Evaluation of methods for multispectral space images classification

Modern approaches and methods are needed to analyze a large volume of Earth remote sensing materials. These may include methods for studying the spectral characteristics of the Earthʼs surface obtained via space systems and geoinformation technologies. High and ultra-high resolution images and development of geoinformation technologies allow reliably identifying the study area. Often, when solving the agricultural production problems, it becomes necessary to determine the intensity and nature of land use. We considered the possibilities of processing multispectral space images using supervised and unsupervised classification in relation to solving problems of agricultural activity. Methods for recognizing the types of agricultural objects using multispectral satellite images were described. Multispectral satellite images were processed independently with subsequent quantitative evaluation of the obtained results. The processing of the obtained materials included several stages: preliminary processing of the satellite image, unsupervised and supervised classification, and assessment of the reliability of the processing results. The formation of class standards was carried out manually using materials received from the farm. An analysis of the spectral separability of the obtained standards for the interpretation of multispectral satellite images was carried out. Several classification methods (ISODATA, maximum likelihood estimation, Mahalanobis distance, the parallelepiped slassification) were studied. A survey from a satellite with a high spatial resolution and vector boundaries of agricultural fields served as input data. The classification was carried out using four Sentinel-2 satellite images and Erdas Imagine 2014 software. The results showed that the selected algorithms classified objects with a probability higher than 70%, but the Mahalanobis distance method gave the most accurate results (92%).