Effective Groundnut Crop Management by Early Prediction of Leaf Diseases through Convolutional Neural Networks

Abstract

Groundnut (Arachis hypogaea L.), is the sixth-most significant leguminous oilseed crop grown all over worldwide. Groundnut, due to its high content of various dietary fibers, is classified as a valuable cash, staple and a feed crop for millions of households around the world. However, due to varied environmental factors, the crop is quite prone to many kinds of diseases, identifiable through its leaves, for which Groundnut producers have to suffer major losses every year. An early detection of such diseases is essential in order to save this significant crop and avoid huge losses. This paper presents a novel Machine Learning based Deep Convolution Neural Network (CNN) model ‘CNN8GN’. The model uses transfer learning technique for detection of such diseases in Groundnuts at an early stage of crop production. A Groundnut real image data set containing a total of 5322 real images for six different classes of Groundnut leaf diseases, captured in the fields of Gujarat state (India) during September 2022 to February 2023, is generated for training, testing and evaluation of the proposed model. The proposed deep learning model architecture is designed on eight different layers and can be used on varied sized images using simple ReLu and Softmax activation functions. The performance of the proposed CNN8GN model on Groundnut real image dataset is examined using a detailed experimental analysis with other six pre-trained models: VGG16, InceptionV3, Resnet50, ResNet152V2, VGG19, and MobileNetV2. CNN8GN results are also examined in detail using different sets of input parameters values. The proposed model has shown significant improvements for disease detection in comparative analysis with 99.11% training and 91.25% testing accuracy.

Plant phenotyping relevance

落花生葉の病害状態を画像から推定するCNN手法の開発・比較検証とデータセット構築が中心であり、植物表現型計測に該当する。

Code and data availability

The paper's key asset is a self-generated Groundnut leaf image dataset (5,322 images, six classes) used to train the CNN8GN model, but it is not publicly deposited; the authors state it will be made available only upon request. No public code, model checkpoints, or dataset URLs are provided.

No evidence-backed public reproduction asset is currently recorded.