Skip to content

HLS-GPT: A Generative Pretrained Transformer (GPT) for Continental-Scale NASA Harmonized Landsat and Sentinel-2 (HLS) Reflectance Reconstruction Across All Bands on Arbitrary Dates

Home / Papers / HLS-GPT: A Generative Pretrained Transformer (GPT) for Continental-Scale NASA Harmonized Landsat and Sentinel-2 (HLS) Reflectance Reconstruction Across All Bands on Arbitrary Dates

HLS-GPT: A Generative Pretrained Transformer (GPT) for Continental-Scale NASA Harmonized Landsat and Sentinel-2 (HLS) Reflectance Reconstruction Across All Bands on Arbitrary Dates

Authors: Junjie Li, Hankui K. Zhang, David P. Roy Date: 2026-06-16 Paper ID: openalex:2606.18115

Summary

HLS-GPT is a large-scale generative pretrained transformer designed to reconstruct Harmonized Landsat and Sentinel-2 (HLS) surface reflectance for any date and pixel location. Utilizing a hierarchical architecture, the model processes 12-month single-pixel time series to capture seasonal and geographic variability across the conterminous United States. By training on a large-scale dataset with a masking strategy, the model achieves highly accurate spectral reconstruction even under conditions of sparse and irregular data availability.

Key Contributions

  • Introduces HLS-GPT, a transformer-based generative model for single-pixel continental-scale surface reflectance reconstruction across all HLS bands.
  • Achieves state-of-the-art performance on HLS time series reconstruction, outperforming conventional methods and the NASA-IBM Prithvi model.
  • Demonstrates robust performance under sparse and irregular observation regimes, maintaining consistent reconstruction accuracy across varying masking levels.

Metadata & Links

url
https://arxiv.org/abs/2606.18115
paper_id
2606.18115
paper_source
openalex
domain
computer-vision
tags
transformertime-seriesremote-sensing
architectures
datasets
concept_slugs
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-06-19T06:47:19Z