Towards ecosystem accounts for Rwanda: Tracking 25 years of change in flows and potential supply of ecosystem services

1Geosciences & Environmental Change Science Center, U.S. Geological Survey, Denver, CO, USA; 2Wealth Accounting and Valuation of Ecosystem Services (WAVES) Partnership, The World Bank, Washington, DC, USA; 3Ernst and Young and Wildlife Conservation Society, Washington, DC, USA; 4World Wide Fund for Nature International, Kigali, Rwanda; 5Wildlife Conservation Society, Kigali, Rwanda; 6CIAT-CGIAR Climate Change, Agriculture, and Food Security, Kigali, Rwanda; 7Rwanda Agriculture and Animal Resources Development Board (RAB), Kigali, Rwanda; 8International Institute of Tropical Agriculture (IITA), Bukavu, Congo; 9State Key Laboratory of Water Environment Simulation, School of Environment, Beijing Normal University, Beijing, China; 10Department of Environmental Management, Institute of Life and Earth Science, Pan-African University, University of Ibadan, Ibadan, Nigeria; 11Department of Applied Economics, University of Minnesota, St. Paul, MN, USA; 12Green Economy Advisory & Research, Kigali, Rwanda and 13Ministry of Finance and Economic Planning, Kigali, Rwanda


| INTRODUC TI ON
As the economies of many low-income African countries continue to grow, there is rising concern that current production and consumption models will undermine their ecological systems and limit the quality of growth on the continent (Egoh et al., 2012;IPBES, 2018;Marques et al., 2019). Governments throughout the region are committing to follow sustainable economic development pathways that maintain their natural capital to secure ecosystem services (ES) that are critical for livelihoods and economic development (Gaborone Declaration, 2019;UNEP, 2015). Across Africa, recent experience and economic models alike illustrate how 'green' investments can improve economic performance while conserving the natural resource base on which African economies and livelihoods depend (UNEP, 2015). Rwanda is among the countries making a commitment to green development within a rapidly growing economy. Rwanda has one of the higher GDP growth rates in Africa (World Bank, 2018a) with ambitions of transitioning into a middle-income country by 2020, as articulated in its 'Vision 2020' plan (Republic of Rwanda, 2000). The Government has pledged to pursue this growth through strategies that will maintain its natural capital through their commitment to the Gaborone Declaration, the national green growth and climate resilience strategy, establishment of a national natural capital committee steering committee, and demonstrating that the maintenance of natural capital is a critical component of its growth accountable governance. EDPRS II targeted 11.5% annual GDP growth; actual annual growth rates from 2010 to 2017 ranged from 4.7% to 8.9% (World Bank, 2018a). Land, water, forests and wildlife are the critical assets on which Rwandans rely for their livelihoods and support to industries such as energy, tourism and agriculture.
Yet, under a business-as-usual scenario, targeted economic growth, increasing population, land scarcity and competing demand for water by various industries is likely to put additional pressure on ecosystems and the services they provide. For instance, development goals to increase energy generation capacity and production of priority crops (maize, wheat, rice, Irish potatoes, beans and cassava) under the national Crop Intensification Program are important components of the country's development strategy. Achieving both will be heavily dependent on the condition of land and water resources (Kathiresan, 2011;Republic of Rwanda, 2013), yet is also likely to impact ecosystem services; the extent of these impacts under different development trajectories is not yet clear. For instance, Rwanda's goal of increasing agricultural self-sufficiency through increased fertilizer use (Republic of Rwanda, Ministry of Agriculture, & Animal Resources, 2013) has not been evaluated against its potential consequences for water quality. Similarly, the future of tourism will depend on forest and wildlife conservation within the country's protected areas; Rwanda's model of revenue sharing with respect to tourism fees has been globally recognized for demonstrating how wildlife conservation and tourism can benefit local communities (Spenceley, Habyalimana, Tusabe, & Mariza, 2010), though challenges remain (Munanura, Backman, Hallo, & Powell, 2016). Rwanda's primary linked environment and development challenge is the management of existing resources to meet the needs of a growing population who depend on natural resources for every aspect of their livelihoods. the evidence base needed to make decisions about where and how to develop without incurring negative trade-offs across industries or depleting the country's natural resource base and losing critical ES.

| Ecosystem service assessments and natural capital accounting for decision support
Decision-making about economic development and natural resource management can benefit from timely and accurate information about ES. Incorporating ES information into natural resource and economic planning can improve the overall value of benefits provided by the landscape and avoid unintended declines in ES provision (Guerry et al., 2015;Nelson et al., 2009;Polasky, Nelson, Pennington, & Johnson, 2011;Zheng et al., 2013). Similarly, incorporation of ES into national-level development planning can improve conservation and development outcomes and increase the likelihood of achieving sustainable development (Griggs et al., 2013;Miteva, 2019).
Conversely, the lack of a unifying data and analytical framework can lead to resource management problems. For instance, agricultural, water or energy ministries may individually plan to use the same resources without considering each others' plans or their combined effects. Ecosystem accounts provide a unifying framework to understand resource availability and use by different industries, helping assess trade-offs to determine optimal development paths (Australian Bureau of Statistics, 2017;Keith, Vardon, Stein, Stein, & Lindenmayer, 2017;World Bank, 2016).
Although ES assessments can be used to inform natural resource management in several ways, their use in natural capital accounts is increasingly common (Vardon, Burnett, & Dovers, 2016). Natural capital accounts comprise the System of Environmental-Economic Accounts (SEEA) Central Framework (SEEA-CF, U.N. et al., 2014a) and the SEEA Experimental Ecosystem Accounts (SEEA-EEA, U.N. et al., 2014b). Rwanda is 1 of 16 nations engaged in the development of natural capital accounts through the Wealth Accounting and Valuation of Ecosystem Services (WAVES) Partnership (World Bank, 2018b). Rwanda has developed land and water accounts as part of the SEEA-CF (Republic of Rwanda, 2018, 2019), which provide time series data on land cover , land use and transaction values (2014)(2015), and physical supply and use and asset tables for water (2012)(2013)(2014)(2015); mineral accounts are in development (Stage & Uwera, 2018). Building on these accounts, ES data organized within the SEEA-EEA can broaden country's understanding of the effects of recent development policies, helping to guide future planning.
Like all accounts, the SEEA-EEA includes sets of tables built using rules that quantify linkages between natural resources and economic production. The SEEA-EEA includes multiple components, including ecosystem extent and condition accounts, physical and monetary supply and use tables, asset accounts, and thematic accounts for land, water, carbon and biodiversity (U.N. et al., 2014b; U.N., 2017). In understanding ES trends to inform the accounts, it is critical to distinguish the potential supply of ES that can be produced by ecosystems from their actual flows to beneficiaries that account for beneficiary locations in relation to ES supply and beneficiaries' levels of demand (Hein et al., 2016 (NESCS, USEPA, 2015). In our case, the modelled ecosystem services of carbon storage, sediment regulation, nutrient regulation and water yield correspond to CICES version 5.1 classes 'Regulation of chemical composition of atmosphere and oceans', 'Control of erosion rates', 'Regulation of the chemical condition of freshwaters by living processes' and the 'Water' division of Provisioning (Abiotic) services, respectively. In NESCS, by contrast, carbon storage is not considered an ecosystem service, as it is not a final ecosystem good or service; carbon would thus be included solely in a SEEA-EEA thematic carbon account (U.N., 2017). All of our modelled sediment regulation, nutrient regulation and water yield metrics would fall under the NESCS ecological end-product 'Water', with their 'Environment' determined by the ecosystem type at the location of their use, and the use and user dependent on the type of water user (e.g. water company or utility, hydroelectric power generator or agricultural irrigator; USEPA, 2015). Important terminological differences also exist between the SEEA-CF and the SEEA-EEA, for which SEEA-CF 'Natural water', a physical input moved from the environment into economic production, corresponds to 'Water provisioning' in the SEEA-EEA Technical Recommendations (U.N. et al., 2014a; U.N., 2017).

| Evidence-based management: Have Rwandan environmental policies protected natural capital?
Making economic planning decisions that protect ES requires evidence on the impacts that alternative land use and resource allocation have on ES. In many fields, including business, education, economic development and medicine, there is movement towards systematically assessing the benefits of various interventions and basing management decisions on this evidence (e.g. Pfeffer & Sutton, 2006;Slavin, 2002;Straus, Tetroe, & Graham, 2011). There have also been calls for evidence-based policy and management in conservation and resource management (Sutherland, Pullin, Dolman, & Knight, 2004). At present, such evidence is unevenly used to inform decisions affecting ES through land use or resource allocation (McKenzie et al., 2014). Evidence-based ecosystem management requires data and models capable of accurately predicting the provision of ES under alternative land use and resource allocation, plus a greater understanding of how policy interventions impact the environment (Karamage et al., 2017). Models are necessary for predicting impacts of potential decisions (Schröter, Remme, Sumarga, Barton, & Hein, 2015), and their validation with field data can instil confidence in their predictive ability. Models can also help to quantify recent historical baselines for ES, that is, how future interventions may lead to divergence from recent trends. Additionally, where available, official statistics can also inform ecosystem accounts, particularly for provisioning ecosystem services .
A retrospective analysis of ES trends aligns well with the goals of the SEEA-EEA, which tracks ES trends and their contributions to specific economic industries, households and government.
Over the past two decades, the Government of Rwanda has enacted policies and legislation governing land use, to ensure sound land use and environmental protection for sustainable development.
Ecosystem accounts can be used to show how effective these policies have been in conserving natural capital and ES. For instance, soil erosion control, increased soil fertility and environmental protection have been emphasized in the major national development frameworks, particularly in Vision 2020 and EDPRS II (Republic of Rwanda, 2000Rwanda, , 2013. Recognizing how high demographic pressure has led to the occupation of marginal areas and rapid soil degradation in fragile ecosystems, Vision 2020 stated clearly that 'to ensure sustainable development, Rwanda will implement adequate land and water management techniques, coupled with a sound biodiversity policy' (Republic of Rwanda, 2000, p. 20). Various laws passed over the last two decades govern soil conservation, land management and general environmental protection, with particular emphasis on erosion control to support more sustainable agriculture.   (REMA, 2015(REMA, , 2017. Various NGOs are also working with partners throughout the country to address Rwanda's environmental, social and economic challenges by selectively raising naturally occurring trees with economic value. Despite these efforts, a lack of nationally consistent data and methods has limited the ability to assess the impact of these investments and policies on ES that are critical to the Rwanda's economic development. In this paper, we used ES models to quantify trends in service can be applied in ecosystem accounting (Schröter et al., 2015); in this case, we used the Integrated Valuation of Ecosystem Services Tradeoffs (InVEST) tool , building on preliminary modelling work done by the Natural Capital Project to apply these models in Rwanda and neighbouring Uganda (Gourevitch et al., 2016). We modelled carbon storage, sediment regulation, nutrient regulation, and annual and seasonal water yield. We summarized results for the nation, its five provinces, 30 districts, and four national parks, quantifying potential supply of these ES. Additionally, for sediment regulation and water yield, we quantified changes in ES flows (Hein et al., 2016) in watersheds upstream of irrigation dams, hydroelectric power dams and water treatment plants. For nutrient regulation, we quantified ES flow changes upstream of water treatment plants, a water user for whom excess nutrients may be problematic. This enables us to identify locations where changing land cover and agricultural practices may affect water supplies for hydroelectric power, irrigation and domestic supply. In these locations, future development plans may need to address how changes to water quantity, quality and timing may impact water security and economic development. Our work thus illustrates the application of ecosystem accounts to evaluate sustainability of ES in a small, rapidly changing African nation illustrative of the continent's rapid economic, population and environmental change (Egoh et al., 2012;IPBES, 2018).

| Study area
Rwanda is a landlocked country situated in the central African highlands with a total surface area of 26,338 km 2 and some 1,385,000 ha  Geological Survey, and Rwandan and US academics. Carbon storage, sediment regulation, nutrient regulation and water yield were identified as ES that would add value for decision-making and were feasible F I G U R E 1 Study area map. District numbers are used to organize districtscale ecosystem accounts by province (Supporting Information Appendix D) to quantify using existing data. Flood regulation was of interest to the government, but we lacked the needed input and calibration data for its modelling. As described below, we did run the InVEST seasonal water yield model, which estimates quick flow -runoff that occurs during or soon after rain events. Although not a flood model, it provides somewhat of a proxy for how the landscape's capacity for rainfall infiltration and flood regulation are changing over time.

| Modelling approach
We used the InVEST 3.3.3 modelling software  to quantify carbon storage, sediment regulation (sediment deliv- Loss Equation (Renard, Foster, Weesies, McCool, & Yoder, 1997) with a connectivity index to model sediment export and retention.
The NDR model uses land-cover specific estimates of nitrogen and phosphorus loading and potential nutrient uptake, combined with the SDR model's connectivity index to quantify nutrient export to downstream water bodies. NDR model outputs include nutrient load (kg N and P applied to or potentially released from each grid cell) and export (N and P reaching downstream water bodies). Nutrient retention (N and P retained by soils and vegetation that is prevented from reaching water bodies) can be estimated as the difference between nutrient loads and nutrient exports, but the model does not calculate nutrient retention on a grid cell basis. The annual water yield model uses the Budyko curve method (Fu, 1981) to model actual evapotranspiration (AET), then subtracts AET from mean annual precipitation to quantify annual water yield. Finally, the seasonal water yield model estimates quick flow (runoff during and immediately following storm events, which can cause problems with flooding, water quality and dry-season water availability; estimated using the Curve Number We summarized results for all five provinces and 30 districts in Rwanda, as well as for the nation's four national parks. We also used spatial data for existing and planned water infrastructure use locations, specifically 33 irrigation dams, 24 hydroelectric dams and 39 water treatment plants to distinguish between the potential supply of water and sediment regulation for the entire nation and actual flows that reach different water users (as well as actual flows of nutrient exports to water treatment plants; Hein et al., 2016). We then evaluated changes in sediment export, phosphorus exports (phosphorus

| RE SULTS
Increases in annual water yield indicate less evapotranspiration (i.e. water regulation through vegetation) and more runoff, but their implications are somewhat ambiguous. Analysis of local recharge and quick flow thus provides a more complete view of changes in water yield . Ecosystem condition tables (quantifying nutrient loads and exports and sediment export), potential supply tables (that account for total production of potential benefits to people) and physical supply tables ( Information Appendix D and shown below for the nation (Table 1).  Figure 4). However, ecosystem extent is a better predictor for simple models like carbon sequestration than for more complex models like sediment and nutrient retention, which depend not just on land cover but also on soils, topography, climate and agricultural practices. We estimated nation-wide water yield at 7.14 billion m 3 in 1990, which increased to 7.42 billion m 3 by 2015. While small, these water yield increases are not necessarily desirable. Less evapotranspiration leaves more water available for surface and groundwater resources, but this water may runoff quickly, contributing to water quality problems and reduced groundwater recharge. To address these limitations, the seasonal water yield model results (described below) add additional nuance to our analysis of hydrologic ES.

| National, provincial and district-level potential ecosystem service supply changes
Water yield in Rwanda strongly follows the country's rainfall gra-

| Potential ecosystem service supply changes in protected areas
Nyungwe NP saw a small decline in carbon storage throughout the study period, from 40.0 to 38.7 MT (Figure 7). Akagera NP wit-

| Ecosystem service flows for irrigation, hydroelectric power and drinking water
Between 2010 and 2015, nation-wide sediment export increased by 38.9%, phosphorus exports increased by 10.6% and quick flow increased by 10.2%. By comparison, sediment export and quick flow in watersheds upstream of all hydroelectric dam sites increased modestly (+14.3% and +8.8%, respectively; Figure 8). For irrigation dam and water treatment plant sites, increases in sediment export were greater than the national average (+43.5% and +47%, respectively), as were increases in quick flow (+10.5% and +12.8%, respectively). Phosphorus exports upstream of water treatment plants increased by 20.2%.
Sites for 14 irrigation dams, 11 hydroelectric dams and 16 water treatment plants had upstream quick flow and sediment export increases above the national average ( Figure 8).

| Caveats
In this paper, we modelled ES in Rwanda using the latest data, a calibrated annual water yield model, terracing data to inform a soil erosion model (Supporting Information Appendices A-C), and applied a seasonal water yield model to better quantify hydrologic ES.
Relative to past ES modelling efforts in Africa that rely on older, coarser resolution, or global data or lack calibration (Gourevitch et al., 2016;Leh, Matlock, Cummings, & Nalley, 2013;Rukundo et al., 2018), results of this study are relatively robust. Still, several important caveats apply.
First, sediment and nutrient load data needed to calibrate soil erosion and nutrient models are scarce in Rwanda (Muvundja et al., 2009;Uwimana, Dam, Gettel, & Irvine, 2018); an attempted calibration using sediment load data for watersheds draining into Lake Kivu (Muvundja et al., 2009) had relatively low predictive power. A national water-quality monitoring programme with adequate spatiotemporal coverage and co-located with stream gages could provide data to calibrate sediment and nutrient models. New water-quality monitoring efforts begun in 2017 by the Ministry of Natural Resources may aid in future model calibration efforts (Christian & Vedaste, 2017).
Second, our analysis of dams and water treatment plants required delineation of their upstream watersheds. We confirmed the

F I G U R E 7 (Continued)
locations of some existing facilities using satellite imagery. However, watershed delineation requires that facilities be accurately located on river flow accumulation lines, and in some cases, the facility location was ambiguous, introducing uncertainty into our analysis.
We suggest a full review of facility locations with Rwandan utility or agency staff to confirm exact water-use locations prior to the use of this information in precise spatial natural resource planning.
Additionally, the use of the national average and 50% increases in sediment export and quick flow (Figure 8)  Finally, given data limitations and unresolved conceptual issues (e.g. the fact that sediment retention is a non-rival service simultaneously benefitting multiple water users in the same watershed), it was not possible to construct a physical use table. This conceptual issue is one of many being addressed by the SEEA-EEA revision process

| Next steps for ecosystem accounting in Rwanda
We quantified ES in Rwanda for four intervals over a 25-year period. While adequate to show long-term trends, 5-to 10-year intervals are suboptimal for integration with more regularly produced national economic accounts, particularly given Rwanda's demand for official statistics in decision-making (Stage & Uwera, 2018).
Land-cover data -a key input to many ES models -are now increas- requires more complex models and underlying data than are currently available. Future ecosystem accounts could also address provisioning and cultural ecosystem services, for instance, tourism (Banerjee et al., 2018) and fuel, timber and food (NISR, 2018b).
Inclusion of more ecosystem services, including provisioning and cultural services, would more fully quantify ecosystems' contribution to Rwanda's economy.
InVEST provided a suitable platform for biophysical modelling of potential ES supply (Hein et al., 2016), and is one of several modelling tools available for ecosystem accounting (Bagstad, Semmens, Waage, & Winthrop, 2013

| CON CLUS IONS
In this paper, we provide an initial view of ecosystem condition and physical ES supply trends for Rwanda that can serve as a foundation for more complete ecosystem accounts, including analysis of additional ES and monetary accounts (U. N. et al., 2014b). Economic valuation to develop monetary supply-use tables is a next step and will require greater integration with both Rwanda's water and national economic accounts. Although simple valuation is currently possible using, for example, the social cost of carbon and water productivity data from the water accounts (Republic of Rwanda, 2019), more sophisticated but data-intensive approaches to value ES like sediment regulation would be more informative for decision-making.
This could incorporate additional key industries such as mining and coffee and tea production.

CO N FLI C T O F I NTE R E S T
The authors declare no conflict of interest. wrote the paper. All authors approved the final version of this paper.

DATA AVA I L A B I L I T Y S TAT E M E N T
A data release containing all relevant datasets will be available within the US Geological Survey Science Data Catalog https ://data.usgs. having >70% tree canopy cover, moderate forests from 40% to 70% cover and sparse forests as 10%-40% cover. Transitions between forest cover classes indicate the crossing of these thresholds, but do not tell us how far a given cell is from that threshold. It is thus possible that some changes in forest density could reflect cells with forest cover near a threshold that crossed it from one time period to the next. 2 RCMRD has produced two land-cover data products for Rwanda -Level I with 6 land-cover classes and Level II with 13 classes. Like any land-cover data, classification error rates will be lower for datasets with fewer land-cover classes. For the 1990-2010 data, Level II accuracy was 80.0% and Level I accuracy was 86.4% (RCMRD-SERVIR Africa, 2013). For 2015, Level II had 77.5% accuracy and Level I 79.6% (RCMRD, 2017) -so relatively little accuracy was lost in moving from 6 to 13 land-cover classes. While the 2015 maps were produced as part of a separate effort than the 1990, 2000 and 2010 maps, the developers of the 2015 data used similar methods and performed accuracy assessments and comparisons with the earlier data to ensure its compatibility. Rukundo et al.'s (2018) trends for ES modeled based on Level I land-cover data for 1990 to 2010 were quite similar to ours, which gives us further confidence that ES trends are robust to the choice of the chosen thematic resolution for land-cover maps. We lacked the data to uniquely parameterize open versus closed grasslands and shrublands in the carbon, RUSLE C factor (SDR model), and Kc and root depth (annual water yield model) lookup tables, and different forest, grassland/shrubland, and cropland types in the NDR model lookup table (Supporting Information Appendix A). In these cases, a more detailed land-cover classification does not help us to make a more refined assessment, but further ecological studies could help to better distinguish between ES provision differences when using Level II land cover in future ecosystem accounts.