Reconstructing cloud-contaminated satellite images with SAR-Optical fusion using spatio-temporal partitioning and multiple linear regression

Python scripts fusion_GEE.py achieves this objective. All outputs will be exported to the Google Drive linked to the GEE account.

Description

• fusion_GEE.py performs GEE-based optical-SAR fusing. It retrieves optical and SAR satellite images from GEE for an user-specified period and AOI, predicts cloudy optical images with corresponding SAR images which are not affected by cloud, and outputs infilled optical images.
• Parameters.json in config folder contains the configuration for all environmental and model parameters.
• The shape or json files in AOI folder contains all case study sites as below:

Name	Landcover and Natural	Area (km^2)	Site Central Longitude (°E)	Site Central Latitude (°S)
North NT	Conservation environments and irrigated perennial horticulture	24	131.2	-12.59
Central West NSW	Grazing native vegetation and dryland cropping	16	148.77	-31.32
Central Tas	Grazing modified pasture and irrigated cropping	7	147.51	-41.92
South West Qld	Irrigated cropping	10	145.73	-27.94
East Gippsland Vic	Production native forests and grazing modified pastures affected by bush fire	5	148.43	-37.64
North Qld	Sugar cane with surrounding rainforest	19	145.8	-17.05
Wheatbelt WA	Dryland cropping, wheat, oats and pasture	19	117.26	-32.89

Flow charts

• SAR-OPT_fusion

Scability

• The algorithm was implemented on Google Earth Engine with a standard (free) account, and the resultant images were exported to Google Drive. Details about the time and storage cost under several different scenarios were summarized below:

Scenario	Image size (pixels)	Number of images	Training and prediction time (mins)	Post-process time (mins)	Size of images to reconstruct
1	248 × 240	207	2	9	87.22 MB
2	500 × 414	437	10	60	802.36 MB
3	2750 × 4671	380	40	Failed (Exceeded GEE memory limit)	34.66 GB

• For application at large spatio-temporal extents, we suggest to:
  1. use other smoothing methods (e.g., spatial instead of temporal smoothers) for post-processing;
  2. download the raw predictions and smooth the data locally according to requirements;
  3. process the large area/timespan as small batches; and/or (iv) apply for larger memory limits from GEE.

Installation

To use GEE, you must first sign up for a Google Earth Engine account.

The project requires earth-engine-api and few other packages to be installed. A conda environment with the required dependencies can be created with

conda create -n GEE_SAR_OPT python = 3.7
conda activate GEE_SAR_OPT
pip install earthengine-api
pip install geopandas
pip install pandas
pip install json

To authenticate and initialize ee

import ee
ee.Authenticate()
ee.Initialize()

Usage

The setup and parameters can be tuned in the Parameters.json in config folder. To implement the gap infilling

cd scripts
python fusion_GEE.py

Output Dataset Paths

Output images and metadata will be saved in a Google Drive folder named according to the parameter "PROJECT_TITLE". GEE will search for the folder name on Google Drive from root to sub directories. If the folder has already existed in a subdirectory, outputs will be saved there, otherwise, GEE will create a new folder under the root path.

Credits

This project was funded by CSIRO Digiscape FSP. It involved Yongjing Mao, Tim McVicar and Tom Van Niel.