Determination of diseases from leaf images using deep learningmethods

Convolutional Neural Networks (CNNs), which form the basis of artificial intelligence, are designed to resemble the structure of human neural networks in Deep Learning (DL) models. Intelligent computers and software systems for prediction, recommendation, and decision making have become very important with the development of artificial intelligence. In many fields (agriculture, health, cybersecurity, defense industry, transportation, etc.), deep learning models can be used to make calculations with high predictive power and accuracy. In particular, deep learning models and analysis systems are widely used to produce high-quality and accurate results in areas where manpower and logic are limited when classifying large and complex data. Training deep learning models using large data sets and automating disease detection processes increase productivity in agriculture and encourage farmers to use the right techniques. The use of these technologies in agriculture increases precision, enabling time- and cost-effective, high-quality crop production. Aerial imagery is accelerating with the development and increased use of unmanned aerial vehicles (UAVs). In addition, the use of high-resolution images taken using advanced cameras in combination with efficient software and image processing techniques in high-tech computer systems makes it possible to automatically detect plant quality and health. This study aims to determine the general health of potato plants using a dataset of potato plant images taken by UAV from agricultural lands and to determine the applicability of this method to other plant species. In addition, the applicability of this method to other plant species is also investigated with a data set consisting of plant images taken with a camera and UAV from agricultural lands in the Southeastern Anatolia Project (GAP) region. The dataset, consisting of healthy, early blight, and late blight leaf images, was retrained by the transfer learning method on VGG19, DenseNet201, InceptionV3, and InceptionResNetV2 deep learning models that were previously trained with the ImageNet dataset. In addition, deep, medium, and shallow features were extracted and retrained with Support Vector Machines (SVM) and K-Nearest Neighbor (K-NN) Machine Learning (ML) models using the feature extraction method by applying hyperparameter optimization. Performance results are compared according to classification metrics. In the feature extraction method, the highest classification success rate was found in the DenseNet201 model with K-NN, and the lowest classification success rate was found in the InceptionResNetV2 model with SVM. According to the transfer learning method, the highest classification success rate was obtained for the Adam-optimized DenseNet201 model, and the lowest classification success rate was obtained for the Sgdm-optimized InceptionResNetV2 model. In general, the InceptionResNetV2 model gave the longest training time, while the VGG19 model gave the shortest training time. In the study, it was concluded that the health of any cultivated plant species can be successfully classified using image data from agricultural lands.