Abstract
Changes in agricultural land use driven by human or natural factors and their monitoring constitute one of the most important challenges in the northern provinces of the country. Different machine learning algorithms exhibit varying performance in regions with specific spatial characteristics. The aim of the present study was to use Sentinel-2 satellite imagery to classify land use in Masal County and to compare the maximum likelihood, minimum distance, Mahalanobis distance, and support vector machine classifiers with linear, polynomial, sigmoid, and radial basis function kernels.
To provide training and testing data for the algorithms, in addition to ground truth points collected using a global positioning system (GPS) and Google Earth maps, spectral indices—including the normalized difference vegetation index (NDVI), the normalized difference built-up index (NDBI), and the normalized difference water index (NDWI)—were used to more accurately determine ground truth information. Model validation was performed using a confusion matrix and by calculating the kappa coefficient and overall accuracy. The results showed that the maximum likelihood algorithm achieved the most accurate classification performance, with an overall accuracy of 97.29% and a kappa coefficient of 0.96, while the support vector machine with a linear kernel ranked second, with kappa coefficient and overall accuracy values of 0.89 and 92.46%, respectively. The weakest performance was observed for the minimum distance algorithm, with an overall accuracy of 72.97% and a kappa coefficient of 0.66. Based on the results of this study, the maximum likelihood classifier provided superior performance in discriminating land use classes in Masal County. The findings of this research are of significant importance for policymakers and provincial managers in increasing awareness of changes in agricultural and forest land use in the region.

EXTENDED ABSTRACT
Introduction 
Land use and land cover change, driven by either anthropogenic activities or natural dynamics, is a critical environmental challenge in many regions of the world, particularly in ecologically sensitive and agriculturally significant areas such as the northern provinces of Iran. Monitoring these changes is essential for sustainable resource management, environmental protection, and regional planning. Masal County, with its diverse landscape including forested areas, agricultural lands, and urban settlements, represents a case where accurate and timely land use classification is vital for informed decision-making. Remote sensing technologies, particularly multispectral satellite data like Sentinel-2, have emerged as powerful tools for capturing spatial and temporal variations in land use. The application of machine learning algorithms has further enhanced the potential for precise classification; however, the performance of these algorithms can vary significantly depending on the spatial characteristics of the study area. Thus, selecting an appropriate classification method is crucial for reliable land cover assessment. This study aims to evaluate and compare the performance of four common classification algorithms, Maximum Likelihood, Minimum Distance, Mahalanobis Distance, and Support Vector Machine (SVM) with various kernel functions (linear, polynomial, sigmoid, and radial basis function), for classifying Land use and land cover in Masal County using Sentinel-2 imagery. 
Material and Methods 
The methodology of this study involved the acquisition and processing of Sentinel-2 satellite imagery covering Masal County. Ground truth data were established through a combination of GPS-based field surveys and interpretation of high-resolution Google Earth imagery. To enhance the accuracy of land use classification, several spectral indices were calculated, including the Normalized Difference Vegetation Index (NDVI), Normalized Difference Built-up Index (NDBI), and Modified Normalized Difference Water Index (MNDWI). These indices helped in distinguishing between vegetation, built-up areas, and water bodies, respectively. Four classification algorithms were tested: Maximum Likelihood, Minimum Distance, Mahalanobis Distance, and SVM with linear, polynomial, sigmoid, and radial basis function (RBF) kernels. Model validation was conducted using confusion matrices, with overall accuracy and Kappa coefficient serving as the primary evaluation metrics.
Results and Discussion 
The comparative evaluation of the four classification algorithms Maximum Likelihood, Minimum Distance, Mahalanobis Distance, and Support Vector Machine (SVM) with various kernels revealed substantial differences in their ability to accurately classify land use types in Masal County using Sentinel-2 imagery and ancillary spectral indices. Among the tested methods, the Maximum Likelihood classifier demonstrated the highest performance, achieving an overall accuracy of 97.29% and a Kappa coefficient of 0.96. This high accuracy indicates the algorithm's strong capability in modeling the statistical distribution of pixels for each land use class, particularly when sufficient and representative training data are available. The confusion matrix for this classifier revealed minimal misclassification between major land cover types, including agricultural lands, forests, built-up areas, and water bodies. The high values for both overall accuracy and Kappa coefficient suggest that the Maximum Likelihood approach is robust for the heterogeneous landscape of Masal County. The SVM classifiers, especially with a linear kernel, also performed well, with an overall accuracy of 92.46% and a Kappa coefficient of 0.89. This suggests that the land cover classes in the study area are largely linearly separable in the feature space constructed from the spectral bands and indices. The SVM with other kernels polynomial, sigmoid, and radial basis function (RBF) showed slightly lower accuracies, indicating that increased model complexity did not necessarily translate to better discrimination in this particular dataset. The linear kernel's strong performance highlights its suitability for remote sensing applications where class boundaries are relatively straightforward. The Mahalanobis Distance classifier achieved moderate results, with overall accuracy and Kappa values lower than those of Maximum Likelihood and SVM, but higher than those of Minimum Distance. This method benefited from considering the covariance among spectral features, which helped in distinguishing classes with overlapping spectral signatures. However, its performance was limited in areas where class distributions were not well separated. The Minimum Distance classifier exhibited the weakest performance, with an overall accuracy of 72.97% and a Kappa coefficient of 0.66. Its reliance on simple Euclidean distances made it less effective in capturing the spectral complexity of the region, leading to higher rates of misclassification, particularly between agricultural and urban classes. Incorporating spectral indices such as NDVI, NDBI, and MNDWI significantly improved the classification results for all algorithms. These indices enhanced the separability of vegetation, built-up, and water classes, reducing confusion and increasing the reliability of the classified maps. 
Conclusions 
The findings of this study underscore the crucial role of algorithm selection in remote sensing-based land-use classification. Among the tested methods, the Maximum Likelihood classifier consistently delivered the highest accuracy and reliability for distinguishing land use types in Masal County, followed closely by the SVM with a linear kernel. The use of spectral indices further enhanced the classification results, emphasizing the value of integrating ancillary data for improved land cover discrimination. These outcomes offer practical guidance for land managers, policymakers, and researchers seeking to monitor and manage land-use changes in northern Iran and similar regions. Accurate land use maps are essential for tracking agricultural expansion, forest degradation, and urban growth, all of which have significant implications for environmental sustainability and regional planning. Future research should explore the integration of additional data sources and advanced machine learning techniques to refine classification accuracy further and support dynamic land use monitoring.
Author Contributions
Momeni Masouleh Seyed Abdollah: Data collection, Software, Visualization, Original Draft
Rahimi-Ajdadi Fatemeh: Conceptualization, Methodology, Validation, Writing - Review & Editing, Supervision, Project administration
Data Availability Statement
This section details where supporting data can be found, typically including links to publicly archived datasets. If no data is reported, "Not applicable" should be stated.
Acknowledgements
The authors would like to appreciate the support of the University of Guilan.
Ethical Considerations
This section states ethical approval details (e.g., Ethics Committee, ethical code) and confirms adherence to ethical standards, including avoidance of data fabrication, falsification, plagiarism, and misconduct.