Early detection of plant diseases is critical to minimizing their adverse effects on agricultural productivity. In particular, machine vision and deep learning approaches (e.g., convolutional neural networks, CNNs) have been increasingly applied for automatic plant disease identification. Although existing deep learning models achieve satisfying classification accuracy, they often consist of millions of parameters that significantly lead to the lengthy training time, prohibitive calculation costs and deployment obstacles at the resource-constrained edge devices. In order to overcome those constraints, we introduce a new deep learning architecture, which uses adaptations of Inception layers and residual connections that can help both with feature extraction and efficiency. In addition, depthwise separable convolutions are used to drastically reduce the amount of trainable parameters with small loss of representational power. We perform training and evaluation of the proposed model on three located benchmark plant disease datasets, PlantVillage dataset, the Rice Disease dataset. Experimental results show that our model achieves state-of-the-art classification accuracy of 99.39% on the PlantVillage dataset, 98.66% on the Rice Disease dataset. In contrast to the state-of-the-art deep learning models, our method obtains higher accuracy with fewer parameters so that it could be better suited for real-time applications on mobile and embedded systems. We explore an application of deep learning with the use of optimized architectures and present the viability of this technique in precision agriculture for faster and more accurate diagnosis of diseases in plants with lower computational load.