LEE- Uydu Haberleşme ve Uzaktan Algılama-Yüksek Lisans

Bu koleksiyon için kalıcı URI

http://hdl.handle.net/11527/19487

Gözat

Konu "artificial intelligence" ile LEE- Uydu Haberleşme ve Uzaktan Algılama-Yüksek Lisans'a göz atma
  • Öge
    Investigating the performance of super-resolved remote sensing images on coastline segmentation with deep learning-based methods
    (ITU Graduate School, 2025) Pala, İlhan ; Algancı, Uğur ; 705231005 ; Satellite Communication and Remote Sensing
    Satellite imagery plays a crucial role in a wide range of applications such as cartography, agriculture, urban planning, environmental monitoring, and disaster management. However, one of the major limitations in Remote Sensing (RS) analysis is the insufficient spatial resolution due to technical constraints of imaging sensors. This limitation affects the accuracy and reliability of tasks such as object detection, classification, and change monitoring. Another significant challenge arises in coastline segmentation, which requires precise boundary detection between land and sea—an inherently dynamic and complex interface that is sensitive to natural processes and atmospheric conditions. The primary objective of this study is to enhance the spatial resolution of Landsat-8 and Sentinel-2 satellite images using Super-Resolution Generative Adversarial Network (SRGAN) and to perform coastal line segmentation on the enhanced images. SRGAN utilizes a generator-discriminator architecture to generate High-Resolution (HR) images from Low-Resolution (LR) images. LR images were generated by downsampling remote sensing images with scaling factors of 2 and 4, matched appropriately, and enhanced by data augmentation techniques including rotation and cropping for training, . Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Cosine Similarity (CS), Correlation Coefficient (CC), and the Relative Dimensionless Global Error in Synthesis (ERGAS) evaluation metrics were used to evaluate image quality. The results indicate that SRGAN demonstrates remarkable performance, especially at a scale factor of 2. Histogram matching was implemented after SR processing to enhance contrast and improve visibility of fine details. The integration of SRGAN and histogram matching process are significantly improved image clarity and interpretability. Thereby supporting more accurate analysis in coastal monitoring applications. In the subsequent phase, three deep learning models with different architectures, U-Net, LinkNet, and DeepLabV3+ were implemented to execute coastal segmentation. The symmetric encoder-decoder design of U-Net enhances the segmentation of thin and complex coastal boundaries by preserving spatial features through skip connections. LinkNet achieves precise and efficient segmentation by minimizing feature loss through direct encoder-decoder connections. In order to refine boundary detection and capture multi-scale contextual information, DeepLabV3+ uses Atrous Spatial Pyramid Pooling (ASPP). Model performance was evaluated using Intersection over Union (IoU), Dice Coefficient, Dice Loss, Accuracy, Precision, Recall, Specificity, and F1 Score. Experiments were conducted on LR, SR, and Gokturk-1 (GT-1) images. Results show that segmentation accuracy improves with super-resolution. DeepLabV3+ achieved the highest performance on LR images, while LinkNet outperformed others on Super-Resolved (SR) images by correcting misclassifications. For GT-1 images, U-Net and LinkNet yielded similar high accuracy, surpassing DeepLabV3+. Coastline length estimation further validated model accuracy, using a manually digitized reference length of 3412.00 meters. On LR images, LinkNet produced the lowest relative error, at 7.29%, which reduced to 6.33% on SR images. On GT-1 images, LinkNet achieved a relative error of just 0.68%, demonstrating its effectiveness in precise coastline segmentation. Overall, the study confirms that both segmentation model choice and image resolution critically affect the accuracy of coastal line detection. The combination of SRGAN-based super-resolution and deep learning segmentation, particularly with the LinkNet model, enables highly accurate and reliable coastline segmentation from satellite imagery.