Development of a deep neural network for segmentation of problem areas of agricultural fields
Abstract
Development of a deep neural network for segmentation of problem areas of agricultural fields
Incoming article date: 08.07.2022Artificial intelligence methods can be used to solve the problems of agricultural production. Assessing the condition of crops in large areas, even with the use of unmanned aerial vehicles, is a time-consuming task. The peculiarities of the task of such an assessment are the multifactorial nature of the analyzed structures, which require the use of a systematic approach at all stages of the study from the formation of a database of color images to the intelligent solution of problems of their analysis. The results of the analysis of the U-net architecture of the INS and its limitations for the problem of image segmentation are presented. The purpose of the study is to substantiate the architecture of the segmentation artificial neural network (INS) to identify problem areas of agricultural fields. The hypothesis of the segmentation network advantage was tested on the DeepLabV3 ResNet50 architecture. Numerical experiments have established that the increase in the accuracy of segmentation of images of agricultural fields is constrained by the limited resolution and accuracy of manual markup dataset. The built architectures can be used as an algorithmic core for creating SaaS systems, while the performance of the used configuration of the INS can be crucial.
Keywords: color images, segmentation task, agropole plots, deep neural network, INS architecture, convolutional layers