Stacking-based and improved convolutional neural network: a new approach in rice leaf disease identification

Abstract

Rice leaf diseases are important causes of poor rice yields, and accurately identifying diseases and taking corresponding measures are important ways to improve yields. However, rice leaf diseases are diverse and varied; to address the low efficiency and high cost of manual identification, this study proposes a stacking-based integrated learning model for the efficient and accurate identification of rice leaf diseases. The stacking-based integrated learning model with four convolutional neural networks (namely, an improved AlexNet, an improved GoogLeNet, ResNet50 and MobileNetV3) as the base learners and a support vector machine (SVM) as the sublearner was constructed, and the recognition rate achieved on a rice dataset reached 99.69%. Different improvement methods have different effects on the learning and training processes for different classification tasks. To investigate the effects of different improvement methods on the accuracy of rice leaf disease diagnosis, experiments such as comparison experiments between single models and different stacking-based ensemble model combinations and comparison experiments with different datasets were executed. The model proposed in this study was shown to be more effective than single models and achieved good results on a plant dataset, providing a better method for plant disease identification.