Detection of diseases in rice plant with deep learning methods and artificial network based NDVI values

Considering the worldwide food consumption, it is seen that rice has an important place. The rice plant is the most cultivated plant after corn and wheat from the grass family. Production loss is experienced in paddy production due to various pests and diseases. There are three main diseases that cause these losses: Leaf Blast (Pyricularia grisea), Brown Spot (Helminthosporium oryzae), and Hispa (Dicladispa armigera). All these disease symptoms are found in the leaves of the plant. The automatic diagnosis of plant disease from leaf image is a topic under development. In this thesis, Convolutional Neural Network (CNN), one of the deep learning models, was used to detect these three diseases of the rice plant. The RGB data set of 3341 rice plant leaves used in the study was obtained from the Kaggle website. With the deep learning model from RGB leaf image of rice plant, three important diseases of rice plant (Brown Spot, Leaf Blast and Hispa) were detected with high accuracy rates. For the early detection of diseases of rice plant, the training of the CNN model was carried out with the most appropriate hyper parameters found as a result of the experiments. For the training of the network, an accuracy of 92.78% was obtained by using RGB image of the rice plant. Normalized Difference Vegetation Index (NDVI) values obtained from multispectral cameras and satellite image are generally used to examine the health status of crops in agricultural areas. High cost and suitability of weather conditions come to the fore in remote sensing techniques used in obtaining NDVI data. Another problem is the high cost of multispectral camera systems integrated into unmanned aerial vehicles (UAV), which is another alternative, and the need for specialists. In this thesis, secondly, a new artificial neural network model is proposed to estimate the NDVI value (nNDVI: Neural network-based Normalized Difference Vegetation Index) from camera systems that provide standard RGB image instead of multispectral cameras. The data set used in the training and testing of the proposed network was obtained from image taken with a multispectral camera from an agricultural field in Switzerland and a farm in Togo. Thanks to this model, NDVI data was obtained from RGB image with an accuracy rate of 92,013%. Finally, within the scope of this thesis, nNDVI data was obtained from standard RGB image of rice plant, and a deep learning model was trained with these data and an accuracy rate of 96.97% was achieved. This success rate obtained for the detection of the disease with leaf image in the rice plant shows the applicability of the method.