Deep Learning Techniques to Detect Nutrient Deficiency in Rice Plants

Abstract

Rice is one of the most widely grown crop in China, Bangladesh, and India. Crop productivity is eventually impacted by nutritional imbalances in plants, which affect their responses and defense mechanisms against abiotic stress, pests, and diseases. A majority of farmers use soil analysis, plant tissue analysis, and visual symptoms to identify nutrient deficiencies. However nutrient deficiency detection using soil analysis and plant tissue analysis requires experts. To reduce the workload of experts and farmers, modern technologies like Artificial Intelligence (AI), Machine Learning, Internet of Things (IoT) systems, and Deep Learning techniques are used. The main objective is to improve the accuracy in identifying and classifying nitrogen, phosphorous and potassium deficiency in rice plants compared to earlier approaches. In the proposed study, five of the most popular pre-trained CNN models namely InceptionV3, VGG16, VGG19, ResNet50, and ResNet152 is used along with SVM to predict the nutrition deficiencies in the Kaggle rice NPK deficiencies dataset. Additionally, the number of images is increased from 1156 to 4963 by performing augmentation on the Kaggle rice NPK deficiency dataset. The proposed system applies pre-trained CNN models (InceptionV3, VGG16, VGG19, ResNet50, and ResNet152) on datasets (with and without augmentation) to extract deep features from the images. These features are then saved in the feature vector. The feature vectors are then fed to the SVM in order to identify any nutrient NPK deficits in the images. Each classifier’s performance is evaluated and contrasted based on its accuracy. The accuracy of all pre-trained CNN models ranges from 91% to 99%. The best result on the dataset without augmentation is 97.40%, and on the dataset with augmentation is 99.05%, with the pre-trained CNN model Resnet50+SVM.

Topics & Concepts