System Development for Liquid Chemicals Point Injection Based on Convolutional Neural Network Models

The authors showed that one of the reasons for the yield loss is poor-quality determination of the infection degree of agricultural crops by pathogens. They proposed a system of liquid chemicals point application. They identified the possibility of calculating the required amount of fertilizers and protective equipment. (Research purpose) To develop a system of liquid chemicals point application for plant protection and nutrition based on a convolutional neural network model. (Materials and methods) The authors analyzed the existing methods of machine learning. When developing the system, they used the U-net-algorithm of convolutional neural networks, as well as data displaying diseases of winter and spring wheat – brown rust and powdery mildew. Each image was cropped by hand and marked up using a specialized Python library. In the course of applying the architecture, the authors experimentally chose the optimal metrics (jaccard metric), the learning rate – 0.0001 seconds, the number of epochs – 300, and other indicators. (Results and discussion) The authors found that when a new, previously unavailable image was submitted to the algorithm, it recognized the disease in a few seconds and returned to the user not only the original image, but also a mask over it. The accuracy of applying the mask to the affected area was determined – 80 percent. They showed that the predicted error on the validation data was 0.18758. In practice, it could differ from the declared one by no more than 10-15 percent. The authors suggested using the algorithm with a vision system. (Conclusions) The authors showed that technical means imperfection for plants chemicalization increased the consumption up to 30 percent relative to the volume required for point application. They developed a neural network algorithm for identifying the affected areas of plants and proposed the concept of a point chemicals application in order to reduce the costs of processing crops. It was determined that the neural network was able to diagnose the affected areas of plants in 1 second.