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

2026-06-16 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

HLS-GPT is a large-scale generative pretrained Transformer model designed for reconstructing NASA Harmonized Landsat Sentinel-2 (HLS) 30 m surface reflectance across all bands, any date, and any pixel location. Addressing limitations of prior deep learning methods, HLS-GPT employs a hierarchical Transformer architecture to manage Landsat and Sentinel-2's differing spectral band configurations, operating on single-pixel 12-month time series. The model was trained using nine years of HLS data from over 0.25 million training pixels across the conterminous United States, employing a random cropping and masking strategy that reconstructs 50% masked reflectance values. Evaluation on over 62,000 independent test pixels demonstrated robust reconstruction, achieving RMSE below 0.026 for all HLS spectral bands, with relative RMSE below 35% for visible bands and below 13% for others. Sensitivity analyses showed modest degradation when 10% to 50% of observations were masked, maintaining all-band RMSE below 0.028. HLS-GPT also outperformed two conventional methods and the NASA-IBM Prithvi model in image reconstruction over nine 109 by 109 km CONUS HLS tiles.

Key takeaway

For remote sensing scientists and environmental modelers requiring complete, high-quality satellite time series, HLS-GPT offers a robust solution for reconstructing missing Landsat and Sentinel-2 reflectance data. You can utilize this model to fill gaps in sparse or irregular observations, improving the accuracy of downstream analyses like land cover classification or phenology tracking. Consider integrating HLS-GPT outputs to enhance the reliability of your continental-scale environmental monitoring and change detection applications.

Key insights

HLS-GPT reconstructs continental-scale Landsat and Sentinel-2 reflectance time series for all bands and dates using a hierarchical Transformer.

Principles

• Hierarchical Transformers can integrate multi-sensor spectral data.
• Masked modeling improves robustness for sparse satellite time series.
• Training on diverse geographic and seasonal data enhances generalization.

Method

HLS-GPT trains on single-pixel 12-month HLS time series, using a random cropping and 50% masking strategy to reconstruct reflectance values from remaining observations.

In practice

• Reconstruct missing satellite data for complex crop phenology.
• Generate complete HLS time series for environmental monitoring.
• Improve data quality for land cover change detection.

Topics

• HLS-GPT
• Generative Pretrained Transformer
• Landsat Sentinel-2
• Surface Reflectance
• Remote Sensing
• Time Series Reconstruction

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Landsat-Sentinel-2 Algal Bloom Mapping Using Vision Transformers: Model Description, Implementation, and Examples — Vision Transformers enable robust, high-resolution algal bloom mapping from harmonized Landsat-Sentinel-2 imagery, outperforming traditional methods under challenging conditions.
• GGPT: Geometry Grounded Point Transformer — GGPT refines dense 3D reconstructions by fusing feed-forward predictions with accurate sparse geometric guidance from an improved SfM pipeline.
• AnyBand-Diff: A Unified Remote Sensing Image Generation and Band Repair Framework with Spectral Priors — AnyBand-Diff generates physically consistent remote sensing images by integrating spectral priors and multi-scale physical constraints into a diffusion model.
• LESSViT: Robust Hyperspectral Representation Learning under Spectral Configuration Shift — LESSViT enhances hyperspectral image model generalization across sensors via efficient low-rank spatial-spectral attention.
• ATT-CR: Adaptive Triangular Transformer for Cloud Removal — ATT-CR uses adaptive triangular attention and feature gating to efficiently remove clouds, reducing complexity and interference.
• ImplicitTerrainV2: Wavelet-Guided Spatially Adaptive Neural Terrain Representation — ImplicitTerrainV2 uses wavelet-guided adaptivity and derivative-aware supervision for efficient, high-fidelity neural terrain representation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.