Temperature Downscaling — Colorado Front Range

Dataset Description

This dataset supports machine learning models for statistical downscaling of land surface temperature (LST). The goal is to downscale coarse-resolution MODIS LST (1 km) using high-resolution predictors (Sentinel-2 vegetation/water indices at 10 m, NLCD land cover at 30 m, and ASTER elevation at 30 m) over the Colorado Front Range during 2022–2023.

Summary

• Domain: Colorado Front Range (~39°N–40°N, 105°W–106°W)
• Period: 2022–2023
• Target variable: MODIS MOD11A1 Land Surface Temperature (1 km, daily)
• Predictors: Sentinel-2 NDVI/NDWI (10 m), NLCD land cover (30 m), ASTER elevation (30 m)

Dataset Preview

Each row in the viewer is one MODIS date × tile sample, with matched data from all four sources:

Column	Description
modis_browse	MODIS LST daytime browse image (available for Jan–Jun 2022)
sentinel2_b04	Nearest Sentinel-2 Red band (10 m) for the matched tile
aster	ASTER DEM elevation tile (30 m, static)
lulc	NLCD land cover for that year (30 m)
split	train, val, or test
date	MODIS acquisition date
modis_tile	MODIS sinusoidal tile (h09v04, h09v05, h10v04)
modis_hdf	Path to the MODIS HDF-EOS science file
sentinel2_date	Date of the matched Sentinel-2 acquisition
sentinel2_tile	Sentinel-2 MGRS tile
sentinel2_b03 / sentinel2_b08	Green and NIR band paths

Note: The MODIS HDF science files (LST data) are the primary target but are not previewable as images. Use the loading code below to access them programmatically.

Dataset Structure

akhot2/downscaling/
├── metadata.csv          # Per-sample catalog (2,151 rows) with all sources matched
├── splits.csv            # Train/val/test assignments for MODIS HDF files
├── ASTER/                # ASTER GDEM v3 — 30 m elevation (static)
│   ├── *_dem.tif         # Elevation tiles (4 tiles: N39–N40, W105–W106)
│   └── *_num.tif         # Observation count / quality
├── bounding box/         # Study area definition
│   └── colorado_front_range_domain.shp (+ .shx, .dbf, .prj, .cpg)
├── LULC/                 # NLCD Land Cover — 30 m, annual
│   ├── Annual_NLCD_LndCov_*.tiff   # 2021, 2022, 2023, 2024
│   └── nlcd_legend.csv
├── MODIS/                # MOD11A1 v061 — 1 km, daily LST
│   ├── MOD11A1.A*.hdf    # HDF-EOS science files (2,151 files)
│   └── BROWSE.*.jpg      # Browse preview images (828 files)
└── Sentinel2/            # Sentinel-2 L2A — 10 m, ~bimonthly
    └── S2*_{B03,B04,B08}.tif   # Green, Red, NIR bands (168 files)

Data Sources

Source	Resolution	Temporal	Format	Files	Role
MODIS MOD11A1 v061	1 km	Daily	HDF-EOS	2,151	LST target (Day & Night)
Sentinel-2 L2A	10 m	~Bimonthly	GeoTIFF	168	NDVI, NDWI predictors (B03, B04, B08)
NLCD	30 m	Annual	GeoTIFF	4 years	Land cover class predictor
ASTER GDEM v3	30 m	Static	GeoTIFF	4 tiles	Elevation predictor

Spatial Coverage & Tile Mapping

MODIS Tile	Sentinel-2 Tiles	ASTER Tile
h09v04	T13TDE, T13TEE	N40W106
h09v05	T13SDD, T13SED	N39W106
h10v04	T13TDE, T13TEE	N40W105

Train / Validation / Test Splits

Splits use a temporal block split to avoid data leakage:

Split	Period	Samples	%
train	2022 Jan–Sep (DOY 1–273)	819	38%
val	2022 Oct–Dec (DOY 274–365)	237	11%
test	2023 (full year)	1,095	51%

Split rationale

• Temporal separation ensures no data leakage between splits.
• Seasonal shift between train (warm months) and val (cold months) tests robustness to seasonal variation.
• Year boundary for test provides a clean, interpretable evaluation of temporal generalization.

Spatial folds

Each sample's modis_tile column (h09v04=0, h09v05=1, h10v04=2) can be used as a spatial fold for leave-one-tile-out cross-validation, enabling evaluation of spatial generalization independently of the temporal split.

MODIS HDF Variables

Each MOD11A1 HDF file contains these Science Data Sets (SDS):

SDS Name	Description	Scale Factor	Units
LST_Day_1km	Daytime LST	0.02	Kelvin
LST_Night_1km	Nighttime LST	0.02	Kelvin
QC_Day	Daytime quality control	—	Bitfield
QC_Night	Nighttime quality control	—	Bitfield
Emis_31	Band 31 emissivity	0.002 (+ 0.49)	—
Emis_32	Band 32 emissivity	0.002 (+ 0.49)	—

To convert LST to Celsius: LST_C = pixel_value × 0.02 − 273.15 (for valid pixels in range 7500–65535).

Uses

Intended uses

• Training and evaluating statistical or deep-learning downscaling models for land surface temperature.
• Studying relationships between LST and land surface characteristics (vegetation, land cover, elevation).
• Benchmarking spatial super-resolution methods on real remote sensing data.

Out-of-scope uses

• This dataset covers a single region (Colorado Front Range) and should not be assumed to generalize globally without further validation.
• The MODIS LST product has known limitations over complex terrain, clouds, and snow-covered surfaces.

Loading the Data

from huggingface_hub import snapshot_download
import pandas as pd

# Download all files
snapshot_download(repo_id="akhot2/downscaling", repo_type="dataset", local_dir="data")

# Load the per-sample metadata
metadata = pd.read_csv("data/metadata.csv")
train = metadata[metadata["split"] == "train"]
val   = metadata[metadata["split"] == "val"]
test  = metadata[metadata["split"] == "test"]

# Example: iterate training samples
for _, row in train.iterrows():
    modis_path = f"data/{row['modis_hdf']}"
    sen2_path  = f"data/{row['sentinel2_b04']}"
    aster_path = f"data/{row['aster']}"
    lulc_path  = f"data/{row['lulc']}"
    # Load with rasterio / pyhdf ...

Code

Notebooks and model code: fresleven/downscaling

License

MIT