Combining Multi-Satellite Optical Images for Improved Land Cover Mapping Using a Random Forest Classifier

The country of Panama is a signatory to the United Nations Sustainable Development Goals (SDG) and has been working towards informing on SDG 15.2.1 progress towards sustainable forest management, which consists of tracking net forest area change on an annual basis. Their efforts have been based on the use of optical imagery to generate land cover maps of the country, which run into coverage challenges as approximately 5% of each annual mosaic is blocked by clouds over areas primarily covered with forests. A way to mitigate this is a consolidated land cover product based on multiple optical datasets with greater coverage and accuracy than land cover maps based on any single source. We employ a decision tree based supervised classification algorithm called a Random Forest (RF) to generate annual land cover classification maps for Panama at 20m resolution using optical data from Landsat-8 OLI and Sentinel-2 MSI. Elevation data from SRTM is also used as a feature in both datasets. An RF model with forest, non-forest, and pasture classes is fitted for Panama and annual forest/non-forest (FNF) classification maps are generated based on the entire image collection for 2019. Training labels are obtained from a 2019 reference image generated by the Ministry of the Environment of Panama (MiAmbiente). Along with land cover maps based on Sentinel-2 MSI and Landsat-8 OLI individually, we consider an ensemble model which combines the inference from the Landsat-8 and Sentinel-2 models post classification. The ensemble model has the highest coverage and also better classification accuracy than models based on individual data sources. Results help support efforts to inform on SDG 15.2.1. Portions of this work were carried out at the University of Maryland, Baltimore County and at the Jet Propulsion Laboratory, California Institute of Technology under contract with the National Aeronautics and Space Administration.