Application of smartphone-image processing and transfer learning for rice disease and nutrient deficiency detection

Abstract

The earliest detection of plant disease is the primary concern of the farming community. The availability of advanced digital cameras and smartphones with improved image acquisition modes and deep learning methods like convolutional neural networks (CNN) can detect plant disease with high accuracy. This study used 2259 smartphone images of various rice (Oryza sativa) plant parts under various classes and 250 real-time validation images for classifying 12 rice diseases and nutrient deficiency symptoms. Different image segmentation techniques like foreground extraction were used to segment affected portions. Additionally, optimization of models and procedures to use them in smartphones with offline working capabilities have been described. Furthermore, a dynamic framework has been developed and demonstrated where the server trains on unseen image data to improve the classification performance and updates the model on falling below a certain threshold level. A comparison was made across different models used for image classification with many supporting metrics to select the best model for transfer learning. The top four performing models were DenseNet201, Xception, MobileNetV2, and ResNet50, with validation accuracies of 0.9803, 0.9778, 0.9756, and 0.9718, respectively. The ResNet50 model was found to best among all for cloud architectures, while MobileNetV2 appeared as the best model for the smartphone application. Finally, the android application, “Rice Disease Detector” compiled with the MobileNetV2 model was tested for multiple disease occurrences in a single capture. More research is warranted to test the application for smartphones with variable configurations.

Keywords

• Image processing
• Smartphone
• Application
• Rice disease detection
• Rice nutrient deficiency detection
• Transfer learning