Document Type : علمی - پژوهشی
Authors
Department of Computer Engineering, Central Tehran Branch, Islamic Azad University, Tehran, Iran
Abstract
Introduction: Detecting changes in remote sensing data is a critical task in environmental sciences, natural resource management, urban planning, and disaster management. Despite recent advancements in this field, many existing methods only address specific challenges and are unable to provide a comprehensive solution for various types of data and applications. These limitations include the inability to handle multispectral, hyperspectral, and radar data and the lack of capability to deliver accurate and timely results using parallel processing and optimized computing resources. Additionally, current methods are often confined to binary change detection and cannot accurately identify the specific types of changes. Therefore, the primary aim of this research is to develop an innovative and comprehensive change detection method that can overcome these limitations and be effectively used in real-world applications.
Materials and Methods: In this study, we propose a novel approach based on the combination of a transformer network and an automated attention model, capable of processing and analyzing remote sensing data with high accuracy and efficiency. This method utilizes multispectral, hyperspectral, and radar data obtained from Sentinel-2, QuickBird, and TerraSAR-X satellites. These data are collected over different time periods and include diverse information such as vegetation changes, land use, and structural changes. The proposed method employs feature fusion techniques using convolutional and transformer layers to integrate information from the data, thereby enhancing change detection accuracy. Additionally, the use of spatial attention mechanisms helps identify spatial relationships between features, focusing on key areas to improve change detection accuracy. The transformer-based network, developed to determine similarity, is enhanced with automated attention mechanisms that capture complex relationships between features over temporal sequences. This capability is especially important for detecting subtle changes that may be overlooked by other methods. For operational implementation, the proposed method was deployed and evaluated on a high-performance system, including a 24-core Xeon E5-2697 v2 CPU, 28 GB of memory, a 200 GB SSD, and a powerful RTX 2080 Ti graphics card with 11 GB of RAM and CUDA 11.
Results and Discussion: The results obtained from this research indicate the superiority of the proposed method compared to existing methods. Evaluations were conducted using metrics such as Precision, Recall, F1-score, Overall Accuracy (OA), and Intersection over Union (IoU). The proposed method outperformed other methods across all these metrics. Notably, overall accuracy (OA) increased significantly, reaching over 95% on some datasets. These results indicate that the proposed method can not only accurately detect binary changes but also identify the types of changed features with high precision. These capabilities are achieved through the use of advanced deep learning techniques and parallel processing. Moreover, the implementation of the SoFRB(Enayati et al. 2023) framework has enhanced the efficiency of the proposed method, enabling the processing of large volumes of data in less time. Our analysis demonstrates that the proposed method has high adaptability with different datasets and can effectively operate under various conditions. Furthermore, this method can serve as an efficient tool for various applications, including environmental monitoring, urban planning, precision agriculture, and disaster management.
Conclusion: The proposed method integrates modern deep learning techniques and parallel processing to significantly improve the accuracy and efficiency of change detection in remote sensing data. The findings of this study show that the proposed method is reliable not only in experimental settings but also in practical applications. Specifically, this method can effectively monitor environmental changes, detect alterations in urban infrastructures, and manage natural and human-induced disasters. These results promise widespread applications of this method in various fields. Future research could include further improvements in different areas, such as model optimization, the use of more diverse and extensive datasets, and the exploration of the impact of newer deep learning and parallel processing techniques.
Keywords