Introduction
Remote sensing has become a valuable tool for acquiring integrated spatial data regarding land cover and land use across various temporal and spatial scales. One of the primary challenges in multi-temporal land cover and land use mapping is the availability and integrity of training data for supervised classification algorithms. Collecting training samples for each class of land cover and land use over different time periods can be time-consuming and, particularly in rapidly changing environments, fieldwork can be challenging. This issue is further exacerbated by the potential spectral and phenological changes in land cover features over time, which can diminish the transferability of training samples. The concept of "migration" or "transfer" of training samples from a reference year to target years has been explored as a means to overcome the limitations of training data. In this context, the use of Google Earth Engine (GEE) has facilitated multi-temporal land cover and land use mapping. GEE's ability to integrate diverse data sources, including Sentinel-2 imagery and a wide range of spectral indices, enables researchers to develop robust and scalable applications for land cover and land use classification. Monitoring changes in land use and cover over time is crucial for understanding and managing the environment. However, when there are limitations in training data for different time periods, this can pose significant challenges. This study presents an innovative approach to classify Sentinel-2 satellite images from different years using a set of reference training samples.
Materials and Methods
In this study, an innovative application was explored using migrated training samples from a reference year (Sentinel-2 imagery from 2019), along with bands from Sentinel-2 images and spectral indices, to classify land cover and land use in the dynamic and ecologically significant mangrove region (Khoran Protected Area). The use of machine learning algorithms within the Google Earth Engine (GEE) framework was examined to achieve high classification accuracy and monitor land cover changes over time. Satellite images from Sentinel-2 covering the study area for the target years 2022 and 2024, as well as the reference year 2019, were retrieved in GEE. Ground truth data, including the location and classification of various land cover types for the reference year, were collected using the European Space Agency's land use maps. Subsequently, high-quality ground truth data and their corresponding image samples from the reference year (2019) were transferred to the target year images using the Spectral Angle Distance (SAD) algorithm. Classification algorithms such as Random Forest (RF), Gradient Boosting Trees (GBT), and Classification and Regression Trees (CART) were employed to classify the target year images using the variable training samples. The classified images were evaluated using various accuracy metrics, including overall accuracy, Kappa coefficient, producer accuracy, user accuracy, as well as commission and omission errors. Finally, the importance of different spectral bands from Sentinel-2 and spectral indices in the classification process was analyzed to identify the most suitable features for distinguishing various phenomena in the study area.
Results and Discussion
The results of this study indicated that among the classification algorithms, the highest accuracy for overall accuracy and Kappa coefficient in the classified images for 2024 and 2022 was achieved using the Random Forest classification, with accuracies of 0.9104 and 0.8742, and Kappa values of 0.8955 and 0.8570, respectively. Additionally, the findings revealed that the area of mangrove forests decreased from 7530.74 hectares to 6546.51 hectares during the study period, representing a reduction of approximately 984.23 hectares, or about 164 hectares per year.
The area of residential and construction zones increased from 72.52 hectares in 2019 to 96.48 hectares in 2024, indicating rapid growth in these land uses over the last two years. The synthesis and summary of various classification methods across different years highlighted the relative importance of bands and indices. Specifically, the EMVI and mNDWI indices demonstrated greater dominance due to their effectiveness in capturing the phenomena of mangrove forests and aquatic areas within the study region. Consequently, for the band combinations aimed at distinguishing various phenomena, the use of green, red, and near-infrared bands, along with the mangrove index and optimized water body index, were identified as the most suitable for the study area. These findings are recommended for similar areas in southern Iran and other mangrove regions. This research employed a simple and effective application by transferring ground truth points from the reference year to their corresponding images as training samples for the target year images, utilizing the Google Earth Engine platform. This approach holds the potential for extension to other regions.
Conclusion
In summary, this study demonstrates the potential of using migrated training samples and advanced machine learning algorithms such as Random Forest, Gradient Boosted Trees, and Classification and Regression Trees, along with spectral indices as auxiliary data, for the accurate classification of multi-temporal satellite imagery. The development of spatial tools, including the online Google Earth Engine (GEE) platform, is essential for the up-to-date management of land uses, particularly in wetland and mangrove areas. In this research, high-quality training samples were successfully migrated from the reference year to the target year, resulting in high classification accuracy using the Random Forest classification algorithm compared to other methods such as boosted regression or regression and classification trees. This method offers a viable solution in multi-temporal land use studies, especially in cases of insufficient or inadequate training samples within the GEE system. For future studies, it is recommended to employ a combination of Euclidean Distance (ED), Spectral Angle Distance (SAD), and K-means clustering for generating variable training samples, and to compare and analyze the classification results obtained through these methods. This approach presents a promising solution for producing up-to-date land use/land cover maps, even in challenging environments with limited training data. The findings of this study can guide future research aimed at monitoring land cover, managing, and effectively protecting valuable natural resources, such as mangrove forests, in other regions. |
