Intelligent Identification and Multiclass Segmentation of Agricultural Crops Problem Areas by Neural Network Methods

The assessment of the crops state in large areas is very labor-intensive. For the intelligent analysis of agricultural fields, databases of RGB color images were formed, marked up for four classes. DNN has been developed for segmentation and identification of agricultural field areas’ problems. The features of this approach are a consistent reduction in the dimension of images, the use of convolutional layers, and the combination (concatenation) of such images with previous ones. The applicability and advantages of the segmentation approach were tested on the DeepLabV3 architecture in combination with ResNet50. Numerical runs of the DNN training procedure revealed that increasing the agricultural field segmentation accuracy according to the “Dice coefficient” criterion is limited by the quality of manual database markup. Due to this, such DNN can be used as the algorithmic core of SaaS systems for which segmentation speed is crucial.