Terra-GPT training data, trained model weights, and demo data

Terra-GPT training data and trained model weights 1. Pretraining data CONUS_scale60_all_tiles_v2_1.csv This file contains the original HLS data with day of year, surface reflectance and quality flag. The data was obtained by sampling every 60th 30-meter pixel across 96 systematically distributed tiles covering the CONUS from year 2015 to 2025 (11 years). Each record contains data for a single pixel location in a specific year. Each row includes pixel information ('tile', 'col', 'row', 'year', 'lat', 'lon', 'tile_id', and 'lid') and 366 daily surface reflectance observations organized as follows: L000, L001, ..., L365, and S000, S001, ..., S365. The surface reflectance values and day of year (DOY) are defined by the column names. For example, L000.doy, L000.coastal, L000.blue, and so on. Missing values were filled with -9999. CONUS_scale60_all_tiles_v2_1.meta_with_split_update_dem_nlcd.csv This file stores the metadata for CONUS_scale60_all_tiles_v2_1.csv file, including 'tile', 'col', 'row', 'year' , 'lat', 'lon', 'tile_id', 'lid', 'set', 'dem', 'slope', 'aspect', 'NLCD', 'homo'. The column 'set' indicates whether a pixel is used for training or testing The column 'NLCD' for year 2025 is -9999 The column 'homo' represents the number of neighboring pixels among the eight surrounding pixels that have the same NLCD value. 2. Fine-tuning data 2.1 Live fuel moisture content (LFMC) HLS_LFMC_CONUS_v2_0_with_split.parquet This file contains the original LFMC training and testing data, including LFMC metadata (Site name, Elevation, IGBP Land Cover ID, lat, lon, Contact, Country, etc.) and the corresponding daily HLS surface reflectance. Each record contains data for a single LFMC site in a specific year. The HLS time series is stored as follows: L000.doy, L000.Date, L000.HLS_Tile, L000.HLSL30_020_B01, L000.HLSL30_020_B02, and so on. The band name definitions are as follows:: L30_names={ 'HLSL30_020_B01':'coastal', 'HLSL30_020_B02':'blue', 'HLSL30_020_B03':'green', 'HLSL30_020_B04':'red', 'HLSL30_020_B05':'nir', 'HLSL30_020_B06':'swir1', 'HLSL30_020_B07':'swir2', 'HLSL30_020_B10':'bt1', 'HLSL30_020_B11':'bt2', 'HLSL30_020_Fmask':'qa',}S30_names={ 'HLSS30_020_B01':'coastal', 'HLSS30_020_B02':'blue', 'HLSS30_020_B03':'green', 'HLSS30_020_B04':'red', 'HLSS30_020_B05':'edge1', 'HLSS30_020_B06':'edge2', 'HLSS30_020_B07':'edge3', 'HLSS30_020_B08':'nir8', 'HLSS30_020_B8A':'nirA', 'HLSS30_020_B11':'swir1', 'HLSS30_020_B12':'swir2', 'HLSS30_020_Fmask':'qa',} The last column, 'split', is used to indicate the training/testing split of the sites. 2.2 Soil moisture retrieval CONUS_ISMN_SM_aligned_P366_HLS-GPT_with_split.parquet This file contains the original ISMN soil moisture training and testing data, including soil moisture metadata (Site name, year, lat, lon, etc.) and the corresponding daily HLS surface reflectance. The data format is the same as that of the LFMC data. The last column, 'split', is used to indicate the training/testing split of the sites. 2.3 Crop type mapping The training and testing data were sampled from https://huggingface.co/datasets/ibm-nasa-geospatial/multi-temporal-crop-classification HLS_Cropland_total_train_scale8_aligned_add_dem.parquet The data were obtained by sampling every 8th pixel across all training chips. Columns include metadata (tile,row,col,lat,lon,year,chip_id,dem,slope,aspect) and corresponding daily HLS surface reflectance. The HLS surface reflectance format is the same as that in CONUS_scale60_all_tiles_v2_1.csv. The column 'value' stores the CDL land-cover values. HLS_Cropland_total_test_scale4_aligned_add_dem.parquet The data were obtained by sampling every 4th pixel across all validation chips. The data format is the same as that of the training data. 2.4 Crop damage The training and testing crop damage events are stored in separate folders. The damaged and non-damaged fields for different years are also stored in separate folders. Each file stores pixel metadata, including label, swathDate, firstDate, bestDate, pixel_id, lat, and lon, as well as the corresponding daily HLS surface reflectance. The HLS time series is stored as follows: L001_DOY, L001_B01, L001_B02, L001_B03, L001_B04, and so on. 3. Trained model weights We provide trained model weights for different tasks and different input. TASK INPUT BANDS MODEL NAME GAP_FILL DEFAULT best_model_hls2026.layer4.METHOD2.BATCH1024.LR0.0005.EPOCH40.L20.1.GAPS0.5.v7_32.h5 GAP_FILL THERMAL best_model_hls2026.layer4.METHOD2.BATCH1024.LR0.0005.EPOCH40.L20.1.GAPS0.5.v7_33.h5 GAP_FILL THERMAL_DEM best_model_hls2026.layer4.METHOD2.BATCH256.LR0.0001.EPOCH40.L20.1.GAPS0.5use_bt_dem.v7_35 FUEL_MOISTURE THERMAL_DEM MC_v6_4.SMmodel.B0004.r0.00001.e10.L0.10000.i5.h5 SOIL_MOISTURE THERMAL_DEM v4_5.SMmodel.B0256.r0.00001.e30.L0.10000.U064.i4.h5 CROP_MAPPING THERMAL_DEM v3_9.layer4.METHOD2.BATCH128.LR0.0001.EPOCH20.L20.1.FT1.use_bt_dem.i0.model.h5 CROP_DAMAGE THERMAL v5_0.layer4.METHOD2.BATCH2048.LR1e-06.EPOCH30.L20.1.FT1.i0.model 1.h5 The input band configurations are defined as follows: DEFAULT: This configuration does not use the L30 thermal bands or auxiliary topographic variables, including DEM, slope, and aspect. THERMAL: This configuration uses the L30 thermal bands but does not use auxiliary topographic variables, including DEM, slope, and aspect. THERMAL_DEM: This configuration uses both the L30 thermal bands and auxiliary topographic variables, including DEM, slope, and aspect. 4. Demo data We provide a small demo dataset for a quick start. HLS_128_patches: Contains three 128 × 128 HLS patches from three HLS tiles (10TDM, 14TMK, and 15TVH). Each patch includes a one-year HLS time series. DEM_128_patches: Contains the corresponding DEM files for the three HLS patches. For more information about the dem tiles, please refer to https://zenodo.org/records/21117230. Expected output: Example outputs generated using the Terra-GPT demo data and demo scripts.