Salma El Fellah

This study leverages advancements in deep learning to enhance the analysis of satellite image time series in marine geoscience, focusing on the prediction of sea surface temperature (SST) and detection of coastal upwelling. By employing Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), including Long Short-Term Memory (LSTM) networks, we introduce a novel approach utilizing Convolutional LSTM (ConvLSTM) and 3D Unet-LSTM models. These techniques provide a nuanced analysis and understanding of complex oceanographic phenomena, specifically coastal upwelling, which significantly impacts marine ecosystems and climate. The adoption of these sophisticated deep learning models has led to a notable improvement in predicting SST fields, achieving a reduction in Root Mean Square Error (RMSE) to 0.038 and an increase in the Correlation Coefficient (CC) to 0.95. This enhancement over the baseline ConvLSTM model, which had an RMSE of 0.045 and a CC of 0.92, underscores our models’ capability to accurately capture the dynamic and intricate nature of coastal upwelling. The results offer promising directions for future research in marine geoscience and remote sensing applications, highlighting the potential of deep learning techniques in interpreting intricate patterns in satellite-derived data and improving predictions in environmental sciences.

Landsat 7 Enhanced Thematic Mapper Plus (ETM+) satellite imagery presents an important data source for many applications related to remote sensing. However, the scan line corrector (SLC) failure has seriously limited the scientific applications of ETM+ data since SLC failed permanently on May 31, 2003, resulting in about 22% of the image data missing in each scene. In this letter, we propose to apply a new method to fill the missing information (gap) in the satellite image without any spatial or spectral constraint applied on the neighborhood of missing pixels. The method can also simultaneously remove various types of noise, including Gaussian noise and impulse noise, affected in the acquisition process. This proposed concept is formulated as a consecutive multilevel Otsu and Two-Threshold Binary Decomposition to detect the boundary of the damaged region (gaps) and then reconstruct the edge of the damaged region based on boundary restoration. In the final step, we applied the well-known matrix completion as a low-rank approximation problem guided by boundary reconstructed. We achieve better approximation of the rank by minimizing the truncated nuclear norm with accelerated proximal gradient line, a method for convex optimization problems. The results show the capability of our approach to predict the missing values accurately in terms of quality, time, and without the need of outside information. This letter also shows encouraging results on both synthetic and real visual data sets.