COMPARATIVE ANALYSIS OF CLASSIFICATION METHODS FOR HIGH-RESOLUTION OPTICAL SATELLITE IMAGES

Abstract

High-resolution satellite image classification is used in various applications, such as urban planning, environmental monitoring, disaster management, and agricultural assessment. Traditional classification methods are ineffective due to the complex characteristics of high-resolution multichannel images: the presence of shadows, complex textures, and overlapping objects. This necessitates selecting an efficient classification method for further thematic data analysis. In this study, a comprehensive assessment of the accuracy of the most well-known classification methods (parallelepiped, minimum distance, Mahalanobis distance, maximum similarity, spectral angle map, spectral information difference, binary coding, neural network, decision tree, random forest, support vector machine, K-nearest neighbour, and spectral correlation map) is performed. This study comprehensively evaluates various classification algorithms applied to high-resolution satellite imagery, focusing on their accuracy and suitability for different use cases. To ensure the robustness of the evaluation, high-quality WorldView-3 satellite imagery, known for its exceptional spatial and spectral resolution, was utilized as the dataset. To assess the performance of these methods, error matrices were generated for each algorithm, providing detailed insights into their classification accuracy. The average values along the main diagonal of these matrices, representing the proportion of correctly classified pixels, served as a key metric for evaluating overall effectiveness. Results indicate that advanced machine learning approaches, such as neural networks and support vector machines, consistently outperform traditional techniques, achieving superior accuracy across various classes. Despite their high average accuracy, a deeper analysis revealed that only some algorithms are universally optimal. For instance, some methods, such as random forests or spectral angle mappers, exhibited strength in classifying specific features like vegetation or urban structures but performed less effectively for others. This underscores the importance of tailoring algorithm selection to the specific objectives of individual classification tasks and the unique characteristics of the target datasets. This study can be used to select the most effective method of classifying the earth's surface, depending on the tasks of further thematic analysis of high-resolution satellite imagery. Furthermore, it highlights the potential of integrating machine learning-based approaches to enhance the accuracy and reliability of classification outcomes, ultimately contributing to more practical applications.