Performance Comparison and Analysis of Noise Reduction Methods on GÖKTÜRK-2 Satellite Images
Özet
Optical satellite imagery represents a critical data source acquired from space, serving a
multitude of application domains. However, these images often suffer from various types of
noise that can significantly impair their quality. Noise in this context refers to the random and
undesired fluctuations in image intensity, which may emerge during the processes of image
acquisition, conversion, transmission, and processing. Noise can reduce image quality and
lead to erroneous conclusions during analysis.
Consequently, it is of paramount importance to assess the efficacy of noise reduction
techniques in satellite images to enhance image quality and ensure more accurate analyses.
This study aims to eliminate thermal noise present in optical satellite images, determine
appropriate noise reduction methods, and compare these methods using specific metrics. The
noise reduction techniques employed encompass spatial domain filtering, transform domain
filtering, variational denoising methods, and deep learning-based methods. The effectiveness
and performance of these methods are evaluated based on several metrics. The tests were
carried out on high spatial resolution Goktürk-2 satellite imagery. In addition, the images acquired from the CalVal (Calibration and Validation) site and SPOT-5 satellite imagery were also used in the experimental tests.
Based on the results, each denoising algorithm has distinct strengths and weaknesses,
excelling in different aspects of image restoration. For preserving structural accuracy
and fine details Expected Path Log-Likelihood (EPLL), Denoising CNN (DnCNN), and
Fast and Flexible Denoising Network (FFDNet) are well-suited. Weighted Nuclear Norm
Minimization (WNNM) and Nonlocally Centralized Sparse Representation (NCSR) offer a
balanced performance, effectively reducing noise while maintaining structural fidelity. The
Wavelet and Bilateral Filter methods are useful for enhancing edge contrast but should be
applied cautiously due to their potential to introduce artifacts. These findings emphasize
the importance of selecting an algorithm based on the specific requirements and desired
outcomes of the task.
