Cloud-computing and machine learning in support of country-level land cover and ecosystem extent mapping in Liberia and Gabon.

Abstract

Liberia and Gabon joined the Gaborone Declaration for Sustainability in Africa (GDSA), established in 2012, with the goal of incorporating the value of nature into national decision making by estimating the multiple services obtained from ecosystems using the natural capital accounting framework. In this study, we produced 30-m resolution 10 classes land cover maps for the 2015 epoch for Liberia and Gabon using the Google Earth Engine (GEE) cloud platform to support the ongoing natural capital accounting efforts in these nations. We propose an integrated method of pixel-based classification using Landsat 8 data, the Random Forest (RF) classifier and ancillary data to produce high quality land cover products to fit a broad range of applications, including natural capital accounting. Our approach focuses on a pre-classification filtering (Masking Phase) based on spectral signature and ancillary data to reduce the number of pixels prone to be misclassified; therefore, increasing the quality of the final product. The proposed approach yields an overall accuracy of 83% and 81% for Liberia and Gabon, respectively, outperforming prior land cover products for these countries in both thematic content and accuracy. Our approach, while relatively simple and highly replicable, was able to produce high quality land cover products to fill an observational gap in up to date land cover data at national scale for Liberia and Gabon.