Research indexJournal article
Оценка точности классификации многозональных космических снимков при использовании пиксельного и объектно-ориентированного подходов
Garafutdinova, Каличкин, Федоров
Региональные геосистемы30 Jun 2026Open access
NobleIDark:/48914/w1/20261105/2636EFAADOI10.52575/2712-7443-2026-50-2-352-365
Оценка точности классификации многозональных космических снимков при использовании пиксельного и объектно-ориентированного подходов
Garafutdinova, Каличкин, Федоров · Региональные геосистемы · 2026
NobleIDark:/48914/w1/20261105/2636EFAADOI10.52575/2712-7443-2026-50-2-352-365
Abstract
The paper presents the results of research focused on assessing the accuracy of multispectral satellite image classification using pixel-based and object-oriented approaches. The studies were conducted on the territories of Elitnaya Experimental Station and Individual Entrepreneur – Head of Peasant (Farm) Enterprise Kovalev S.M. land use areas in Novosibirsk region. Multispectral satellite images from Sentinel-2A with 10 m spatial resolution per pixel were used to implement various approaches for land cover type classification. The pixel-based approach was implemented based on supervised classification using parametric methods of maximum likelihood, parallelepiped, and Mahalanobis distance in ERDAS Imagine 2014 software package. We applied the object-oriented approach in two stages. First, image segmentation was conducted using the OBIS algorithm, and then the segmented images were classified. Segmentation was performed using SAGA GIS version 8.5.1. Machine learning models were used for classification of the obtained segmented image: Random Forest (RF), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP). In the course of the study, parametric classifiers using the pixel-based approach showed generally unstable results, which varied from satisfactory overall classification accuracy to low and very low, especially for individual classes of spatial objects. The maximum likelihood method proved to be the best, showing satisfactory overall classification accuracy (82.0-88.0%) and high Cohen's Kappa coefficient values. Non-parametric classifiers using the object-oriented approach demonstrated more stable results for both land use territories under analysis. The best among them was the MLP model with overall classification accuracy of 96.0% for Elitnaya Experimental Station land use and 91.0% for Individual Entrepreneur – Head of Peasant (Farm) Enterprise Kovalev S.M. land use, as well as high Cohen's Kappa coefficient values. The research findings show that the object-oriented approach with preliminary image segmentation is the most effective strategy for multispectral satellite image classification.