---
license: cc-by-4.0
language:
- en
size_categories:
- n<1K
---
# Dataset Summary:
This dataset consists of Harmonized Landsat and Sentinel-2 multispectral reflectance imagery and MERRA-2 observations centered around eddy covariance flux towers and the corresponding Gross Primary Productivity (GPP) data at the towers. Its purpose is to serve as a finetuning dataset for geospatial foundation models for the task of regressing GPP flux observations from HLS and MERRA-2 data.

# Dataset Structure:
The dataset consists of: <br/>
(1) HLS 6-band Tiff files of dimension 50x50x6, with the center of the chip colocated with flux tower locations, <br />
(2) 10-dimensional vector of MERRA-2 variables for each chip (1x1x10) recording temperature, soil moisture, heat flux, radiation, precipitation at the flux towers, <br />
(3) Daily GPP data derived from the eddy covariance measurements using the night-time partitioning approach at 37 flux tower sites distributed globally spanning 2018 to 2021. There are a total of 975 instances. MERRA-2 data and GPP flux observations are recorded as csv files, with a row corresponding to each HLS chip.

## HLS Band Order:
1, Blue, B02 <br />
2, Green, B03 <br />
3, Red, B04 <br />
4, NIR, B8A <br />
5, SW 1, B11 <br />
6, SW 2, B12


# MERRA-2 observations:
1. [M2T1NXSLV] T2MIN,
2. [M2T1NXSLV] T2MAX,
3. [M2T1NXSLV] T2MEAN,
4. [M2T1NXSLV] TSMDEWMEAN,
5. [M2T1NXLND] GWETROOT,
6. [M2T1NXLND] LHLAND,
7. [M2T1NXLND] SHLAND,
8. [M2T1NXLND] SWLAND,
9. [M2T1NXLND] PARDFLAND,
10. [M2T1NXLND] PRECTOTLAND

# Data Splits:
The dataset consists of 975 chips and we split the dataset based on years to create train test splits. Given the relatively small size of the dataset, we use a leave-one-year-out-cross-validation approach to train and evaluate. The number of observations vary across years.
In this repo, we have used three years for training and one year as test.

# Dataset Creation:
The flux observation sites recorded in our csv file guide our HLS and MERRA-2 data preparation process. HLS data is preselected with a 25% maximum cloud threshold and 75% minimum spatial threshold. We note the lat/long of the flux stations and extract 50x50 size HLS chips around each station and retain scenes with less than 2% snow cover and less than 5% cloud cover. HLS reflectance is transformed by the scaling factor. We record the daily hourly mean of each abovementioned MERRA-2 data at each of the flux location sites. We then apply a quality check on GPP flux station readings and retain those that meet 60% quality threshold to create the final dataset by selecting corresponding input HLS and MERRA-2 data for these flux stations.

# Source data:
1. HLS Imagery is from V2.0 of HLS. A full description and access to HLS may be found at https://hls.gsfc.nasa.gov/ <br/>
2. MERRA-2 data is from V5.12.4. Full description and to the MERRA-2 datasets can be found at https://gmao.gsfc.nasa.gov/reanalysis/MERRA-2/data_access/ <br/>
3. Eddy covariance data were obtained from the AmeriFlux FLUXNET and ICOS Warm Winter 2020 datasets. (List of eddy covariance flux sites used in the study can be found in the csv file).

# Citation