Neural Operator-Based Continuous Tensor Representation for Thick Cloud Removal in Multiresolution Remote Sensing Images

Abstract

The availability of multitemporal remote sensing images (MTRSIs) provides new opportunities for thick cloud removal. An MTRSI acquired from a sensor such as Sentinel-2, Landsat-8, or MODIS usually consists of spectral bands with different spatial resolutions. However, existing methods are mainly designed for single-resolution data and cannot directly handle MTRSIs with different spatial resolutions. To address this challenge, we propose a neural operator-based continuous tensor representation, termed NOCTR, for thick cloud removal in MTRSIs with different spatial resolutions. By leveraging the inherent capability of neural operators for powerful continuous modeling, NOCTR can genuinely explore the intrinsic spatial-spectral-temporal structure of MTRSIs with different spatial resolutions. In particular, NOCTR factorizes MTRSIs with different spatial resolutions into spatial abundance maps and spectral signatures faithfully generated by tailored spatial and spectral-temporal neural operators. Extensive experiments on simulated and real datasets demonstrate that NOCTR achieves superior performance on multiresolution MTRSIs, particularly in terms of spatial and spectral fidelity.