Özet:

Maize is used as a staple food in many parts of the world. Maize is a good source of starch that can be used to make oil and high fructose corn syrup. Since maize and maize flour are very vital products, early diagnosis of some diseases such as Blight, Rust, and Gray Leaf Spot can prevent the deterioration of product quality by improving the plants. In this way, both the quality of the product and the amount of product obtained can be increased. In this study, feature extraction was performed from corn leaf images using Convolutional Neural Network (CNN) based VGG-19, DenseNet-201, and NasNet-Large models. In order to eliminate the imbalance in the data set, the balance was adjusted with the Synthetic Minority Over-Sampling Technique (SMOTE) method. Principal Component Analysis (PCA), one of the dimension reduction methods, was used to reduce the number of features. Support-Vector Machines (SVMs) algorithm was used to classify diseases in maize leaves. With the GridSearchCV approach to improving the performance of the algorithm, the Kernel function and Box constrain hyperparameters have been optimized to identify diseases in corn leaves. The experimental results obtained were tested on the publicly accessible Kaggle corn or maize leaf disease dataset. In the experimental results obtained, 94.5%, 94.4%, 94.3%, and 96.2% accuracy rates and a weighted average of 94.3% were achieved for the 4 classes, respectively, in the classification of the images with only CNN and feature extraction with LibSVM. Using the proposed method, 96.7%, 96.7%, 96.7%, and 97.8% accuracy rates and 96.7% weight average were achieved for 4 classes with LibSVM, respectively. Thus, according to the classification accuracy obtained without optimization in the classification accuracy obtained with the proposed method, it was observed that there was an improvement of 2.2% for the first class, 2.3% for the second class, 2.4% for the third class, and 1.6% for the fourth class, and 2.4% in the weighted average.