A country’s economic progress relies heavily on the growth of the agricultural sector. However, plant diseases represent a major obstacle, affecting both the yield and quality of crops. In the absence of specialized expertise and when dealing with low-contrast images, the reliable identification of these plant diseases proves particularly challenging and requires considerable time. Recently, several intelligent solutions based on deep learning, particularly convolutional neural networks have been proposed for the early and accurate detection of plant diseases. However, the majority of studies focus solely on detection accuracy while neglecting the computational aspect, which can pose problems when deploying solutions in resource-limited environments. The objective of our study is to conduct a comparative analysis between five convolutional neural network architectures to determine which one combines high accuracy with reduced computational cost. Our approach trains the five architectures under the same conditions and on the same dataset, while collecting accuracy performance metrics and computational cost indicators to establish a comprehensive comparative study. Experiments conducted on the PlantVillage dataset comprising 38 categories show that MobileNet is the most balanced architecture, combining good detection performance with an accuracy of 96.87% and a minimal computational cost of 0.33 GFLOPs and two million trainable parameters. The VGG16 and AlexNet architectures achieve accuracies above 98%, but they are very resource-intensive. These experimental results make MobileNet the most suitable solution for plant disease detection in a resource-constrained context and the most appropriate for precision agriculture applications.