Lightweight Hybrid AI Models for Rice Pest Classification Using Deep Feature Extraction

With the global population expected to reach 9 billion by 2050, improving early pest detection is essential to protect rice yields and support long-term food security. Traditional methods of identifying pests in the field are slow and often inaccurate, creating a need for faster and more reliable tools. This research focuses on developing lightweight hybrid AI models that classify rice pests using deep feature extraction, making the system efficient enough to run on resource-limited devices used by farmers. Using the IPI02 dataset, this study evaluated several computer-vision models including MobileNetV2, DenseNet201, and ConvNeXtXLarge. Instead of relying on full end-to-end deep learning, the study used a hybrid approach where deep features were extracted from these networks and passed into classical machine learning classifiers. The best results came from pairing ConvNeXtXLarge feature ex-traction with a lightweight Multilayer Perceptron (MLP), achieving a validation accuracy of 66.90 %. The findings show that deep feature extraction can improve accuracy while keeping the model compact enough to be suitable for potential real-world deployment. This system can help farmers detect pests earlier, choose targeted treatments, and reduce unnecessary pesticide use, supporting more sustainable rice production and contributing to global food -security efforts.