Deep Feature Gaussian Processes for Single-Scene Aerosol Optical Depth Reconstruction

Shengjie Liu (University of Southern California), Lu Zhang (University of Southern California)

Data and code for the published paper in IEEE Geoscience and Remote Sensing Letters: https://doi.org/10.1109/LGRS.2024.3398689

GitHub repo at https://github.com/skrisliu/dfgp

Key Packages

	pytorch==2.0.1
	gpytorch==1.11
  

Tested on Python 3.9, Ubuntu 18.04.6 LTS, with 1080 Ti 11GB GPU. Running on CPU is possible but significantly slow.

This repo includes the MODIS-LA data, with trained CNN network and deep features. For the EMIT-BJ data, download from Google Drive.

Demo on MODIS-LA

data/modis/im.npy         # features, 240*300*13, the last two dimensions are xy coordinates
data/modis/aod.npy        # label, 240*300
data/modis/trainmask.npy  # train mask, 240*300
data/modis/testmask.npy   # test mask, 240*300
data/modis/fea64.npy      # deep features, 240*300*64
data/modis/cnn.pt         # trained CNN model
  

Data required to run the MODIS-LA demo are included in this repo.

DFGP

Python demo12_modis_dfgp.py
  

DFGPs

Python demo12_modis_dfgps.py
  

Baseline Methods

CNN

Python demo11_modis_cnn.py
  

Requires networks.py and rscls.py to clip the image and load the network.

@article{liu2024deep,
  title={Deep Feature Gaussian Processes for Single-Scene Aerosol Optical Depth Reconstruction},
  author={Liu, Shengjie and Zhang, Lu},
  journal={IEEE Geoscience and Remote Sensing Letters},
  year={2024},
  publisher={IEEE}
}