Deep Learning for Sea Surface Temperature Reconstruction under Cloud Occlusion

TL;DR

This work tackles the problem of reconstructing sea surface temperature (SST) in cloud-covered regions by leveraging deep learning, specifically U-Net and Vision Transformer architectures, trained on MODIS-Aqua nighttime L3 SST data at 4 km resolution over the Italian seas. A cloud-occlusion generator enables training and evaluation in the absence of ground truth, and the best-performing model (U-Net64 with four past days as input) achieves substantially lower RMSE than traditional gap-filling methods, notably outperforming the DINCAE approach by about 20% RMSE in a Northern Adriatic test. Subtracting an unbiased seasonal climatology and learning SST residuals further enhances training stability and predictive accuracy. The method demonstrates strong potential for operational SST reconstruction in cloud-affected imagery, with robust performance when applied to Copernicus L3S inputs and promising avenues for extension to the full Mediterranean and integration of additional microwave data.

Abstract

Sea Surface Temperature (SST) reconstructions from satellite images affected by cloud gaps have been extensively documented in the past three decades. Here we describe several Machine Learning models to fill the cloud-occluded areas starting from MODIS Aqua nighttime L3 images. To tackle this challenge, we employed a type of Convolutional Neural Network model (U-net) to reconstruct cloud-covered portions of satellite imagery while preserving the integrity of observed values in cloud-free areas. We demonstrate the outstanding precision of U-net with respect to available products done using OI interpolation algorithms. Our best-performing architecture show 50% lower root mean square errors over established gap-filling methods.

Figures (11)

• Figure 1: The white polygon describes the region of our investigation, with latitude between 35.33$^\circ$ and 46.0$^\circ$E and longitude between 7.92$^\circ$ and 18.58$^\circ$N.
• Figure 2: Histogram relative to the distribution of MODIS-AQUA nightly temperatures, cumulative overall spatial positions and all years.
• Figure 3: Large Spatial Gradients ($>1.5^\circ$C) relative to different days of the year. Units are $^\circ$C
• Figure 4: On the left, the nighttime SST data from a sample day, in this case 10/05/2022; in the middle, the climatology for May 10; on the right, the climatology adjusted (shifted) to the mean of the specific day.
• Figure 5: Basic U-Net. In our terminology, this U-Net has a [32,64,128,256] structure, meaning that it is composed of three downsampling blocks progressively halving the spatial dimension, and increasing the channel dimension to 64, 128 and 256. The initial spatial dimension is 256x256. The initial number of channels is 7, corresponding to three input days with the associated masks and the land-sea mask.