DELAG

S. Liu, S. Wang and L. Zhang, 2025. "Daily Land Surface Temperature Reconstruction in Landsat Cross-Track Areas Using Deep Ensemble Learning With Uncertainty Quantification," in IEEE Transactions on Geoscience and Remote Sensing, doi: 10.1109/TGRS.2025.3643985.

Contact email: [email protected]

This is a GitHub repo at github.com/skrisliu/delag

Code and data update ongoing: last updated 20251216

Minimum Run [NYC]: Task #1 Clear-Sky With Real-World Cloud Patterns

git clone https://github.com/skrisliu/delag.git
cd delag
curl -L https://github.com/skrisliu/delag/releases/latest/download/nyc.7z -o nyc.7z
7z x ./nyc.7z -o./nyc
python nyc_task1_01.py # Run eATC, get 200 predictions.
python nyc_task1_02.py # Based on 200 predictions, Run GP. 

Environment & Package Setup

Run these commands in sequence (terminal/Anaconda Prompt):

conda create --name py39c python=3.9    
conda activate py39c 
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117   
pip install gpytorch==1.13  
pip install numpy==1.26.4   

Preprocessing

(Folders are auto-created for each step)

1. Unzip all Landsat data for the target year. [auto-create unzip folder]
2. (Optional, for ldn & hkg) Reproject data to the same UTM zone.
3. Clip data to target UTM x/y bounds. [auto-create clip folder]
4. Reorder data files as [YYYYMMDD]_[LC08/LC09] (date_sensor). [auto-create order folder]
5. Build datacubes: lsts, clearmasks, meanbands, era5lst. [auto-create datacube folder]

Ready-to-Use Training/Testing Dataset

Datasets and Predictions

NYC

1. NYC Dataset: nyc.7z
2. NYC Predictions LST 2023: nyc_pred_2023.zip

London (LDN)

1. LDN Dataset: ldn.7z
2. LDN Predictions LST 2023: ldn_pred_2023.zip

Hong Kong (HKG)

1. HKG Dataset: hkg.7z
2. HKG Predictions LST 2023: hkg_pred_2023.zip

Get Datasets [NYC as Example]

1. The NYC dataset is hosted via GitHub Releases: nyc.7z
2. After downloading, unzip it to the ./nyc folder to run the scripts:

curl -L https://github.com/skrisliu/delag/releases/latest/download/nyc.7z -o nyc.7z
7z x ./nyc.7z -o./nyc

Experiments: three settings

We evaluate DELAG's performance across three experimental setups:

Task #1: Clear-Sky Situations With Real-World Cloud Patterns

For a day with observed data, use another day's real-world cloud patterns

NYC: Predict index= 91, cloud index=155
LDN: Predict index=249, cloud index=169
HKG: Predict index=321, cloud index=177

NYC Task #1

python nyc_task1_01.py # Run eATC, get 200 predictions. 
python nyc_task1_02.py # Based on 200 predictions, Run GP. 

Task #2: Heavily-Cloudy Situations

Reconstruction under heavily cloudy situations.

Task #3: Indirect Validation via Estimating Near-Surface Air Temperature

Indirectly evaluate performance through LST data's capability to estimate near-surface air temperature.

1. The nyc.7z file should be downloaded and unzip to ./nyc
2. Download the full prediction data [~1.6GB]
3. Unzip to current folder
4. Run nyc_task3_plot_airtemp.py to estimate air temperature from the predicted surface temperature.

# curl -L https://github.com/skrisliu/delag/releases/latest/download/nyc.7z -o nyc.7z
# 7z x ./nyc.7z -o./nyc
curl -L https://github.com/skrisliu/delag/releases/latest/download/nyc_pred_2023.zip -o nyc_pred_2023.zip
7z x nyc_pred_2023.zip
python nyc_task3_plot_airtemp.py

Project Task Status

Code cleaning in progress

✅ Dependencies
✅ Preprocessing Code
✅ Task #1: Clear-Sky Situations With Real-World Cloud Patterns
✅ Task #1: Data Ready
❌ Task #2: Under Heavily-Cloudy Situations
❌ Task #2: Data Ready
✅ Task #3: Indirect Validation via Near-Surface Air Temperature
✅ Task #3: Data Ready

The Paper

Liu, Shengjie, Siqin Wang, Lu Zhang. Daily Land Surface Temperature Reconstruction in Landsat Cross-Track Areas Using Deep Ensemble Learning With Uncertainty. Quantification. IEEE Transactions on Geoscience and Remote Sensing, 2025. https://doi.org/10.1109/TGRS.2025.3643985

@article{liu2025daily,
    author = {Liu, Shengjie and Wang, Siqin and Zhang, Lu},
    title = {Daily Land Surface Temperature Reconstruction in Landsat Cross-Track Areas Using Deep Ensemble Learning With Uncertainty Quantification},
    journal = {IEEE Transactions on Geoscience and Remote Sensing},
    year = {2025}, 
    doi = {10.1109/TGRS.2025.3643985}
}