Meeting Documents

Structured Noise in AMSR-E SST Fields and Its Impact on Their Deconvolution

Abstract

AMSR-E fields, like those of many other satellite-borne microwave instruments, are grossly oversampled. In the case of AMSR-E sea surface temperature, SST, fields, the footprint of the fields is order 45x65 km sampled on a 10x10 km grid—every 10x10 km region of the sea surface impacts the SST of approximately 35 pixels. Given the significant oversampling, it should be possible, with a small amount of additional information, to deconvolve the existing fields to obtain a true 10x10 km SST product. Given that the MODIS instrument flies on the same satellite providing a 1x1 km field in cloud free regions, there are often enough cloud-free MODIS pixels—the additional information needed—to invert the AMSR-E fields. Unfortunately, even a minuscule amount of noise—white noise in the milli-Kelvin range—in the AMSR-E fields results in inversions with significantly more noise than in the underlying geophysical fields. To address this we have used a U-Net machine learning model to perform the deconvolution. We trained the model with output from the LLC-4320 run of the ECCO ocean general circulation model and we tested the trained model with LLC-4320 simulated fields as well. The results were very encouraging even when we added significant white noise to the simulated fields. However, when applied to real AMSR-E fields the results were very poor. The cause of the poor performance was traced to the fact that the noise in the AMSR-E fields is not white, the fields are quite 'lumpy' in the along-scan direction and the model trained with simulated data including white noise 'learned' to augment the gradients found in the AMSR-E field. In this presentation, we discuss the characteristics of the noise in the AMSR-E fields, the addition of noise with the same spectral characteristics as that observed, retraining of the U-Net model and the results of the application of the trained model to actual AMSR-E data.