Navigating deep learning strategies for large-area land cover mapping using very-high-resolution imagery in Senegal: Validation Data

Rapid advances in deep learning for land cover classification of trees, shrubs and very small agricultural fields using very high-resolution satellite data (< 2 m), has tremendous potential for resolving current challenges in quantifying land cover change in sub-Saharan African (SSA), due to growing demand for food resources. We conducted experiments with different training strategies for scaling up UNet convolutional neural network models for regional land cover mapping with multispectral WorldView (WV)-2 and –3, imagery in three distinct regions of Senegal which has complex seasonal wet/dry conditions and cropland-savanna mosaics. 

The validation exercise of this research consisted in validating more than 70,000 km² across Senegal. The infrastructure was setup in the NASA SMCE system with a total of twelve George Mason University (GMU) students participating as operators. These operators validated more than 59 WV-2 and -3 images, each consisting of 200 stratified points in 5,000 x 5,000-pixel images. This effort resulted in a total of ~35,000 aggregated observations that are available through the eo-validation API for public consumption. Each validation point from this dataset has three individual observations.