Multi-season Sentinel-2 level 2a image stack classification using Random Forest in Google Earth Engine

Script for Google Earth Engine to obtain a multi-season image stack and classify it (with user-defined ground truth) using Random Forest machine learning approach

Requirements

For running, the script requires user-defined Area Of Interest (spatial polygon) and ground truth data (spatial polygons with numerical classes). Both files should be loaded into GEE environment as separate assets.

For correct displaying of classified imagery one should manually edit the palette, using desired class color codes and ajust the range of coloring according to the number of classes.

Outputs

1. Classified satellite imagery for user-defined area of interest (exportable to Google Drive as .tiff raster file)
2. Variable importance chart and values
3. Overall accuracy metric
4. Cohen's Kappa metric
5. Confusion matrix (explortable to Google Drive as .CSV file)

Example of results

Southern Buh river valley, Ukraine (47.96 N, 31.03 E).

Contributors

Oleh Prylutskyi - idea and code

Dariia Shyriaieva - idea and ground truth

Anton Biatov - code

References

Shyriaieva, D., Prylutskyi, O. (2021). Exploratory analysis of the spectral reflectance curves of habitat types: a case study on Southern Bug River valley, Ukraine. In: 63rd IAVS Annual Symposium: Book of Abstracts, p. 153.