Methodology for assessing the quality of multispectral space imaging data in landscape element monitoring

Abstract

To determine the types and current condition of agricultural fields, methods of automated processing of multispectral space imagery are increasingly being used. One of the most relevant tasks is semantic segmentation of landscape elements within the studied scene based on machine learning algorithms. Various algorithms can be used to solve this task, but the problem of evaluating the quality of processing results performed by a specific method remains open. This paper discusses the indicators that characterize the quality of results of imagery data thematic processing when monitoring the condition of agricultural fields, using fields designated for forage preparation as an example. The methodology for assessing the quality of processed multispectral space imagery data is presented. A list and numerical values of basic quality indicators for identifying the condition of agricultural fields, considering ground survey data and hyperparameter values in machine learning algorithms, are provided. Generalized quality indicators for processing results are proposed. The role of a well-founded choice of initial data for evaluating the quality of processed imagery results is highlighted. The mathematical apparatus of fuzzy clustering is applied when forming the initial data, and the degree of membership of landscape elements to a selected cluster is taken into account when refining the initial data. The presented methodology can also be applied to determining the types and forecasting the yield of agricultural crops, detecting diseases, and solving other agricultural production tasks.