Mapping Perennial Crops in Africa: A Case Study of Oil Palm in Ghana

Abstract

Forests in Sub-Saharan Africa are experiencing some of the highest rates of deforestation and degradation in the world, with most natural forest species being replaced by cropland and plantation monoculture. In this work, a method was developed that combined the Synthetic Aperture Radar (Sentinel-1) and optical satellite imagery (Sentinel-2) data to accurately map natural forest and perennial crops (oil palm) in Ghana. This was done using all three variables including spatial, spectral, and temporal variables to assess the most important variables in characterizing oil palm and natural forest, as well as the added value of sentinel-1 SAR data in a sentinel-2 optical-based classification. In this workflow, the Gray level co-occurrence matrix (GLCM) was calculated as representing textural/spatial variables, a yearly median composite to represent the spectral variables, and raining and dry season composites of Normalized Difference Vegetation Index  (NDVI) and Normalized Difference Moisture Index (NDMI) to represent the temporal variables for the Sentinel-2 data. In terms of the SAR data, rainy and dry season composites of NDVI and NDMI were calculated. With all these variables together, a characterization of the study area was conducted based on reference data of the land use land cover classes including oil palm, natural forests, and croplands (others) using Random Forest classifier. The variable importance of the Random Forest model was investigated to identify the top 10 most important variables. Results from this study showed that spectral variables followed by spatial variables are the most important and need to be considered when characterizing oil palm and natural forest, which is consistent with some pieces of literature. The use of sentinel-2 data achieved an acceptable classification accuracy (75%); whereas, sentinel-1 SAR further increased the accuracy (up to 85%) as compared to sentinel-2 only.