Self-Supervised Representation Learning on Remote Sensing Pixel Time Series with Patch-Based Masking

We build on the patch-segmentation and channel-independence concepts introduced by PatchTST (Nie et al., 2023), adapting them to remote-sensing pixel time series for self-supervised representation learning. In our approach, each spectralband/index (e.g., Red, NDWI) is treated as a univariate sequence, split into ”patch tokens”, then processed by a 600k-parameter transformer encoder. We randomly mask a substantial fraction of the patch tokens and train the transfomer model to reconstruct the missing segments, capturing both short sub-seasonal dynamics within each patch and multi-seasonal context across the entire time series. Empirical results show that the learned representations transfer effectively to downstream tasks such as land-cover classification and anomaly detection.We also discuss future directions, including foundation-scale training and integration of additional other sensor modalities.