Multiclass recognition of crops by a recurrent deep learning neural network with convolutional layers based on high-resolution color aerial photographs

The relevance of the study is due to the need to create an intelligent system for monitoring the state of crops during the growing season on significant areas. The scientific problem of operational monitoring of the state of crops with subsequent recognition of their condition from color images obtained from unmanned aerial vehicles (UAVs) can be solved by means of neural network algorithms of deep learning that provide analysis and processing of graphical information. As part of the computer implementation of a recurrent neural network with convolutional layers, the authors have formed a dataset of color images of crops obtained using UAVs. In order to solve the problem of class imbalance of the resulting dataset, its augmentation was carried out by artificially adding new images obtained on the basis of existing ones. For the training of a recurrent neural network implemented in Python, including convolutional layers, training and test samples were formed, with the marking of images in four classes: healthy vegetation ("healthy"), affected vegetation ("affected"), soil, unseeded field ("soil") and other objects ("other"). The recognition results obtained in the course of the study can be used to create hybrid architectures with subsequent software implementation of a complex of neural network models that allow identifying patterns of growth and development of various groups of crops.