DEVELOPMENT OF A NOVEL MULTI-CRITERIA METHOD USING DEEP LEARNING AND OPTIMIZATION FOR IMAGE CLASSIFICATION Tsolmonbayar.Sh1, Bayanjargal.D2*, Batchuluun.Ts1, Tsolmon.R3, Davaajargal.J2, Selenge.M1 and Bayanmunkh.N4,5
1School of Engineering and Applied Science, National University of Mongolia, Ulaanbaatar, Mongolia
2School of Information and Electronic Engineering Applied Matematics National University of Mongolia, Ulaanbaatar, Mongolia (*bayanjargal[at]num.edu.mn)
3School of Art and Sciences Physics department, National University of Mongolia, Ulaanbaatar, Mongolia
4Centre for Policy Research and Analysis of Ulaanbaatar Municipality, Ulaanbaatar, Mongolia
5Mongolian Geospatial Association, Ulaanbaatar 15141, Mongolia
Abstract
Recent advancements in remote sensing image analysis have increasingly utilized deep learning models, resulting in notable improvements in classification accuracy and computational efficiency. Among these approaches, hybrid methods that combine deep learning with optimization techniques have shown superior performance over conventional single-model algorithms. In this study, we propose a novel classification algorithm called Multi-Criteria Mean Clustering (MCMC). This method integrates deep learning-based feature extraction with a multi-objective optimization framework, enabling it to better capture the diverse characteristics of high-dimensional and heterogeneous remote sensing data. By considering multiple criteria-such as spectral separability, spatial coherence, and class distribution-MCMC enhances clustering robustness and interpretability. The proposed method was applied to a case study in Dornod Province, Mongolia, a region along the Siberian forest boundary known for its complex land cover structure and ecological significance. We used Sentinel-2B multispectral imagery to perform land cover classification. To validate the classification performance, results from MCMC were compared against NDVI-based ground truth data. Correlation analysis revealed a 98% agreement between the MCMC outputs and the NDVI-derived reference map. Additionally, MCMC was benchmarked against two commonly used techniques: Mini-Batch K-Means, known for its scalability, and Random Forest, a widely adopted supervised classification method. Comparative results showed that MCMC either matched or exceeded the performance of these methods, particularly in terms of class boundary delineation and intra-class homogeneity. These findings demonstrate the potential of the MCMC approach in addressing the limitations of existing clustering and classification techniques, especially for complex and heterogeneous remote sensing datasets.
Keywords: Multi-Criteria, deep learning, optimization, classification
