A super-ensemble approach to map land cover types with high resolution over data-sparse African savanna landscapes

Abstract

Accurate and timely land cover products are critical inputs for landscape planning, and provide key information for biodiversity conservation and food security. However, poor mapping quality and low resolution are considerable issues in existing land cover maps over the African savanna, where land use is complex and changing rapidly, and necessary ground-truth data are sparse and hard to obtain. To overcome this problem, to make optimal use of existing maps, and to minimize manual training data collection, we developed a three-stage ensemble method to make land cover maps. In the first stage, we extracted the consensus of multiple existing land cover products to generate fragmented pixel-wise training labels. In the second stage, we translated pixel-wise training labels to image-wise labels using Random Forest (RF) as a “gap-filling model”, with temporal features extracted from Sentinel-1 time series, raw bands, and vegetation indices derived from PlanetScope basemaps. These image-wise labels were scored and edited by humans and the quality information was used in the next stage. For stage three, we trained a U-Net network based upon these image-wise labels, using Sentinel-1 time series and raw bands of PlanetScope basemaps as image features. Using the information on label quality, a quality-weighted loss function was used in the network to reduce the impact of noise in the training labels. Using Northern Tanzania as a case study, the results demonstrate that ensembles of existing land cover maps provide a useful source of data for developing improved land cover maps over hard-to-classify, data-sparse landscapes. The Random Forest “gap-filling model” had an overall accuracy of 80.26% on our independent test dataset with 7 classes. The final U-Net model had an overall accuracy of 83.57%. This approach can be readily applied to other regions and extents (e.g., regional, global) and other data sources (e.g., Sentinel-2).