For experienced farmers and biologists, detecting and identifying plant diseases is usually fairly straightforward: they are familiar with the most common problems related to their crops and can use visual inspection, possibly in combination with other contextual aspects, to identify the cause from the perceived symptoms. This manual method of disease detection, however, does not scale very well and depends on the availability of such experts.
In this paper, the authors review the literature on plant disease detection and compare different approaches. They group the approaches into two larger categories: approaches based on convolutional neural networks (CNNs) and conventional approaches. The latter consists mostly of a combination of clustering methods like k-means for segmentation and support vector machines (SVMs) for classification. The former mostly use adaptations of widely used CNN architectures for image processing, such as AlexNet, GoogLeNet, or Inception, typically in combination with transfer learning. One of the main comparison criteria used is the accuracy of the model. All CNN approaches mentioned reach accuracies higher than 93 percent, several of them more than 99 percent. Combined with transfer learning, CNNs can achieve high accuracy with fast training times. For most conventional models, no accuracy figures are given; only two reach 90 percent or higher.
Overall, I found the paper moderately informative. It is short at five pages, and the selection of articles clearly covers only a relatively small subset of the publications on this topic. Especially the use of deep learning techniques has resulted in hundreds of related publications over the past few years, so it is difficult to achieve wider coverage. As a starting point for a deeper descent into the rabbit hole, however, it offers a good overview.