Surface reflectance products from optical satellites often show systematic inconsistencies across acquisition dates because atmospheric correction is performed independently for each image. These residual biases distort the apparent temporal behavior of invariant ground targets, reducing the reliability of multitemporal reflectance analysis.
The authors propose DeepMGCCA, an unsupervised deep multiview pseudoinvariant feature learning framework that integrates Deep Generalized Canonical Correlation Analysis with Multivariate Alteration Detection. Instead of date-specific linear projections, it uses a shared convolutional encoder to learn deep invariant features directly from pixel-weighted canonical variates via an iterative reweighting strategy. A reconstruction regularizer preserves surface reflectance information and prevents degenerate collapse that would suppress genuine temporal variability.
Experiments were conducted on four annual Landsat-8/9 and Sentinel-2 acquisitions across four regions of interest. Among four harmonization transforms derived from the deep invariant features, a monotone piecewise-linear transform based on feature quantiles consistently achieved the largest bias reduction, lowering between-date spectral angle by up to 2.54 and RMSE by 0.015, while raising PSNR by up to 4.04 dB.
The results highlight the potential of deep invariant feature learning for removing surface reflectance biases, while also underscoring the importance of preserving physically meaningful temporal and spectral information during nonlinear representation learning. This is a single arXiv preprint; no independent replication or comparison with other recent methods is reported in the source.