Whittaker-Henderson smoother for long satellite image time series interpolation

Researchers at CESBIO have reimagined a classic signal processing tool for the AI era, transforming the Whittaker-Henderson smoother into a differentiable neural layer. This innovation tackles two persistent limitations in satellite image time series analysis: the need for manual per-pixel parameter tuning and the assumption of uniform noise across time. By embedding the smoother within a neural network framework, the system can now automatically infer optimal smoothing parameters while adapting to heteroscedastic noise—where noise levels vary throughout the temporal dimension. This allows the degree of smoothing to adjust locally along the time series, potentially better handling phenomena like seasonal changes or temporary cloud cover.

The technical implementation leverages a sparse, memory-efficient architecture using Cholesky factorization to exploit the symmetric banded structure of the underlying linear system. Benchmarks demonstrate this approach substantially outperforms standard dense linear solvers in both speed and memory consumption on GPU hardware. The system was validated on satellite image time series (SITS) covering French metropolitan territory from 2016 to 2024, confirming feasibility for large-scale processing. However, researchers noted reconstruction differences with traditional homoscedastic baselines remained limited, suggesting the transformer architecture used for parameter estimation may lack the temporal acuity needed to capture abrupt noise variations like single-day cloud contamination—indicating an area for future improvement in temporal modeling precision.