3D-DINEOF: Extension of Decomposition Dimensions of Data Interpolating Empirical Orthogonal Functions

Optical and infrared satellite remote sensing data are often severely affected by cloud cover, sun glint, swath limitations, and other environmental or instrumental factors, leading to large gaps that hinder climate analysis and ocean monitoring. To mitigate this challenge, a variety of gap-filling techniques have been developed. Among them, the Data INterpolating Empirical Orthogonal Functions (DINEOF) method is one of the most widely adopted. Conventional DINEOF methods struggle with spatiotemporal inconsistencies and often fail under high missing data rates. Here, we present 3D-DINEOF, an enhanced approach that applies iterative empirical orthogonal function decomposition sequentially along longitude, latitude, and time to exploit multi-dimensional correlations. This strategy preserves intrinsic spatiotemporal structures and substantially improves gap-filling performance. Applied to a two-year sea surface temperature dataset, 3D-DINEOF significantly outperformed both conventional DINEOF and Data Interpolating Convolutional AutoEncoder (DINCAE), accurately recovering fields with over 95% missing data. Beyond restoring large-scale patterns, it also captures fine-scale variability in dynamic regions, demonstrating its robustness for reconstructing complex oceanographic datasets under extreme data loss.