Early detection of rice crop diseases remains crucial for global food security and sustainable agriculture. While deep learning approaches have shown promising results in disease detection, their deployment in resource-constrained agricultural environments presents significant challenges. This study presents optimizations on lightweight convolutional neural networks (CNNs) for real-time rice disease detection, with a particular focus on the recently introduced MobileNet-V4 architecture. A comparison of MobileNet-V4 is made against established architectures including MobileNet-V2, MobileNet-V3, and ResNet34 across multiple deployment scenarios. These deployment scenarios include high-performance GPUs, standard CPUs, and resource-constrained edge devices (Raspberry Pi 5). Simulated results demonstrate that MobileNet-V4 achieves superior performance, with an average inference time of 39 ms on a CPU and 88.91 ms on a Raspberry Pi 5, while maintaining a detection accuracy of 97.84 % across 12 common rice diseases. Further the model’s performance is enhanced through various training strategies, including focal and label smoothening loss to address class imbalance, and augmentation techniques such as random flipping, rotation and jittering to improve generalization. The described implementation achieves real-time performance on edge devices, making it particularly suitable for field deployment and establishing MobileNet-V4 as an efficient solution for practical precision-agriculture monitoring systems. The code for this study is available at: github.com/lars-uav/LARS-MobileNet-V4.