Denoising VIIRS and Sentinel-2 MSI ocean color imagery for improved floating algae monitoring using noise-simulation-aided deep learning

Abstract

The floating algae index (FAI) images derived from Visible Infrared Imaging Radiometer Suite (VIIRS) and Sentinel-2 Multispectral Instrument (MSI) have been widely used to monitor open ocean and coastal floating algal blooms, but they often suffer from complex and variable noise with different strengths, orientations, and distributions. While deep learning methods are effective at reducing noise, adapting them to sensor-specific noise variations remains challenging. Such adaptation typically requires large-volume and high-quality training data to adjust models across different satellite sensors. Here, we propose a two-step denoising process: 1) simulating noise using spatial frequency domain information to generate customized representative training data from limited samples, and 2) training the state-of-the-art Multi-scale Image Restoration Network (MIRNet), which integrates multi-scale residual learning and attention mechanisms, for optimal performance. The method was tested on medium-resolution VIIRS FAI data, degraded by stripe noise, and high-resolution Sentinel-2 MSI FAI data affected by glitter noise. By applying noise simulation, 4,320 training data were generated from 91 real VIIRS samples, and 11,200 training data from 58 real MSI samples. The optimized MIRNet denoising model effectively reduced various types of noise while preserving the spatial details of ocean surface algae features. The average peak signal-to-noise ratio (PSNR) and Structural Similarity Index (SSIM) of denoised VIIRS and MSI FAI images significantly improved over the original noisy images. Noise intensity on FAI and biomass values were estimated for VIIRS and MSI images, showing that noise on MSI FAI led to ∼10 % biases in biomass estimates, whereas the effect on VIIRS biomass estimates was negligible. The MIRNet was also tested on VIIRS Color Index images and improved the image quality even without retraining. Overall, the proposed noise-simulation-aided deep learning method can effectively enhance ocean color image quality, enabling more accurate ocean color remote sensing.