Neural Network-Based Classification for Automated Powdery Mildew Detection in Modern Tomato Greenhouses

Modern greenhouses are characterized by scale, standardization, and efficiency. Some indoor agricultural facilities spread across over
$100000~{\mathrm {m^{2}}}$
while being inhabited by dozens of species in a fragile balance. Timely monitoring is crucial for these greenhouses, and both speed and precision are required to minimize the potential yield loss. Human inspection is prone to error and time-consuming, thus expensive. On the other hand, collecting samples and analyzing them in the laboratory outside the agricultural facility is even more time-consuming and economically infeasible. Consequently, inspection by an autonomous robot that performs visual plant analysis gains traction. This work is devoted to the data acquisition and training of a neural network-based classifier that solves the problem of powdery mildew identification. The data collection was performed with user-grade RGB cameras while in motion. The infection identification accuracy of 85% was reached with a small computational load on the robot’s computer, which proves that the target problem could be solved under challenging circumstances. In order to carry out the experiments, an autonomous robot capable of greenhouse monitoring was developed and tested in a real environment. It is supposed to be deployed in modern greenhouses equipped with rails.