Above-ground biomass mapping in West African dryland forest using Sentinel-1 and 2 datasets - A case study

Highlights

• Sentinel-2 found to be a better predictor of AGB than Sentinel-1.

• Combination of Sentinel-1 and 2 produce better results than either of them.

• Indices and biophysical parameters useful than spectral bands in AGB mapping.

• Dry season images achieve better results in AGB mapping than rainy season images.

• Open access satellite data can assist developing countries to attain SDG targets.

Abstract

The Sudanian Savanna (SS) of West Africa is characterized by tropical savannas and woodlands. Accurate estimation of AGB and carbon stocks in this biome is important for addressing sustainable development goals as the information can aid natural resource management at varied spatial scales. Previous AGB mapping efforts focused on humid forests, with little attention on savannas. This study explored the use of annual monthly time-series of Senitinel-1 (S-1) and Sentinel-2 (S-2) data to map AGB in the SS. Backscatter, spectral reflectance, and derivatives (vegetation indices and biophysical parameters) were combined with field inventory data in a Random Forest regression to map AGB. Eight experiments were conducted with different data configurations to determine: (1) the potential of S-1 and S-2 for AGB mapping, (2) optimal image acquisition period for AGB mapping, and (3) contribution of image derivatives to improving the accuracy of AGB mapping. The predicted map was validated with 40% of the inventory data. Uncertainty in the AGB was assessed using mean absolute error, root mean squared error, coefficient of determination and symmetrical mean absolute percentage error. Results show that about 90% of the study area have low AGB stocks of less than 90 Mg/ha. Compared to S-1 (RMSE: 78.6; MAE: 25.6), S-2 achieved better prediction accuracy (RMSE: 60.6; MAE: 19.2), although combination of the two according to seasonality produced the best results (RMSE: 45.4; MAE: 16.3). Images acquired in the dry season were found to be more useful for predicting AGB than those of rainy season. Also, stress-related vegetation indices and a red-edge dependent normalized difference vegetation index not tested in previous AGB studies using Sentinels were found to be significant contributors to the superior performance of S-2. Since biomass is a finite resource, our results can provide valuable information on the sustainable use of biomass and energy security including studies on carbon cycling and ecosystem functions in the region. The demonstrated possibility of using open access earth observation data to map and monitor AGB in data scarce regions is useful and beneficial to attaining SDG indicators 15.2.1 (sustainable forest management) and 15.3.1 (proportion of land that is degraded over total land area). Further work on developing species-specific wood densities and allometric equations is required to improve AGB and carbon stock estimation in the SS.

Introduction

West Africa has been identified as a climate change hotspot where an increased probability of hazards, vulnerability, and exposure meet (Heubes et al., 2013; Sylla et al., 2015). Climate change and variability (CCV) and its effects are inducing significant land use/land cover (LULC) changes in the sub-region, resulting in unprecedented deforestation rates, degradation of arable lands and deterioration of ecological systems (Zoungrana et al., 2018). These challenges call for appropriate adaptation and mitigation strategies to reduce the adverse effects of CCV on the sub-region's socio-ecological systems. Several global programs have been established to tackle the effects of CCV at varying scales.

The United Nations Sustainable Development Goals (SDGs) aim to, among other things, protect, restore and promote sustainable use of terrestrial ecosystems by ensuring sustainable management of forests and reverse land degradation (SDG 15). Monitoring the spatio-temporal changes in biomass and carbon stocks is essential for the realization of several SDGs and global programs such as Reducing Emissions from Deforestation and Forest Degradation (REDD+) (Herold et al., 2011; Saatchi et al., 2011). Above and below ground biomass stocks are important sub-indicators for achieving land degradation neutrality target (15.3.1) (UNEP/CBD/SBSTTA, 2016) and sustainable forest management (FAO, 2017). Consequently, determination and monitoring of carbon stocks in forests and other land uses are important for reducing the effects of CCV and achieving sustainable development in the sub-region (Asner et al., 2013; Skutsch and Ba, 2010).

Remote Sensing (RS) data are suitable for mapping the spatial distribution of above-ground biomass (AGB) or carbon stocks over large areas. Previous mapping efforts at the scale of Africa mostly used coarse spatial resolution RS imagery (ca. 0.5 to 1 km) with limited ground reference data (Avitabile et al., 2016; Baccini et al., 2012; Dobos et al., 2001; Baccini et al., 2008). Few studies recently used relatively high resolution (30 m) optical (Baccini et al., 2017) and Synthetic Aperture Radar (SAR) data (25 m) (Bouvet et al., 2018) for large scale mapping. Despite these improvements, existing products/maps give conflicting and inconsistent AGB estimates, especially for the savanna and woodland regions of dry forests (Bouvet et al., 2018). The fragmented and heterogeneous nature of tropical dry forest areas require further efforts by way of data and methodological approaches to improve AGB estimation. Exploring the complementary use of higher resolution optical and SAR data, and incorporating significant and representative field inventory data, can enhance the accuracy and consistency of AGB estimation in tropical dry forests.

The Copernicus program of the European Commission (Moreno et al., 2012) provides open access high resolution optical (Sentinel-2, S-2) (Drusch et al., 2012) and SAR (Sentinel-1, S-1) (Torres et al., 2012) data for terrestrial research applications. Recent studies have explored these datasets, singularly or complementarily, to map forest AGB stocks in a variety of biomes. Chen et al. (2018) used S-1, S-2 and their derivatives (e.g. texture, spectral indices, biophysical variables) to map forest AGB in Jilin Province, northeast China. One parametric (Geographically weighted regression, GWR) and three non-parametric machine learning algorithms (Support Vector Machines for Regression, SVR; Random Forest, RF and Artificial Neural Networks, ANN) were compared in modeling AGB. They found both datasets to be suitable for forest AGB estimation, especially textural properties from S-1 and biophysical variables from S-2. SVR outperformed the other tested algorithms. A similar study by (Chen et al., 2019, Chen et al., 2019), which combined S-1, S-2 and elevation data from the Shuttle Radar Topographic Mission (SRTM), further emphasized the suitability of S-1 and S-2 for forest AGB mapping. However, they found RF to have outperformed GWR, SVR, ANN and linear multiple regression. Pandit et al. (2018) applied S-2 data and derived spectral indices to map AGB in sub-tropical buffer zone community forests in Nepal using RF algorithm. They found that compared to the use of only spectral bands, the inclusion of vegetation indices improved AGB prediction accuracy. The study noted the possible contribution of red-edge derived vegetation indices in enhancing the performance of the regression. An integration of Unmanned Aerial Vehicle (UAV) data, S-1 and S-2 was performed by Navarro et al. (2019) to map AGB in a mangrove plantation in Senegal. Results of a UAV-based field AGB estimation procedure (response) were regressed against predictors from S-1 (VH backscatter) and S-2 (spectral bands and vegetation indices) to predict AGB values for the entire study area. SVR was selected for the modelling and prediction. Comparison of the models using the Akaike information criterion (AIC) showed S-1 to perform better than S-2, although a combination of the two datasets produced the best results. The study further revealed that the spectral indices from S-2 were more important than spectral bands in AGB estimation. Other studies that used S-1 and/or S-2 affirmed their suitability for AGB estimation (Chen et al., 2019, Chen et al., 2019; Haywood et al., 2018; Jay Labadisos Argamosa et al., 2018).

This study seeks to contribute to the existing knowledge on the suitability of S-1, S-2 and their derivatives for AGB mapping in the tropical dry forest of the Sudanian Savanna (SS) agro-ecological zone of West Africa. Four main vegetation types are considered – agroforestry parklands, shrub savannas, woodland/tree savannas and forest. Previous efforts at AGB estimation in sub-Saharan Africa focused on tropical humid forest areas (Akindele and LeMay 2006; Bakayoko et al., 2012; Lindsell and Klop, 2013) due to their high carbon storage potential (193–200 tons carbon/ha). On the other hand, dry forests and their associated vegetation types, estimated to have carbon storage potential of 17–70 tons carbon/ha have received less attention (Gibbs et al., 2007; Skutsch and Ba, 2010). But there are several reasons why tropical dry forests can no longer be neglected in terms of improving AGB estimation and understanding its dynamics: (1) their estimated coverage of about 22% of Africa's land area (Simons et al., 2001), (2) widespread degradation from high population growth vis-à-vis scarce resources (Campbell et al., 2007) and (3) the signing on of most countries in this biome to the REDD + program and international conventions such as the Paris Agreement (Skutsch and Ba, 2010).

To improve biomass and carbon stock estimation in the tropical dry forest regions of the SS in West Africa, we utilized high resolution multi-temporal optical (S-2, Landsat 8) and annual monthly time-series SAR (S-1) data together with a comprehensive field dataset of inventory plots to map AGB stocks (White, 1986). S-1 and S-2 were selected because they are new, open access and have superior spatial, spectral and temporal resolution compared to other open access datasets. Landsat 8 (L-8) data were used to fill-in areas that S-2 data were not available. RF (Breiman, 2001), a machine learning non-parametric algorithm, was used for modeling AGB. Previous studies have confirmed the superior performance of non-parametric machine learning algorithms (MLA) (e.g. ANN, SVR, RF) over parametric algorithms (GWR, linear regression) (Chen et al., 2018; Wålinder, 2014). This is mainly due to the ability of the former to handle complex non-linear relationships between variables from multi-source data. However, comparisons between MLAs in estimating forest AGB and other terrestrial properties (e.g. soil) have been mostly inconclusive. For example, whereas Chen et al. (2018) found SVR to outperform RF and ANN in forest AGB estimation, Chen et al., 2019, Chen et al., 2019 found RF to outperform SVR and ANN in a similar study. In addition, MLA comparative studies that tested for significance found the differences between them to be minimal and statistically insignificant (Adam et al., 2014; Freeman et al., 2016). These notwithstanding, RF was selected in this study due to the relative ease of tuning, robustness against noise and its inherent ability to extract variable importance measures.

Apart from testing the suitability of S-1 and S-2 for AGB mapping in the SS, this study advances the knowledge in AGB estimation using multi-sensor satellite data in three unique ways. First, we investigated the optimal image acquisition period (within a year) for AGB estimation by analyzing annual multi-temporal (monthly) time-series of S-1 and S-2. This knowledge can reduce image processing efforts in future mapping exercises as well as financial burden of image acquisition if commercial sensors are to be used. Second, we demonstrated the complementary use of S-1 and S-2 according to rainfall seasons in the SS. Specifically, experiments were set up to determine the optimal combination of S-1 and S-2 data for improved AGB mapping. Third, a red-edge dependent index and stress-related vegetation indices (Thenkabail et al., 1994) that have not been tested with S-1/S-2 data in recent AGB estimation studies were used. To our knowledge, no study has tested S-1 and S-2 data for AGB mapping in the SS biome. Thus, we focused on the following objectives: (1) determine the potential of S-1 and S-2, individually and in combination, to map AGB in the SS agro-ecological zone of West Africa, (2) determine optimal image acquisition period for AGB modelling and (3) investigate the contribution of derivatives (e.g. indices and biophysical parameters), to AGB mapping, and (4) establish the importance of accurate AGB maps to achieving targeted indicators in the SDGs.