Satellite Image Denoising Using Deep Learning: Autoencoder Decoder Application

Through deep learning Autoencoder Decoders, it is possible to clean noisy or damaged image data received from satellites. Two models with a PSNR of 25.6 dB and 25.54 dB were generated using a dataset consisting of 610 weather satellite images retrieved from the National Oceanic and Atmospheric Administration (NOAA) Weather (WX) satellites. These images were acquired via the ground station at the University of Perpetual Help System DALTA - Las Pinas campus. Initially, only a few alterations were made to the code in order to work with later Keras versions. However, this led to crashes during training as well as noticeable darkening in the predictions. The original code also used accuracy as the training metric, instead of PSNR. The rewrite rebuilt the model using updated Keras 3 API as well as introduced a PSNR metric. Doing so eliminated the darkening and crashes entirely.