Assessment of Land Use Changes Based on the Integration of Machine Learning Method and Spectral Angle Mapper Algorithm Using Training Samples Migration: A Case Study of Anzali Wetland Basin

ABSTRACT
Given the significance of land use changes in spatial planning and the conservation of critical ecosystems such as wetlands, this study aims to analyze land use changes in the Anzali Wetland basin by integrating the Spectral Angle Mapper (SAM) algorithm with the Random Forest (RF) classifier, utilizing dynamic training samples within the Google Earth Engine (GEE). For this purpose, harmonized Sentinel-2 imagery from 2019–2023 and six spectral indices were employed to enhance classification accuracy. By collecting 500 ground points in the base year and using spectral angle difference analysis, new training samples were generated for 2021 and 2023, and classification maps were produced using the RF algorithm. The results show that over these five years, the most significant land use changes were a decrease in water bodies and an increase in wetlands and built-up areas. The modeling outcomes demonstrated an overall accuracy and kappa exceeding 87% for the study period. Additionally, the water body class exhibited the highest user and producer accuracy, exceeding 90%. The results of the relative importance of bands and indices also highlight their role in enhancing the accuracy of the generated maps. It was found that the green, blue, and red bands, along with the MNDWI, had the greatest effect on land use discrimination and the transfer of training samples. Based on the research findings, the hybrid method, incorporating dynamic sampling and automated sample generation, can effectively improve the accuracy of land use classification in wetlands. Therefore, it is a reliable and applicable method for future studies in other wetland basins.
Wetlands are among the most significant aquatic bodies that interact with both natural and human ecosystems, providing diverse ecosystem services. Over the past century, more than half of the world's wetlands have disappeared, despite their ecological significance. Anzali International Wetland, which is listed under the Ramsar Convention, is one of the wetlands experiencing degradation due to stress factors such as climate change and human activities. These pressures have resulted in a decline in both the quantity and quality of its water body, leading to habitat loss and environmental deterioration. Understanding and analyzing land use changes in the watershed draining into the wetland, combined with spatial planning and environmental management, can help to mitigate wetland degradation.
Thanks to the progress in satellite sensor technology, the assessment of land use changes has become increasingly feasible, offering significant time and cost savings compared to traditional methods. However, selecting an efficient classification method and ensuring its accuracy remain critical challenges. A review of previous studies indicates that although supervised classification techniques generally outperform other methods, no universally optimal approach has yet been identified for accurately classifying land use in wetland watersheds. Furthermore, while the Google Earth Engine (GEE) cloud platform offers distinct advantages over software like ENVI, it has been underutilized in wetland studies. Additionally, newer integrated approaches, such as combining machine learning algorithms with the Spectral Angle Mapper (SAM) method—designed to detect spectral differences between land cover types—have not been specifically applied to monitoring wetland changes.
Another limitation of previous studies is the uniform approach to training sample collection, which is typically conducted manually through multiple ground-truth surveys. The classification models in these studies rely on predefined land use labels, potentially limiting adaptability. To address these gaps, this study, for the first time, utilizes a time series of harmonized Sentinel-2 imagery with a 10-meter resolution to assess land use changes in the Anzali Wetland watershed. Moreover, it is the first study to implement a new hybrid methodology on this platform by integrating the SAM algorithm with the Random Forest machine learning classifier, incorporating dynamic training samples to automatically generate land use classification maps for target years based on a base-year map. Additionally, the study evaluates the relative importance of spectral bands and indices to determine their contribution to land use classification within the study area.
 
Methodology
Summer season data (2019–2023) for the Fumanat sub-watershed were collected using the GEE platform and harmonized Sentinel-2 imagery. To enhance land use class separability, various spectral indices and bands were incorporated into the dataset. A total of 500 reference data points for the base-year map (2019) were sampled via Google Earth, corresponding to different land use classes. Subsequently, the SAM algorithm was applied alongside reference data, preprocessed base-year and target-year (2021 and 2023) images, and spectral angle calculations for land use classes to generate new training samples for the target yearsOf these samples, 70% were utilized for training the Random Forest model, while the remaining 30% were used for accuracy assessment based on overall accuracy, kappa coefficient, and other validation metrics. Finally, the relative importance of spectral bands and indices was evaluated based on their impact on classification performance.
 
Results and discussion
The results indicate that the new hybrid approach enhances the accuracy of land use classification maps in complex environments compared to previous methods. The mean values of overall accuracy and kappa coefficient demonstrate a high classification accuracy exceeding 85%. Furthermore, since the user’s and producer’s accuracy for all land use classes remained above 70%, confirming reliable training sample transfer and effective class separability. Additionally, the transfer of the training sample and the automated model training helped minimize human-induced errors in sample collection. The estimated land use area and percentage of changes over the study period revealed a 43% increase in built-up areas, while the Anzali Wetland water body experienced a decline of more than 28%. Analyzing the importance of spectral bands showed that in 2019, the red band had the highest effect on classification accuracy and sample transfer, while in 2021, the blue band played a more significant role. In 2023, the Modified Normalized Difference Water Index (MNDWI) proved to be the most influential factor in distinguishing land use classes and optimizing sample transfer accuracy.
 
Conclusion
The application of the SAM algorithm and spectral angle analysis between the base image and target images facilitates the automated generation of dynamic training samples. This approach significantly enhances the separability of land cover features in classification mapping, particularly in environments with high land use complexity and diversity, such as wetlands, compared to static training sample selection. Additionally, integrating this method with the Random Forest classifier improves model accuracy in land use classification. The observed land use change trends in the study area highlight the urgent need for conservation as well as sustainable restoration initiatives for Anzali Wetland, one of Iran’s and the world's most critical ecosystems. Therefore, it is recommended to conduct ecological capacity assessments and implement governance-based policies with the participation of stakeholders to ensure informed decision-making. Furthermore, revising spatial planning strategies and shifting economic policies that promote urban and industrial expansion should be prioritized to enhance the resilience of vulnerable resources such as Anzali Wetland. These measures are essential for halting its degradation and initiating long-term restoration initiatives.
 
Funding
This work is based upon research funded by Iran National Science Foundation (INFS) under project No. 4033513