A Comparative Study of Noise Removal Filters for Satellite Imagery
Satellite images pick up noise from detectors, electronics, radar physics and calibration drift, and no single filter copes with all of it. Published comparisons are narrow in imagery, noise model, metrics and cost reporting, so their rankings cannot be read across studies. This paper compares the main families of noise removal filters used on satellite imagery: mean and Gaussian smoothing, median filtering, local-statistics filters, wavelet and curvelet thresholding, non-local methods and convolutional neural networks. The comparison draws on published experiments with Landsat 8, Sentinel-2, radar, aerial and planetary images, supplemented by recent 2026 studies, and is organized by noise type, fidelity, preservation of edges and texture, and computing cost. Its contributions are an evidence-based statement of the research gap, a multi-criteria comparison in which every reported difference is re-derived from the source tables, evidence-graded selection guidelines, and a list of what a fair benchmark would need. The evidence divides by noise type. Median-type filters scored best on impulse noise, Lee and Kuan filters were the strongest simple options for radar speckle, and learned denoisers scored highest on simulated speckle but need training data that satellite programs rarely have. For additive Gaussian noise the studies disagree. Stripe noise survived almost every general-purpose filter that was tried. Because the studies differ in imagery, noise level and parameter choices, the paper does not rank filters across studies.
Authors
- Mohak Rajendra More (ORCID: https://orcid.org/0009-0007-5468-1082)
- Sujal Virendra Khadatkar
- Atharva Sanjay Dhupkar
Institutions
- Dr. Babasaheb Ambedkar Technological University (IN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24421281
- Primary Topic
- Image and Signal Denoising Methods
- Type
- preprint