More information about the project
Support for climate change mitigation policies has been mounting over the past two decades, both within countries as well as internationally. The loss of forest cover accounts for between 12 and 15 percent of the annual anthropogenic emissions of greenhouse gases – the second largest source after fossil fuel combustion (Canadell et al., 2007). Forest conservation is deemed to be a cost-effective way to mitigate climate change (Nabuurs, 2007). Not only are the opportunity costs of forest conservation relatively limited, there are also substantial co-benefits in the form of improved local climate regulation, better water storage, and biodiversity conservation (Canadell et al., 2007; Stern, 2007). Economic activities affecting loss of forest cover, especially in the drylands of Africa, include – but are not limited to – agricultural expansion, overgrazing, forest fires, demand for firewood and charcoal, over-exploitation of non-wood forest products, and mining (Griscom et al., 2017). In addition, the increasing need to grow food for a growing population around the world, coupled with the still widespread use of non-sustainable production practices, translates into increasingly serious forest degradation which threatens the livelihoods of both current and future generations. Halting deforestation has thus become a central objective in the climate policy of international agencies. A key example in point is the United Nations' initiative “Reducing Emissions from Deforestation and Forest Degradation” (REDD+) (Ministry of the Environment and Sustainable Development, Government of Burkina Faso, 2012). This initiative has gained wide popularity since the 2015 Paris Agreement, and several new countries have joined (or are preparing to join) this initiative.
Effective protection of forest resources requires detailed knowledge about the status of the resources, as well as the capacity to monitor changes. More importantly, the implementation of conservation policies such as payments for avoided deforestation, or the assessment of the impacts of forest conservation programs in general, demands the ability to regularly estimate – and as accurately as possible – the size of the resource stock as well as the changes therein. Global datasets of land cover, including tree cover, are now publicly available, including Global Forest Watch at 30 m resolution (based on Hansen et al., 2013), the ESA Land Cover CCI at 300 m; the Global Land Cover dataset at 30 m from China for 2000 and 2010; and Global, Landsat-based forest-cover change from 1990 to 2000 from Kim et al. (2014); for an overview, see Tsendbazar et al. (2014). Global land cover datasets are critical and cost-effective sources of information when national mapping capacity is not available yet. But national mapping is considered more accurate, as it can better account for the local circumstances (Global Forest Observation Initiative, 2014). Definitions of both land use and forest cover can vary largely with the context.
In this paper we present a low cost, easy to implement alternative approach to estimate forest cover relying on 10-m resolution Sentinel-2 imagery, highlighting the potential of large temporal and spatial capabilities using free publicly-accessible platform and data. This method will be especially useful for those countries that lack national mapping capacity to estimate their forest inventory, including many of the more arid countries in Africa. We leverage recent advances in satellite imagery technology to analyze large sets of images, and develop a land cover map for twelve Burkina Faso gazetted forests. The paper contributes to the available literature on satellite based image classification to map forest cover, especially in the drylands of Sub-Saharan Africa (SSA). We show the use of different spectral bands and remote sensing-derived indices to run a multi-spectral-based assessment at the pixel level, which is our spatial analysis unit. In addition, we present the accuracy metrics of our estimations at three different probability thresholds indicating how true positive and true negative rates change across them.
This research is implemented as part of DIME's impact evaluation support to the Forest Investment Program (FIP) in Burkina Faso, a targeted program of the Strategic Climate Fund set up under the Climate Investment Funds (CIF)3 (Climate Investment Funds, 2014), from which Burkina Faso has benefited. The project includes the Gazetted Forest Participatory Management Project for REDD+ (PGFC/REDD+) financed through the African Development Bank (AfDB) (African Development Bank Group, 2013), which is aiming to conserve forest cover in Burkina Faso. The information provided in this paper will enhance Burkina Faso's ability to plan, implement and monitor the success of their FIP, and provide lessons for other countries in the region that are also part of the FIP initiative.
In this paper, we present tree cover estimates for the 12 gazetted forests in Burkina Faso that are targeted by the FIP project, between March and April 2016. The method used relies on a multi-spectral image classification at 10-m resolution, improving the prediction power, and therefore the mapping precision, compared to Landsat-based classifications. Sentinel-2 imagery contains thirteen spectral bands, four of which are sensed at a 10-m resolution: red, green, blue and near infrared (NIR) bands (European Space Agency, 2017). Goldblatt et al. (2017) have found that higher-resolution images improve the classification accuracy of ecosystems with relatively little tree crown cover, like forests in arid or semi-arid areas, which cannot be detected with Landsat imagery.
Image classification is performed using a random forest algorithm as classifier, as well as the Google Earth Engine (GEE) platform which combines a multi-petabyte catalog of satellite imagery and geospatial datasets with planetary-scale analysis capabilities allowing faster GIS and remote sensing computing. The general classification approach is based on the construction of a ground truth dataset that leverages on a false color composite image to label pixels with higher accuracy, and then uses a k-fold cross-validation approach to determine the probability of a trained pixel to be classified as tree-covered based on its spectral signature. The output is a collection of binary rasters that covers the total area of each one of the 12 gazetted forests of interest, indicating the existence of trees and displaying the accuracy rate of the results.
This paper is organized as follows. Section 2 describes the area of study indicating the agro-ecological characteristics of the forests of interest as well as the method and data used for the classification. Section 3 cover the calculations and the data construction process while section 4 presents the classification results. The fifth section concludes the paper with a discussion of the results.
The above content is cited from https://www.sciencedirect.com/science/article/pii/S2352728517300891