Deep Neural Networks for Classifying Nutrient Deficiencies in Rice Plants Using Leaf Images

Abstract

Nutrients are vital in ensuring expected crop growth and yield quality. Accurate identification of nutrient deficiencies in plants is essential to provide appropriate supplements of fertilizers. Manual inspection of symptoms and identifying nutrient deficiencies is a tiresome task requiring higher expertise. This paper aims to design and develop a computationally efficient deep-learning model to classify plant nutrient deficiencies accurately. This paper presents an image-based deep-learning framework for nutrient deficiency identification. Three deep learning models, namely the Xception model, vision transformer, and multi-layer perceptron-based (MLP) mixer model, were trained to identify nitrogen (N), phosphorous (P), and potassium (K) deficiencies in rice plants from red-green-blue (RGB) images. The model performance is tested on nutrient deficiency symptoms in rice plants dataset available publicly on Kaggle. All three models achieved nutrient deficiency classification accuracy greater than 92%. The Xception model achieved the highest average accuracy of 95.14% at the cost of approximately 1.2 million total trainable parameters, much less than the vision transformer and MLP mixer model. The Xception model performs better than the other two models in classifying nutrient deficiencies with the least number of total trainable parameters. In the future, these neural networks can be trained and extended to accurately detect and segment nutrient-deficient crop areas in large fields to supply precise fertilizer supplements.