HLS-GPT trained model and training data

This collection contains the trained model (.h5) and the training and test data (.npy) for the HLS-GPT annual time series reconstruction for Conterminous United States (CONUS).

More details can refer to the paper:

Li, J., Zhang, H. K., and Roy, D. P. (2025). HLS-GPT: A Generative Pretrained Transformer (GPT) Model for Accurate Harmonized Landsat and Sentinel-2 (HLS) Reflectance Time Series Reconstruction. In review.

1.best_model_GAPS_P0.5_versionv7_27.h5

Trained model weights for HLS-GPT, the structure of the model please refer to Li et al., 2025.

2.AlignedCONUS_v1_6_training_test_data.zip

We collected 11-year (2013-2023) good-quality (not cloud, cloud shadow or ice/snow) HLS reflectance time series from 312,950 pixel locations evenly distributed across 96 HLS tiles (every other tile across CONUS) and randomly split them into training (80%, 250,360 pixels ) and test (20%, 62,590 pixels) locations. In our paper, we only used 9-year data from 2015 since the first Sentinel-2 satellite launch. 

The training or testing input x (i.e., the HLS reflectance) is stored as a 3D matrix with dimensions n × ( (176+176)×11) ×13 in AlignedCONUS_v1_6_multi_year_train.npy and AlignedCONUS_v1_6_multi_year_test.npy, respectively,

The first dimension, n, represents the number of training or test pixel locations.

The second dimension, (176+176)×11, represents the length of the 11-year time series (2013-2023):

The first 176 represents the Landsat data, with a maximum of 176 dates.

The second 176 represents the Sentinel-2 data, with a maximum of 176 dates.

The sequence is as follows:  176 Landsat time points for year 2013, 176 Sentinel-2 time points for year 2013, 176 Landsat time points for year 2014, 176 Sentinel-2 time points for year 2014, …, 176 Landsat time points for year 2023, and 176 Sentinel-2 time points for year 2023.

For each record, the Landsat and Sentinel-2 time series are temporally aligned—identical dates in the same order—which is essential for concatenating sensor-specific Transformer feature vectors. This alignment introduces only minimal computational overhead at inference.

Time series with fewer than 176 observations were padded with -9999.

The third dimension, 13, corresponds to spectral information, including year, day of year (DOY), and normalized reflectance. The reflectance bands are ordered as follows: four visible bands, one near-infrared (NIR) band, two shortwave infrared (SWIR) bands, three red-edge bands (Sentinel-2 only), and one broad NIR band (Sentinel-2 only). For Landsat data, the last four bands are filled with -9999. The mean and std normlization file is included in https://github.com/junjieliwhu/HLSGPT.

In addition, the AlignedCONUS_v1_6_random_selected_one_year_test.npy stores 4,157,97 one year time sereis (4,157,97× (176+176) ×13) used in the paper for the model evaluation that was extracted from testing data AlignedCONUS_v1_6_multi_year_test.npy with random starting dates. This is in case users want to develop more advanced models for comparison purpose. 

3. AlignedCONUS_scale60_all_tiles_v1_6_filtered.parquet

The training and testing files were generated by processing AlignedCONUS_scale60_all_tiles_v1_6_filtered.parquet.

AlignedCONUS_scale60_all_tiles_v1_6_filtered.parquet is derived from our previous work (https://zenodo.org/records/14715402), except that this dataset filters out pixels that are located in the ocean or outside the CONUS. For an introduction to the original data, please refer to https://zenodo.org/records/14715402