant Disease Classification

Deep learning with convolutional neural networks represents the most used approach in recent years in the classification of leaves' diseases. The literature has extensively addressed the problem using laboratory-acquired datasets with a homogeneous background. In this article, we explore the variability factors that influence the classification of plant diseases by analyzing the same plant and disease under different conditions, i.e., in the field and in the laboratory. Two plant species and five biotic stresses are analyzed using different architectures, such as EfficientB0, MobileNetV2, InceptionV2, ResNet50 and VGG16. Experiments show that model performance drops drastically when using representative datasets, and the features learned from the network to determine the class do not always belong to the leaf lesion. In the worst case, the accuracy drops from 92.67% to 54.41%. Our results indicate that while deep learning is an effective technique, there are some technical issues to consider when applying it to more representative datasets collected in the field.