ارزیابی مناسب‌ترین الگوریتم یادگیری ماشین در طبقه‌بندی تصاویر ماهواره‌ای و آشکارسازی تغییرات کاربری/ پوشش اراضی در حوضه آبخیز گرگان‌رود

Introduction: Knowledge of land use/land cover (LULC) ratios and their temporal change trends constitutes a fundamental prerequisite for environmental planning. Understanding land use change is of particular importance in the context of land development processes. The application of remote sensing science to land use studies and the extraction of land use maps has proven successful in change detection and the formulation of associated management policies. In recent years, various algorithms have been developed for classifying different land use categories using remote sensing imagery. Consequently, identifying the optimal algorithm for a given classification approach is critical to obtaining accurate outputs. Among probability-based algorithms, Maximum Likelihood is the most common and accurate technique and is recognized as one of the most precise classification methods. The Support Vector Machine (SVM) algorithm is less sensitive to multidimensional phenomena. As a result, it is considered an appropriate method for the classification of multispectral and hyperspectral data. One of the advantages of the SVM algorithm is its ability to produce an optimally classified map using a limited number of training samples. Accordingly, this reduces costs and increases classification speed. Random Forest is another machine learning algorithm, based on a complex ensemble of decision trees. In this method, each classification is derived from a random vector that is independent of the input. Each tree assigns a separate class to every vector that exhibits the highest correspondence with that class.
Material and Methods: In this research, Landsat 8 (OLI) multitemporal image data from 2013, 2017, and 2019 were utilized to extract land use/land cover (LULC) change detection in the Gorganroud basin, located in Golestan Province. The Landsat 8 data within the study period represent the most stable images for extracting and mapping LULC change detection. Radiometric corrections were applied using the Chavez and Dark Subtraction methods. Furthermore, the Histogram Matching algorithm was employed to prepare satellite imagery for processing purposes. In this method, recent panchromatic data, along with intensity components extracted from RGB multispectral data, were matched using the histograms of the data. To assess the accuracy of the extracted maps, an error matrix was utilized. Accuracy assessment required ground truth images or regions of interest, which were obtained from a field survey conducted in 2017. In the error matrix, the raw data were compared with the classified data. In most studies, the Kappa coefficient is employed to evaluate the accuracy of results derived from different classification methods. To investigate and detect changes, the histogram matching method of multitemporal images was first applied based on a reference year, followed by the image difference method.
Results: Evaluation of the results derived from accuracy indices, along with comparison to the locally derived map obtained through in-person field monitoring, demonstrates the high accuracy of the algorithm applied in this study. Furthermore, it confirms the effectiveness of Landsat 8 imagery in extracting land use/land cover (LULC) maps. The results reveal that, between 2013 and 2019, forest lands experienced a slight decrease. Dry lands also underwent gradual changes. Bare soil areas increased significantly; however, after two years, their area expanded further, indicating a reduction in natural plant cover. Should this trend continue, the region will face forest loss and conversion to rainfed agricultural lands.
 
 In the Maximum Likelihood method, rainfed lands, forests, and areas with scattered plant cover followed distinct trends. The forest change trend remained stable, whereas residential areas, rainfed lands, and scattered plant cover areas increased gradually. Based on the accuracy values obtained from the Random Forest and Support Vector Machine methods, it can be concluded that these methods are more accurate than the Maximum Likelihood method. In contrast, wetlands and forests exhibit a declining trend.
Discussion and Conclusion: The study area is facing forest destruction and an increase in dry lands at higher elevations. Consequently, land use/land cover change detection must be taken into account in the future to prevent and control flooding in the region. Accordingly, it can be stated that during the study years, land cover in the region has changed gradually; however, no pronounced or significant changes have occurred over the short term. If the current trend continues, it is predictable that forests may convert to agricultural or rainfed lands. Such a transformation could lead to increased water consumption and the decline of wetlands. As water resources diminish, unused lands are likely to expand due to a reduced capacity for water supply. The results of comparing the three classification methods—Random Forest, Support Vector Machine (SVM), and Maximum Likelihood—indicate that Random Forest and SVM yield more comparable results, whereas Maximum Likelihood produces different outputs relative to these two methods. Based on the extracted maps, it is evident that Random Forest and SVM provide more accurate and reliable results. From these findings, it can be concluded that in steeper areas, forests are decreasing while plowed lands are increasing, which may contribute to flooding in the region. Therefore, future water supply for the residents is likely to become a critical issue.