This repository provides the code and data that accompanied our paper titled Mapping Construction Grade Sand: Stepping Stones Towards Sustainable Development chosen as a poster for COMPASS 2023 and a paper at KDD / Fragile Earth Workshop 2023.
There are 2 levels of data aggregation that we performed:
-
Aggregated geocoordinates and grain size information from various academic sources This dataset can be found in a Google Sheet here. Each sheet contains data from a separate source.
-
The positions from the dataset are fine-tuned, along with a timestamp (if the SGR desposits are not available on the original date), and a set of spectral bands are grabbed, object based image analysis (OBIA) is conducted and the median value of those bands are stored. These are available in
data/labelsa. The base dataset is:
gt-bands.xlsx, which contains all S-1 and S-2 median band values for the cluster identified using SNIC for each sampleb. The other modfied dataset that could be of use are
gt-bands-resampled-s10-dw.xlsxthat also has the Dynamic World (DW) class probability layers as bands. Other variations includegt-bands-resampled-s15-dw.xlsxorgt-bands-resampled-s15.xlsxwhich are resampled versions of the base dataset with different superpixel parameters for the SNIC algorithm, with or without the DW bands.
If you would like to simply use the best trained random forest (RF) model to visualize any region on the planet, within Google Earth Engine (GEE), which is open access and requires no additional computation, except having a GEE account, you can use this publicly available GEE file to run those inference steps. Simply hit "Run" on the command bar on top. Beware that at lower zoom levels (more area shown), you may encounter errors, and will have to zoom in to reduce the computer burden.
The best performing random forest (circa July 30, 2023) can be found here. This was trained using the code in /validation/global_validation.ipynb. You can use this earth engine asset and generate predictions directly in GEE.