Predicting fertilizer treating of maize using digital image processing and deep learning approaches

The quality and quantity of maize yields are declining as a result of several structural issues with Ethiopia's traditional maize producing system. The lack of soil fertility, which is frequently hard to see visually from the maize leaves, is a major reason for this decline. An automated approach to identify and categorize fertility problems in maize plants is desperately needed to address this issue. The goal of this study is to develop a model for the recognition and classification of fertilizer treatment for maize based on maize leaf images, using deep learning algorithms to facilitate and improve the recognition and early control of fertilizer treatment for maize. The datasets utilized for this study were collected from various farming areas in the East Gojjam Zone, specifically the Hulet Ejju Enessie Woreda, comprising 4000 images of normal and deficient maize leaves. Through data augmentation techniques, this dataset was expanded to 16,000 images. A Convolutional Neural Network (CNN) with VGG16 and VGG19 architectures, along with a SoftMax classifier, was employed to analyze and classify the images into eight distinct categories based on their characteristics. Image enhancement and segmentation were performed using Gaussian filtering and Canny edge detection techniques. Hyperparameters, including image size, epoch number, batch size, dataset training and testing split ratio, and learning rate, were used to enhance the model's performance. The experiments with an image size of 224 × 224, 60 epochs, a batch size of 32, an 80/20 dataset split ratio, and a learning rate of 0.001 showed significant improvements in the classification model's performance. Ultimately, the best result was achieved with an accuracy of 95% in VGG16. VGG16 performed better than VGG19 in predicting fertilizer treatment for maize due to its lower complexity, which minimizes the risk of overfitting and enhances generalization, especially with smaller datasets.