Abstract: Rice diseases have a critical influence on yield, accurately identifying disease types and taking timely control measures are essential for minimizing economic losses．With rise of smart agriculture, precise identification and monitoring of plant diseases based on image technology has become critical．To achieve higher accuracy and reduce computational load in rice disease identification, an improved ResNeXt50 model, named ECA-ResNeXt, was proposed．Initially, ResNeXt network depth was reduced to 35 layers, a number of channels in initial layers was adjusted, convolutional channels were reduced, and standard convolutions were replaced with depthwise convolutions, effectively reducing a number of floating-point operations, parameter count, and storage requirements．Secondly, integration with efficient channel attention（ECA）module played a key role in improving model's feature representation capabilities．Experimental results showed that ECA-ResNeXt achieved an accuracy of 99.83% in rice disease identification, with a floating-point operations volume of only 0.054 GFLOPs, model parameters of 0.054×10⁶, and model size of 0.593 MB, demonstrating significant computational and storage efficiency．Compared to other classic convolutional neural networks, such as ResNet18, ResNet101, ResNeXt50, EfficientNet-b4, MobileNetV2, and MobileNetV3-Small, ECA-ResNeXt outperformed them in several evaluation metrics, including accuracy, precision, recall, and F1 score, particularly exceeding 99% in both precision and recall．In terms of transfer learning, ECA-ResNeXt’s performance in rice disease identification was further improved by pre-training on Plant Village dataset．Finally, an efficient rice pest and disease detection system was developed．Experimental validation confirmed that ECA-ResNeXt was highly efficient and resource-saving in rice disease identification, and showed great potential for practical applications．