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

Introduction: Knowledge of land-use distribution and its temporal change trends constitutes one of the fundamental requirements in environmental planning and management. Understanding land-use change plays a critical role in analyzing land development processes and supporting sustainable resource management. The application of remote sensing technology in land-use studies, particularly for the extraction of land-use maps, has achieved significant success in change detection and the formulation of effective management policies. In recent years, numerous algorithms have been developed for land-use classification using remotely sensed imagery. Therefore, selecting the most appropriate classification algorithm is essential for achieving accurate and reliable results. Among probabilistic classification methods, the Maximum Likelihood algorithm is one of the most widely used techniques and is recognized as a highly accurate approach for image classification. The Support Vector Machine (SVM) algorithm is relatively insensitive to high-dimensional data characteristics and is therefore considered an effective method for the classification of multispectral and hyperspectral datasets. One of the main advantages of SVM is its ability to produce optimal classification results using relatively small training samples, which reduces operational costs while increasing classification efficiency. The Random Forest algorithm represents another powerful machine learning approach based on an ensemble of decision trees. In this method, each tree is constructed using randomly generated subsets of input data and features. Each tree independently assigns a class label, and the final classification is determined based on the class receiving the highest level of agreement among the trees.
 
Materials and Methods: In this study, multitemporal Landsat 8 Operational Land Imager (OLI) imagery acquired in 2013, 2017, and 2019 was used to extract and analyze land use/land cover (LULC) changes in the Gorganroud Basin, located in Golestan Province. Landsat 8 imagery during the study period provided stable and reliable data for mapping and detecting LULC changes. Radiometric corrections were performed using the Chavez atmospheric correction method and the Dark Object Subtraction (DOS) technique to reduce atmospheric effects and improve image quality. Furthermore, to prepare satellite imagery for processing and ensure radiometric consistency between multitemporal datasets, the histogram matching algorithm was applied. In this approach, the histogram of each image was adjusted relative to a reference image to minimize radiometric differences. Panchromatic data, together with intensity components derived from RGB multispectral imagery, were normalized through histogram matching to improve comparability among datasets. To evaluate the accuracy of the classified maps, an error (confusion) matrix was employed. Accuracy assessment requires reference data or regions of interest (ROIs), which were obtained through field surveys conducted in 2017. Within the error matrix framework, classified data were compared with reference (ground truth) data to quantify classification performance. The Kappa coefficient, widely used in remote sensing studies, was calculated to assess and compare the accuracy of different classification methods. For change detection analysis, multitemporal images were first normalized using histogram matching based on a selected base year. Subsequently, the image differencing technique was applied to identify and quantify land-use changes over time.
 
Results: Evaluation of the classification results based on accuracy assessment indices, together with comparisons against locally derived maps obtained through field surveys, demonstrates the high accuracy of the algorithms applied in this study. The findings also confirm the effectiveness of Landsat 8 imagery for extracting reliable land use/land cover (LULC) maps. The results indicate that between 2013 and 2019 forest areas experienced a slight decrease. Dryland areas showed gradual and relatively minor changes over the study period. In contrast, bare soil areas increased significantly, particularly after the initial two-year interval, suggesting a reduction in natural vegetation cover. If this trend continues, the region may face substantial forest degradation and potential conversion of forested areas into rainfed agricultural lands. The Maximum Likelihood classification method exhibited different change patterns for rainfed lands, forests, and areas with scattered vegetation cover. Under this method, forest changes appeared relatively stable, while residential areas, rainfed lands, and scattered vegetation areas showed gradual increases over time. Comparison of classification accuracy metrics among the applied algorithms indicates that the Random Forest and Support Vector Machine methods achieved higher accuracy than the Maximum Likelihood approach. Additionally, the analysis reveals a declining trend in both wetland and forest areas throughout the study period.
 
Discussion and conclusion: The results indicate that the study area is experiencing forest degradation accompanied by an expansion of dryland areas, particularly in higher elevations. These land-use changes should be carefully considered in future land management and planning strategies, as they may increase flood susceptibility in the region. Although land-cover changes during the study period occurred gradually, no abrupt or large-scale transformations were observed within the relatively short time frame analyzed. However, if the current trend continues, forested areas are likely to be converted into agricultural or rainfed lands. Such transitions may lead to increased water consumption, contributing to the degradation and shrinkage of wetland ecosystems. The reduction of water resources may further result in the expansion of unused or abandoned lands due to limited water availability for cultivation and other human activities. Comparison of the three classification methods—Random Forest, Support Vector Machine (SVM), and Maximum Likelihood—demonstrates that Random Forest and SVM produced highly consistent results, whereas the Maximum Likelihood method generated comparatively different outputs. Accuracy assessments derived from the classified maps clearly indicate that Random Forest and SVM provide more reliable and accurate classification performance. The analysis further reveals that forest cover has decreased in steep terrain, while cultivated and plowed lands have expanded. Such land-use transformations may intensify surface runoff and increase flood risk in the basin. Consequently, continued land-cover change may pose significant challenges for future water resource availability and sustainable water supply for local residents.