"Performance analysis of Support Vector Machine, Random Forest, and Maximum Likelihood algorithms in land use classification of the metropolitan area of Mashhad."

Introduction: Considering that the value and usability of any map produced from satellite images depend on its accuracy, evaluating the accuracy of satellite image classification methods is of great importance. Therefore, this research aims to analyse the performance of Support Vector Machine (SVM), Random Forest (RF), and Maximum Likelihood Classification (MLC) algorithms in identifying land use and land cover (LULC) in the metropolitan area of Mashhad. Numerous algorithms have been developed for satellite image classification to date, and their performance varies under different conditions. For this reason, this study first identifies the most commonly used algorithms through a review of previous research, and then, by assessing the characteristics of various classifiers, selects the three algorithms: Support Vector Machine, Random Forest, and Maximum Likelihood. There are various studies regarding the performance of different classification algorithms, each yielding different results. Given that multiple studies have shown that LULC mapping accuracy is related to time and location, and that each of these studies has emphasized the accuracy of different algorithms, their results cannot be generalized to the geographical conditions of Iran. On the other hand, there has not been sufficient research in the geomorphological conditions of Iran to assess the accuracy of classification algorithms, and most studies validating these algorithms have been conducted in case studies outside of Iran. Therefore, considering the differences in algorithm results under various conditions, examining the accuracy and performance of these algorithms focusing on the extensive and diverse metropolitan area of Mashhad may yield novel and noteworthy findings.

Materials and Methods: The present research is applied in terms of purpose and descriptive-analytical in terms of nature. Data collection in this study has been conducted through a documentary-library method. In this study, images from the OLI sensor on the Landsat 8 satellite were used. The classification of satellite images was performed in two stages: preprocessing and processing. After assessing the accuracy of the classification using the Kappa coefficient, confusion matrix, coefficient of variation, and User's accuracy and Producer's accuracy coefficients, the best algorithm for classifying land uses in the metropolitan area of Mashhad was determined in five classes: 1- Built-up areas, 2- Barren land, 3- Mountainous areas, 4- Green spaces, and 5- Water bodies.

Results and Discussion: The results from the evaluation of standard deviation (SD) and coefficient of variation (CV) regarding the area share percentage in a LULC class by various algorithms indicate that barren lands were classified with higher accuracy, while water bodies and green spaces were classified with lower accuracy. The examination of U_Accuracy and P_Accuracy coefficients shows that the overall accuracy of the classification for all studied algorithms falls within the range of good to excellent. However, a more detailed examination of these algorithms reveals that the greatest challenge in class identification lies in built-up areas, mountainous regions, and green spaces, whereas the identification of barren lands faces fewer challenges. The Kappa coefficient and analyses based on the confusion matrix also demonstrate the variation in accuracy among each LULC classifier. The differences in the accuracy of the classifiers used are marginal, but these slight variations hold significant importance in the context of LULC planning. Given that these marginal differences are evident in sensitive land uses such as built-up areas and green spaces, selecting an algorithm with the highest accuracy and lowest error is of special importance.



Conclusion: The results of the Kappa coefficient evaluation and confusion matrix analyses indicate that the SVM approach has greater overall accuracy and a higher Kappa coefficient compared to RF and MLC methods. Specifically, the algorithms achieved overall accuracies of 0.93, 0.88, and 0.80, respectively. Therefore, Support Vector Machine demonstrates the highest accuracy and least error among the studied classifiers. Considering that numerous studies have shown that LULC mapping accuracy is related to time and location, it is suggested that future research analyse the accuracy of classifiers under different morphoclimatic and geomorphic conditions.