Crop Detection by Deep Learning Models From Agricultural Area Images Obtained with Drone

Creative Commons License

Arvas M. M. , Özdağ R.

1 st International Conference on International Conference on Interdisciplinary Applications of Artificial Intelligence (ICIDAAI), Yalova, Turkey, 21 - 23 May 2021, pp.13

  • Publication Type: Conference Paper / Summary Text
  • City: Yalova
  • Country: Turkey
  • Page Numbers: pp.13


Deep learning models based on artificial neural networks have led to significant developments in many areas related to artificial intelligence, primarily image processing, in recent years. One of the main reasons why deep learning models are preferred and successful is that they perform end-to-end learning. Thanks to end-to-end learning approaches, they can learn from the data itself a special set of attributes belonging to the relevant problem. In this study, the problem of detecting crops through images in agricultural areas by Drones, which are Unmanned Aerial Vehicles, was taken as basis. The purpose of the study is to classify the crops in the agricultural areas to be determined and make a distinction according to their types and thus to check whether the support given to the farmers by the state is used correctly on the crops in the area. Therefore, it is planned to prevent farmers from requesting state support over uncultivated agricultural areas. In addition, identifying products that will threaten human health by making illegal agricultural crops is among the objectives of this study. In the study, the data set will be created by accessing the images of the crops in the agricultural area with the Drone. However, in the initial phase, a total of 8775 images were used, which are the closest to the agricultural crop images and consist of 7500 training and 1275 test data provided by the ImageNet data set. First of all, feature extraction was made over ImageNet with the VGG16 architecture. A 10-layer Ensemble Neural Network (ENN) model, consisting of ResNet, Convolutional Neural Network (CNN) and VGG-16 architectures, was created. Transfer learning was performed on the Imagnet dataset using the ENN model, basic CNN and single layer VGG-16 architectures. A comprehensive analysis of the test data determined that the ENN model achieved a success rate of 99.8% and gave optimal results compared to the basic CNN and single-tactic VGG-16 architecture.