Keywords

1 Introduction

Plant photosynthesis on land takes up about 120 billion tons of carbon (C) per year, which is equivalent to 440 billion tons of CO2 (Friedlingstein et al. 2020). This is about 10 times more than the global annual CO2 emissions of 43 billion tons of CO2 (Friedlingstein et al. 2020). About half of this uptake is used by plants as respiration to maintain their metabolism and for nutrient uptake (Gonzalez-Meler et al. 2004). The rest is available as net primary production (NPP) to grow new biomass, replace leaves and fine roots, transfer sugars to mycorrhizal fungi in the soil, and produce root exudates and biogenic volatile organic compounds (Chapin et al. 2011).

NPP is the carbon gained by plants using photosynthesis at the ecosystem level after subtracting the respiration costs, and can thus be calculated as the gross primary production (GPP) minus plant autotrophic respiration (Ra) (Chapin et al. 2011). NPP is important for providing fundamental resources for all animals and fungi on Earth including ecosystem services for people such as food, fibre, and timber production (Melillo et al. 1993; Abdi et al. 2014; Ardö 2015; Pan et al. 2015). The importance for human society can also be illustrated through the large fraction of NPP used by humans, which has been estimated as human appropriation of net primary production (HANPP) (Fetzel et al. 2012). HANPP is currently estimated to be about 25% of global NPP (Haberl et al. 2007; Niedertscheider 2011; Abdi et al. 2014; Andersen and Quinn 2020).

The main direct drivers of NPP include: temperature, precipitation, solar radiation, the CO2 concentration in the atmosphere, nutrient availability, and vegetation structure such as the amount of leaves per ground area (Heisler-White et al. 2008; Reeves et al. 2014; Gao et al. 2016; Feng et al. 2019; Ji et al. 2020; Zhang et al. 2021). Atmospheric CO2 influences NPP both directly as the photosynthesis of plants with the common C3 photosynthesis partly CO2-limited, and indirectly, as many plants reduce stomatal conductance under enhanced CO2 levels, which can lead to more conservative water use (Archibald et al. 2009; Reeves et al. 2014; Xu et al. 2016). In savannas, these plant-physiological CO2 effects can lead to complex changes in vegetation dynamics and fire as plants with C4 photosynthesis, which are most savanna grasses, benefit much less from increasing CO2 than woody plants, which can lead to woody encroachment (Midgley and Bond 2015). Terrestrial NPP patterns are expected to change in the future in response to these drivers and human population dynamics, thus necessitating assessment of NPP sensitivity to climate and other environmental change (Mohamed et al. 2004; Reeves et al. 2014). The overall vegetation NPP in warm dry regions, such as most of southern Africa is commonly mostly limited by precipitation or the amount of available moisture (Nemani et al. 2003; Hickler et al. 2005; Ji et al. 2020).

Given that NPP is one of the most variable components of the terrestrial C cycle, ecologists aim to make accurate estimates of this component when conducting research on terrestrial ecosystems, C cycles, and climate change (Sala and Austin 2000; Yu et al. 2018). Answering important questions concerning the global C balance, and predictions of the effects of global climate change rely on estimates of this fundamental quantity (Sala and Austin 2000). However, directly measuring GPP and NPP in the field is close to impossible and hence researchers resort to estimating the vegetation production components (Clark et al. 2001; Chapin et al. 2011; Peng et al. 2017).

There are various ways to estimate NPP across large extents and this includes: (1) field surveys (and subsequently extrapolating field measurements for local NPP to larger regions, using a vegetation map), (2) Earth Observation-based products, which are also informed by field survey, and (3) process-based ecophysiological modelling (Ruimy et al. 1994; Zhao et al. 2005). In this chapter, we focus on the latter two. Temporal trends, particularly if shown by both of these two approaches, will highlight whether there is increase (greening) or decrease (browning) of NPP in the study region (Zhu et al. 2016).

Southern Africa is one of the regions identified as most vulnerable to climate change (Pan et al. 2015; Ranasinghe et al. 2021; Chap. 3). Precipitation is expected to decrease over the summer rainfall region of southern Africa, and with the drying effect of increased temperature will thus lead to a robust and pronounced decrease in soil moisture over the region (IPCC 2021; Chaps. 6 and 7). In the summer rainfall areas, this is due to the El Niño/Southern Oscillation phenomenon (ENSO) which is negatively correlated with the amount of rainfall during the summer season in southern Africa (Malherbe et al. 2015; Chap. 6). Furthermore, there is high confidence in projected mean precipitation decreases in west southern Africa and medium confidence in east southern Africa by the end of the twenty-first century (Ranasinghe et al. 2021; Chap. 7).

Climate change will challenge agriculture, forestry, water systems, health, and the adaptive capacity of the natural ecosystems of the region (Pan et al. 2015). Robust projections of NPP will be highly relevant to meeting these coming challenges. However, in order to produce such projections, a solid understanding of the current NPP dynamics and drivers must be established. To this end, we seek to shed light on the following research questions: (1) what are the current NPP spatial distribution and temporal trends in southern Africa? (2) how consistent are the estimates of NPP from different methods? (3) what are the drivers of the dynamics that are producing these spatial patterns and temporal trends? and (4) can process-based models give robust estimates of future NPP by capturing these dynamics? Therefore, in this chapter, we: (1) examine and compare spatial and temporal NPP patterns for southern Africa derived from a Dynamic Global Vegetation Model (DGVM) and Earth observation-based estimates and (2) investigate some possible drivers (climate variables and atmospheric CO2 concentration) of NPP in the region. (3) Furthermore, we subset our data using a well-known biome map to determine the NPP patterns and drivers for different ecosystems.

2 Materials and Methods

2.1 Study Region

The southern African region (here defined as −35.0 S to −20.0 S and 13.5 E to 35.0 E) is located on the southernmost part of the African continent consisting of several countries, namely: Angola, Botswana, Lesotho, Malawi, Mozambique, Namibia, South Africa, Swaziland, Zambia, and Zimbabwe. Southern Africa has both low-lying coastal areas, and mountains with varied terrain, ranging from forest and grasslands to deserts. Furthermore, the region has diverse ecoregions that includes grassland, bushveld, Karoo, savanna, and shrublands (Rutherford et al. 2006; Schoeman and Monadjem 2018). The climate across southern Africa varies from arid conditions in the west to humid subtropical conditions in the north and east, while much of the central part of southern Africa is classified as semi-arid (Cooper et al. 2004; Daron 2015). Despite the wide range of climate types, agriculture is a critical sector for all of the economies of southern African countries, so the effects of climate change on NPP and the knock-on effects on agricultural productivity, ecosystem service development, and food security are highly relevant across the study region (Gornall et al. 2010).

2.2 Data Sources

2.2.1 Earth Observation Data

MODIS and other Earth observation missions are essential tools for the development and evaluation of Earth system models predicting global ecosystem changes, which are an important information source for political decision-makers (Simmons et al. 2016; Chaps. 24 and 29). The use of Earth observation data allows the monitoring of different ecosystem variables (e.g. vegetation/biomass changes, surface moisture dynamics, etc.) with high spatial resolution and short temporal intervals (Gao et al. 2013). Earth observation data from various sources has become a valuable tool for analysing vegetation productivity in combination with in situ NPP estimates (Zhao et al. 2005; Fukano et al. 2021). A variety of light use efficiency (LUE) models have been developed (Running et al. 2004; Zhang et al. 2015) to calculate GPP from measured absorbed photosynthetically active radiation (APAR) (Xiao et al. 2019). Earth observation data have also been integrated with machine learning approaches (Xiao et al. 2008; Jung et al. 2009) and process-based models (Hazarika et al. 2005; Liu et al. 2019) for quantifying C fluxes (e.g. GPP and NPP). Autotrophic respiration can be estimated using modelling approaches that use daily climate variables and estimated biomass, and this can then be subtracted from GPP to derive NPP (Clark et al. 2001; Ardö 2015). In addition to estimates of these fluxes, the so-called vegetation indices (formed by combining two or more spectral bands) have been developed to characterise different aspects of vegetation from Earth observation data (Masoudi et al. 2018). Studies have shown the normalised difference vegetation index (NDVI) to be a good proxy for NPP at high spatial resolution (Zhao et al. 2005; Pachavo and Murwira 2014; Cui et al. 2016). NDVI is expressed as:

$$ \textrm{NDVI}=\frac{\left(\textrm{NIR}-\textrm{RED}\right)}{\left(\textrm{NIR}+\textrm{RED}\right)}\kern0.5em $$
(26.1)

where NIR and RED are reflectance values in the near-infrared and red wavelengths, respectively (Tucker 1979). NDVI values range from +1.0 to −1.0, where negative values may be representative of cloudy conditions or areas over water bodies. Areas of barren rock, sand, or snow show very low NDVI values of 0.1–0; sparse vegetation such as shrubs and grasslands or senescing crops may show moderate NDVI values of 0.2–0.5; and high NDVI values of 0.6–0.9 correspond to dense vegetation found in temperate and tropical forests or crops at their peak growth stage (Higginbottom and Symeonakis 2014).

2.2.1.1 Earth Observation Platforms and Products

This study primarily utilised time series information from MODIS sensors onboard the satellites TERRA and AQUA (Minnett 2001; Yang et al. 2006; Cao 2020). Both platforms have sun-synchronous orbits with a revisit time of 1–2 days (Savtchenko et al. 2004). Data is acquired in 36 spectral bands with wavelengths ranging from 0.4 to 14.385 μm (Cao 2020).

For this study, NPP was taken from the MOD17A3 (UM Collection 5) annual totals. The MOD17 products are the first MODIS operational data sets to regularly monitor global vegetation productivity (Zhao et al. 2005; Yu et al. 2018). Details of the MODIS NPP derivations are provided in the section below.

NDVI estimates were taken from the MOD13C2 Collection 6, 16-day product (at monthly intervals) which is at 1 km spatial resolution and is provided in 0.05 degree geographic climate modelling tiles (Solano et al. 2010). The 16-day product was processed to mean annual values in R. This dataset has been used for modelling global biogeochemical and hydrologic processes in both global and regional climates (Didan 2015). Furthermore, the data have been utilised in studies to characterise land surface biophysical properties and processes, including primary production and land cover conversion (Solano et al. 2010; Didan 2015).

In addition, NDVI time series from the Advanced Very-High-Resolution Radiometer (AVHRR) sensor on board of the National Oceanic and AtmosphericAdministration's (NOAA) polar-orbiting satellites was used to investigate NDVI long-term trends. Mounted on a polar-orbiting satellite it acquires images of the visible, near-infrared, and thermal infrared parts of the electromagnetic spectrum (Kidwell 1995; Sus et al. 2018). The sensor has a spatial resolution of approximately 1.1 km at satellite nadir and covers the time period from 1981 to 2015 (Trishchenko et al. 2002). NOAA AVHRR is a widely used sensor due to its long-term monitoring period and retrieval of various land surface parameters such as land cover/use dynamics, NDVI, Land Surface Temperature (LST), and Albedo (Forkel et al. 2013; Wang et al. 2020; Gulev et al. 2021; Urban et al. 2013).

2.2.1.2 MODIS NPP Derivation

The MOD17 algorithm is based on the original LUE logic of Monteith (1972). Input data for the model include climatic variables such as temperature, solar radiation, vapour pressure deficit (VPD) from meteorology dataset from NASA global modelling and assimilation office (Running et al. 2004). MOD15 leaf area index and fraction of absorbed photosynthetically active radiation (FAPAR) products are also utilised (Running et al. 2004; Zhao et al. 2005). Land cover classification from MODIS MCD12Q1 data product is used (Running et al. 2004). A Biome Parameter Lookup Table (BPLUT) containing values of ε_max (Eq. 26.2) was derived and later updated (Running et al. 2004; Zhao et al. 2005). The table contains different vegetation types, temperature, and VPD limits and other biome-specific physiological parameters for respiration calculations (Running et al. 2004). The different vegetation types obtained from the land cover type 2 classification include: evergreen needleleaf forest, evergreen broadleaf forest, deciduous needleleaf forest, deciduous broadleaf forest, mixed forests, closed shrublands, open shrublands, woody savannas, savannas, grasslands, and croplands. Environmental multipliers represent limitations by low temperature and high VPD, and autotrophic respiration is estimated with a Q10 relationship (Zhao et al. 2005; Ardö 2015). The MOD17 algorithm calculates daily GPP as:

$$ \textrm{GPP}={\varepsilon}_{\textrm{max}}\times 0.45\times {\textrm{SW}}_{\textrm{rad}}\times \textrm{FPAR}\times f\left(\textrm{VPD}\right)\times f\left({T}_{\textrm{min}}\right)\kern0.5em $$
(26.2)

where εmax is the maximal, biome-specific light use efficiency (g C MJ−1), SWrad is incoming short-wave radiation [assuming 45% to be photosynthetic active radiation (PAR)], FPAR is the fraction of photosynthetically active radiation, f(VPD) and f(Tmin) are linear scalars reducing GPP due to water and temperature stress (Ardö 2015). The model estimation utilises the fact that GPP is closely related to the APAR and that APAR can be measured continuously using Earth observation sensors (e.g. MODIS) (Cui et al. 2016; Xiao et al. 2019). The FPAR is estimated as a function of NDVI, derived from the standard MODIS land product (MOD15) (Eqs. 26.3 and 26.4) (Running et al. 2004; Zhao et al. 2005; Gonsamo and Chen 2017).

$$ \frac{\textrm{APAR}}{\textrm{PAR}}=\textrm{NDVI}\kern0.5em $$
(26.3)
$$ \textrm{FPAR}=\frac{\textrm{APAR}}{\textrm{PAR}}=\textrm{NDVI}\kern0.5em $$
(26.4)
$$ \textrm{GPP}={\varepsilon}_{\textrm{max}}\times \textrm{PAR}\times \textrm{NDVI}\times f\left(\textrm{VPD}\right)\times f\left({T}_{\textrm{min}}\right) \vspace*{-6pt}$$
(26.5)
$$ \textrm{PsnNet}=\textrm{GPP}-{R}_{\textrm{ml}}-{R}_{\textrm{mr}}\kern0.5em $$
(26.6)

NPP is calculated annually as:

$$ \textrm{NPP}={\sum}_{i=1}^{365}\textrm{PsnNet}-\left({R}_{\textrm{mo}}+{R}_{\textrm{g}}\right)\kern0.5em $$
(26.7)

where PsnNet is the maintenance respiration by leaves (Rml) and fine roots (Rmr) and is calculated daily. Rmo is the annual maintenance respiration by all other living parts except leaves and fine roots, Rg is the annual growth respiration (Zhao et al. 2005).

The products were projected to a geographic grid while resampling to 0.5 degree using “average” resampling type in order to match the resolution of the climate input data used to drive our DGVM. Grid cells without valid MOD17 NPP (MOD12Q1 land cover barren, water, or urban) were masked out from the LPJ-GUESS data in order to make the data sets comparable with identical spatial extent, land cover classes, and number of grid cells. The datasets were aggregated (annual sums for NPP, annual means for NDVI) to a 14-year annual time series from 2002 to 2015.

2.2.2 Dynamic Vegetation Models

A class of ecosystem models known as dynamic global vegetation models (DGVMs) have been developed to simulate vegetation dynamics and biogeochemical cycling either at regional or global scales (Prentice et al. 2007; Sitch et al. 2008; Smith et al. 2014). DGVMs represent basic ecophysiological processes, such as photosynthesis, plant and soil respiration, C allocation, and plant growth, competition between plant types for resources (commonly light and water, increasingly also nutrients) and disturbances such as fire (Ardö 2015; Hantson et al. 2016). Simulating the impacts of climate change on ecosystems and feedback from ecosystems on climate, in particular via the terrestrial carbon cycle, has been a research priority (Prentice et al. 2007; Kelley et al. 2013). Subsequently, the representation of land-use has also received attention and has been integrated into DGVMs (Bondeau et al. 2007; Lindeskog et al. 2013; Pugh et al. 2019; Drüke et al. 2021).

DGVM’s largest potential lies in process-understanding (Hickler et al. 2005) rather than short-term predictions, which can be even more accurate with empirical approaches. DGVMs are powerful tools to quantify spatial and temporal variations in ecosystem C fluxes and to analyse the underlying mechanisms of NPP at large scales (Tao et al. 2003; Yu et al. 2018). DGVMs have the potential to accurately explain how ecosystem processes will interact in future climatic conditions, CO2 concentration, nitrogen deposition, land-use changes, and soil conditions (Melillo et al. 1993; Luo et al. 2004; Ardö 2015). DGVMs are driven with climate and other environmental data (e.g. soil properties and nutrient deposition) which can be either historical (observed) data sets or projections of past/future environmental conditions.

2.2.2.1 LPJ-GUESS Model and Setup

The Lund–Potsdam–Jena (LPJ) model has been developed as a process-based DGVM which can efficiently represent the land–atmosphere interaction and potentially be applied for broader global problems (Gerten et al. 2004; Sitch et al. 2003). The Lund–Potsdam–Jena General Ecosystem Simulator (LPJ-GUESS) framework was originally developed to add a more detailed representation of vegetation dynamics through a “forest-gap model” to the LPJ DGVM (Smith et al. 2001). Thus, LPJ-GUESS is an individual (or cohort) based model which combines biogeography and biogeochemistry typical of a DGVM with a comparatively more detailed individual and patch-based plant functional type (PFT) representation of vegetation structure, demography, growth, mortality, reproduction, carbon allocation, and resource competition (Smith 2001; Sitch et al. 2003). The model now includes an interactive nitrogen cycle (Smith et al. 2014), which can limit photosynthesis and is so important to constrain future potential CO2 fertilisation effects (Hickler et al. 2015), and a representation of agricultural land and management (Lindeskog et al. 2013).

In the framework, productivity is simulated as the emergent outcome of growth and competition for light, space, and soil resources among woody plant individuals and a herbaceous understory in a number of a replicate patches (typically 15–50) representing “random samples” of each simulated locality or grid cell (Smith 2001). Natural, cropland, and pasture land cover types are distinguished and their fractional covers are prescribed from the dataset by Hurtt et al. (2011). Within the cropland land cover type, crop fractions from the MIRCA database (Portmann et al. 2010) are used and nitrogen fertiliser application rates from Zaehle et al. (2010) are prescribed.

In this study, we used the standard version of the cohort-based LPJ-GUESS model using 20 replicate patches at a spatial resolution of 0.5° × 0.5°. Climate forcing data from the CRU JRA v2.0 dataset (details below) were used. Land-use dataset by Hurtt et al. (2011) was included. The global plant functional types (PFTs) analysed in our model were: boreal needleleaf evergreen (BNE), boreal shade-intolerant needleleaf evergreen (BINE), boreal needleleaf summergreen (BNS), temperate needleleaf evergreen (TeNE), temperate broadleaf summergreen (TeBS), shade-intolerant broadleaf summergreen (IBS), temperate broadleaf evergreen (TeBE), tropical broadleaf evergreen (TrBE), tropical shade-intolerant broadleaf evergreen (TrIBE), tropical broadleaf raingreen (TrBR), C3 grasses (C3G), C4 grasses (C4G). The model output includes GPP, NPP (kg C m−2 year−1), respiration, carbon pools, burnt area fraction, and potential vegetation among other outputs. The model was run on a daily time step. All simulations were initialised with a 500 years spinup to allow vegetation, soil carbon and nitrogen pools to build up from “bare ground” to a “steady state” and then the full transient period (1901–2018) was simulated. Fire was enabled through the SIMFIRE-BLAZE fire model (Knorr et al. 2016; Nieradzik et al. 2015).

2.2.3 Meteorological Data

Precipitation, temperature, and solar radiation data from the CRU JRA v2.0 dataset were used for both driving the LPJ-GUESS simulations and for investigating the correlation between NPP and its potential drivers. As a basis, this dataset starts with the Japanese reanalysis (JRA) (Harada et al. 2016; Kobayashi et al. 2015) data produced by the Japanese Meteorological Agency (JMA). This is then adjusted, where possible, to align with the monthly values of the CRU TS 3.26 dataset (Harris and Jones 2019), a gridded land surface dataset based on meteorological station data produced by the Climatic Research Unit (CRU). The data availability spans from January 1901 to December 2017. The dataset is a 6-hourly, gridded time series of ten meteorological variables and is intended to be used to drive models of the global land surface and biosphere such as DGVMs. The variables are provided on a 0.5 × 0.5 degree grid.

Fig. 26.1
A biome map of SANBI plots latitude versus longitude. It represents the regions of Albany Thicket, desert, fynbos, grassland, Indian Ocean coastal belt, nama-karoo, savanna, and succulent karoo.

South African National Biodiversity Institute (SANBI) 2006 biome map by Rutherford et al. (2006)

Full size image

2.3 Data Analysis

NDVI time series from MODIS and AVHRR products along with MODIS NPP products were used to assess the vegetation productivity. NPP was simulated to assess whether the LPJ-GUESS model agrees with MODIS and AVHRR estimates by following the inter-annual variation of remotely estimated NPP and NDVI. This was to evaluate the model’s capability to reproduce past data and ultimately adopted for future predictions in southern Africa.

The potential driving factors and trend analysis of NPP and NDVI were conducted per biome, according to the South African National Biodiversity Institute (SANBI) 2006 biome map (Fig. 26.1) (Rutherford et al. 2006). The biome map is made up of nine well-established biomes in South Africa which includes: Savanna, Grassland, Nama Karoo, Succulent Karoo, Fynbos, Albany Thicket, Forest, Indian Ocean Coastal Belt (IOCB), and Desert (Rutherford et al. 2006). This was to show the ecological and climatic variability experienced across the South African region. Although the study area covers southern Africa and not just South Africa, the SANBI biome was used because it has been studied extensively. Furthermore, the study focuses on the dynamics and drivers of NPP per biome where actually the South African biomes extend into the upper regions of the study area.

2.3.1 Analysis Software

The R statistical programming language was used for processing and for statistical analysis of the data. MODIS R package (Mattiuzzi et al. 2017) was used to download and process MODIS data. We used the DGVMTools R packageFootnote 1 to perform comparisons, analysis, and plotting of the spatially explicit simulated and remotely sensed NPP distribution across the study region. DGVMTools is a high-level framework for processing, analysing, and visualising DGVM data output which easily interfaces with both the raster package and base R functionality. The ggplot2 package (Wickham 2016) was used for additional plotting and linear trend analysis.

2.3.2 Time Series Analysis

The NDVI and NPP time series were analysed along with the key climate variables (precipitation, temperature, and solar radiation) over the period 2002–2015. The region experienced extreme rainfall events in the year 2000 which were far outside its normal variability (Dyson and van Heerden 2001; Smithers et al. 2001). Accordingly, the years 2000 and 2001 were excluded from the analysis since including them was found to produce spurious trends and thus produced misleading analysis and conclusions. The AVHRR NDVI time series were analysed from 1982 to 2015 in order to gain some perspective of the longer NDVI trend. As these variables have different units and magnitudes, we derived aggregated standardised anomalies following Seaquist et al. (2008) and this is expressed as:

$$ \textrm{Standardised}\ \textrm{anomaly}=\frac{\Big(\boldsymbol{x}-\underset{\_}{\boldsymbol{x}\Big)}}{\textrm{sd}} $$
(26.8)

where x is the annual mean values, \( \underset{\_}{x} \) is the mean of the annual mean values, and sd is the standard deviation.

Linear trends were fitted to NDVI and NPP time series to determine the productivity long-term trend (Higginbottom and Symeonakis 2014). Due to the large inter-annual variation and relatively short time series of our data, most trends are not statistically significant. The Mann-Kendall test (Mann 1945) was used to quantify the significance and only p < 0.05 was considered. We also analysed the response of LPJ-GUESS simulated NPP to atmospheric CO2 concentration by comparing the standard LPJ-GUESS simulation to one with atmospheric CO2 concentration fixed from 1901 at its corresponding value of 296.4 ppm.

2.3.3 Statistical Correlation Analyses

We used Pearson’s product moment correlation coefficient (r) to quantify the level of agreement between LPJ-GUESS and MODIS NPP. To examine the relationship between different variables (NPP, NDVI, CO2, and different climate factors) Spearman’s rank correlation coefficient was used. This is because we are interested in monotonic relationships between these variables which do not necessarily need to be linear. Spearman’s coefficient is insensitive to any nonlinearity that could undermine the detection of a monotonic relationship between the variables. The coefficient ranges from −1 to +1 and was examined using the “PerformanceAnalytics” R package (Peterson et al. 2020). Large positive values indicate strong agreement, large negative values indicate strong disagreement, and values near 0 indicate random agreement (Seaquist et al. 2008; Bon-Gang 2018).

Fig. 26.2
2 color-coded maps, labeled L P J GUESS on the left and MODIS on the right, illustrate the net primary production, with nearly identical values.

Spatial distribution of modelled NPP and remotely sensed NPP averaged over time (2002–2015) to evaluate spatial patterns of NPP over the southern African region. The plot presents a strong linear relationship between LPJ-GUESS and MODIS of r = 0.85 and RMSE = 0.1446 (kg C m−2 year−1)

Full size image

3 Results

3.1 NPP Geographical Patterns

The broad spatial patterns of NPP are as expected, with higher NPP in regions of higher rainfall and lower NPP in areas that experience extreme heat and receive less rainfall. In general, NPP simulated by LPJ-GUESS and estimated by MODIS showed similar spatial patterns, with Pearson’s r = 0.85 and Root Mean Squared Error (RMSE = 0.1446 kg C m−2 year−1). The main difference is lower LPJ-GUESS simulated NPP values along the coastal regions (Fig. 26.2). Contrastingly, LPJ-GUESS showed higher NPP than MODIS as one moves in from the coast. The inter-annual dynamics of NPP on a gridcell level correspond reasonably well. This is indicated by Pearson’s correlation coefficients of the time series of the individual grid cells with high statistical significance (p < 0.05) over most parts of the study region (Fig. 26.3).

Fig. 26.3
Two color-coded maps plot latitude and longitude, with one map displaying Pearson correlation coefficients and the other exhibiting P-values.

Pearson correlation and significance level of modelled NPP and Earth observation-based NPP per gridcell, averaged over time (2002–2015). The plot presents a high correlation of LPJ-GUESS and MODIS NPP over most parts of the region with great significance level (p < 0.05) within the correlated areas

Full size image
Fig. 26.4
A line graph plots N P P versus year. It illustrates fluctuating trends for L P J GUESS and MODIS. L P J GUESS has the highest value.

Temporal distribution of LPJ-GUESS modelled and MODIS-estimated NPP (kg C m−2 year−1) averaged over the whole southern Africa study region

Full size image
Fig. 26.5
9 line graphs plot productivity measures versus year. The graphs are labeled desert, succulent karoo, nama-karoo, fynbos, Albany Thicket, grassland, savanna, forest, and the Indian Ocean coastal belt. The lines represent L P J GUESS N P P, MODIS N D V I, and MODIS N P P.

Absolute values and linear trends of NPP (kg C m−2 year−1) and NDVI (unitless) per biome

Full size image

3.2 NPP Temporal Development

When averaged over the whole study region, LPJ-GUESS simulated and MODIS-estimated NPP showed good agreement in terms of the overall magnitude and inter-annual variation (Fig. 26.4). The inter-annual variation was found to be large, annual values ranged from a maximum of over 0.5 kg C m−2 year−1 in 2006, to less than 0.35 kg C m−2 year−1 (minima in 2003 and 2015). The time series also revealed a small tendency of the LPJ-GUESS model to simulate higher values than estimated with MODIS. However, when applying the Mann-Kendall test the trends showed no significance.

Considering NPP times series per biome (Fig. 26.1; Chap. 3) gives a more nuanced view of the regional disparities between LPJ-GUESS and MODIS (Fig. 26.5). The grass-dominated biomes (Grassland and Savanna) showed good agreement in NPP magnitude between LPJ-GUESS and MODIS, the tree-dominated biomes (Forest, Albany Thicket, and Indian Ocean Coastal Belt) and Fynbos showed consistently higher NPP in MODIS than LPJ-GUESS. This pattern is reversed for the most arid biomes (Nama Karoo, Succulent Karoo, and Desert) where LPJ-GUESS simulates higher NPP. Both LPJ-GUESS and MODIS NPP showed a similar increasing but weak trend for the majority of the biomes (Fig. 26.5). However, on a per grid cell basis, the majority of the region shows no significant trends for both the LPJ-GUESS and MODIS NPP (Fig. 26.6b).

Fig. 26.6
Four color-coded maps display the trend in N P P and the significance in N P P trends in L P J GUESS and MODIS datasets.

(a) Trend in modelled and Earth observation NPP. (b) The significance in trends for both modelled and Earth observation NPP. The majority of the region shows no significant trends for both products

Full size image

Examination of simulated LPJ-GUESS NPP over a longer period (1901–2015) showed high inter-annual variation against a backdrop of increasing NPP (Fig. 26.7b) caused by rising atmospheric CO2 (Fig. 26.7a). According to the model, NPP has increased by 18% since 1901. This is in agreement to the increasing trend for 2002–2015 (Fig. 26.4). This longer term trend in NPP is significant (Mann-Kendall p < 0.00005).

Fig. 26.7
2 line graphs plot N P P and standard anomalies versus year. Graph 1 exhibits a stable trend for L P J-GUESS fixed C O 2 and an increasing trend for L P J-GUESS standard C O 2. Graph B displays increasing trends for A V H R R N D V I and L P J-GUESS N P P.

(a) Annual average NPP (kg C m−2 year−1) across the whole study region simulated by LPJ-GUESS with historically varying atmospheric CO2 concentration and CO2 fixed at values from the year 1901 (296.3785 ppm) and (b) Aggregated standardised anomalies of historical LPJ-GUESS NPP (1901–2015) and AVHRR NDVI (1982–2015). This plot evaluates the trends and Inter-annual variability (IAV) of NPP over a longer time span

Full size image

The longer time series of AVHRR data averaged over the whole study area showed a strong increase in NDVI from 1982 to 2015 and a Mann-Kendall test of trend significance revealed high statistical significance with a p < 0.0048 (Fig. 26.7b). Comparison of NPP and NDVI showed increasing trends (Figs. 26.5 and 26.7b) for all biomes except for decreasing trends in the desert and flat trends in the IOCB for MODIS NDVI and NPP. However, the trends showed no significance for the Mann-Kendall test except for MODIS NPP in the Fynbos (p = 0.037) and forest (p = 0.016) (data not shown). The decadal scale fluctuations were apparent in the longer term time series with some periods of increasing and declining NPP, thus indicating that the increasing NPP observed for 2002–2015 is not necessarily a new phenomenon or a result of global change, but could simply be a consequence of decadal scale climate variability.

3.3 NPP and NDVI Correlations

The LPJ-GUESS and MODIS NPP showed high Pearson correlation coefficient (ranges from 0.68 to 0.82) throughout the biomes except for Fynbos (0.37) (Table 26.1). The correlations between the two products showed significance of p < 0.001 while Fynbos showed no significance. Succulent Karoo (0.49), Albany Thicket (0.50), and IOCB (0.16) showed no significance when comparing LPJ-GUESS NPP and MODIS NDVI while high and significant (p < 0.001) correlation was observed for the rest of the biomes (ranges from 0.64 to 0.91). The IOCB biome (0.36) showed low and not significant correlation when comparing MODIS NPP and NDVI, while the rest of the biomes showed significantly high correlations (ranges from 0.61 to 0.93).

Table 26.1 Spearman rank correlation coefficient for aggregated standardised anomalies (MODIS NDVI, MODIS NPP, LPJ-GUESS NPP, precipitation, temperature, and solar radiation) per biome, from 2002 to 2015. The Pearson correlation was applied for the LPJ-GUESS and MODIS NPP comparisons. The table shows the value of the correlation plus the significance level as stars. Each significance level is associated to a symbol: p-values (0, 0.001, 0.01, 0.05, 0.1, 1) refers to symbols (***, **, *, †,, "") respectively
Full size table

3.4 Potential Driving Factors of NPP

Precipitation showed similar inter-annual variation to that of LPJ-GUESS and MODIS NPP per biome, as demonstrated by significantly high correlations with both MODIS (0.56–0.95) and LPJ-GUESS NPP (0.81–0.95) (Fig. 26.8 and Table 26.1). However, there were two notable exceptions to this: the correlations between MODIS NPP and precipitation in the Desert (0.42) and Fynbos (0.42) biomes were low and not significant. In contrast, solar radiation and temperature showed negative correlations with MODIS and LPJ-GUESS NPP across all biomes, although they were not significant in most cases (Table 26.1). Of all the climate drivers, temperature showed a significantly increasing trend with the Mann-Kendall p < 0.049 for Succulent Karoo and 0.009 for Desert while the rest were not significant. This is consistent with global warming in areas of high inter-annual variability and relatively short time series (as is the case here).

Fig. 26.8
9 line graphs plot standardized anomalies versus year. The graphs are labeled desert, succulent karoo, nama-karoo, fynbos, Albany Thicket, grassland, savanna, forest, and the Indian Ocean coastal belt.

Aggregated standardised anomalies of precipitation, temperature, and solar radiation in comparison with both the modelled and Earth observed NPP per biome

Full size image

The model simulation in which atmospheric CO2 concentration was held constant at 1901 levels (Fig. 26.7a) showed that increasing CO2 concentration is responsible for a large increase in NPP. This effect was built to be around +18% (+0.064 kg C m−2 year−1) by 2015. LPJ-GUESS NPP showed a slight decline without significant trend when the effects of increasing atmospheric CO2 were removed.

4 Discussion

4.1 Comparison of NPP from Modelling and Earth Observation Products

The DGVM LPJ-GUESS and MODIS NPP products show similar spatial patterns, temporal trends, inter-annual variability, and good correlation (Figs. 26.2, 26.3, and 26.4). The good correspondence and correlation is encouraging as both methods are largely independent and only share some common input variables (e.g. daily precipitation and temperature), but from different data sets. The correspondence between MODIS17 and LPJ-GUESS here appears to be slightly better than in comparison with an earlier LPJ-GUESS version applied to simulate all of Africa by Ardö (2015), which did not include land-use and nitrogen limitation to vegetation growth and had a different fire module.

Both the MODIS and LPJ-GUESS estimates are to a greater or lesser extent model derived estimates and so subject to uncertainty in the parameters used, the input data and, in the case of LPJ-GUESS, uncertainty in the process representations. The MODIS estimate might be expected to be the more reliable as it uses Earth observation FPAR as an input (which means it is more constrained) (Myneni et al. 1999, 2002; Running et al. 2004), while FPAR in LPJ-GUESS emerges from complex equations that govern the growth of and competition between plant functional types (PFTs) and disturbances (Smith 2001; Sitch et al. 2003). It should also be noted that the MODIS NPP algorithm has been criticised, e.g., for its assumed temperature effects (Medlyn 2011), and the representation of environmental controls of NPP is not very sophisticated, see above. However, given that MODIS NPP uses Earth observation FPAR and so is in principle better constrained, we can interpret the comparison of LPJ-GUESS with MODIS as an evaluation of LPJ-GUESS to some extent. As we find good correspondence between the two, we can conclude that past and future projections for southern Africa are also reliable.

LPJ-GUESS is commonly used to derive future projections to guide climate adaptation and mitigation measures, so this result is promising and indicates that LPJ-GUESS can be fruitfully applied to the study area. As a small caveat, it should be noted that the per biome comparisons between MODIS and LPJ-GUESS did show some disparities in the more arid and the more tree-dominated biomes. This is not surprising as we applied the standard global LPJ-GUESS configuration here and a global model cannot be expected to reproduce all the features of a diversely vegetated region such as southern Africa. Future studies would benefit from using a regionally parameterised version of LPJ-GUESS, in particular the inclusion of shrub PFTs to better represent the arid regions.

All analysed time series products (LPJ-GUESS and MODIS NPP, and MODIS and AVHRR NDVI) show increasing trends for all the biomes except for the desert which showed a decreasing trend and grasslands which showed no trend (Figs. 26.5 and 26.7b). One should note that the inter-annual variability for our study region is large compared to the magnitudes of these trends and so, when tested, these trends did not show statistical significance. This is consistent with other results that have pointed out that vegetation trends are mostly not statistically significant for our study region and the MODIS operational period (Samanta et al. 2011; Cortés et al. 2021). However, an examination of the trends plotted spatially on a per gridcell basis (Fig. 26.6a) shows regional hotspots of statistical significant greening and browning. Furthermore it should be noted that even within a particular biome, different regions can show opposite trends (consider, for example, LPJ-GUESS in the western and eastern parts of the Fynbos biome in Fig. 26.6a) which is not apparent from the biome averaged results (Fig. 26.4). The LPJ-GUESS derived NPP over the whole southern Africa region shows a statistically significant rise since 1901 but with high IAV and decadal scale fluctuations due to internal climate variability or oscillations. Therefore, we can conclude the future NPP trajectory will be governed by both climate variability and long-term trends resulting from changing drivers due to global change.

4.2 Drivers of NPP Patterns and Trends

The spatial pattern of NPP, shorter-term temporal trends, and inter-annual variability appear to be mainly driven by gradients or changes in precipitation (Figs. 26.2, 26.4, and 26.8). This can be seen, for example, with the high NPP occurring along the eastern coastal parts of the region (Fig. 26.2). This region receives the highest rainfall while the western and central part of the region experiences high temperatures and less rainfall (Botai et al. 2018). As outlined in the introduction, we expected this dominant role of precipitation as a driver of NPP variations. This is in agreement with the study conducted by Zhu et al. 2016, where they concluded that the greening in South Africa over the period from 1982 to 2009 is primarily driven by increasing precipitation. Recent climate projections show decreasing precipitation trends in southern Africa both for low and high warming scenarios, with corresponding increases in aridity and drought (IPCC 2021). Higher levels of warming result in stronger precipitation decreases, with the highest percentage decreases projected in the west of the region, which is already arid (IPCC 2021, Fig. SPM.5c). Our results indicate that, up to 2015, precipitation decreases have not yet caused widespread negative impacts on NPP as we can see slightly increasing trends in Figs. 26.4 and 26.7a, although there are some regional exceptions (Fig. 26.6a). The potential effects of future drying on NPP and the consequences for ecosystems, ecosystem services, agriculture, and forestry remain an open question. To tackle this question, predictive methods such as DGVMs, with more detailed or regionally parameterized representations of land-use, or more empirical modelling approaches must be applied in combination with future climate projections and, ideally, informed by experimental work.

The observed high negative correlation with solar radiation is likely due to the fact that precipitation is associated with cloud cover and therefore decreased solar radiation. This indicates that whilst solar radiation is essential for photosynthesis and therefore NPP, it is not a limiting factor to the same degree as precipitation and so is in fact not a driver of NPP in the study region (Nemani et al. 2003). The same was found by Hickler et al. (2005) (Fig. 3) for the Sahel region.

In our study region, NPP shows a negative correlation with temperature. This is consistent with the fact that most of the study region is in the subtropical to warm temperate climate zones, so photosynthesis is unlikely to be temperature limited (not applicable for winter rain regions in the south west) and no positive correlation is expected (Nemani et al. 2003). Instead, this negative correlation may occur due to a combination of increased autotrophic respiration rates that accompany higher temperatures, a net reduction of photosynthesis due to decreased efficiency and heat stress at higher temperatures, and decreased soil water availability due to stronger atmospheric moisture demand. Among all the climatic variables analysed, only temperature in the Succulent Karoo and Desert biome showed statistically significant trends (increasing) from 2002 to 2015. Taking these findings together and putting them in the context of global change, it is likely that temperature will have an escalating and negative effect on NPP in these biomes in the coming decades, at least under moderate warming without unprecedented heat stress for the vegetation, which is not represented in the applied model version.

According to LPJ-GUESS, the plant-physiological effects of increasing atmospheric CO2 have been large and positive between 1901 and 2015 (Fig. 26.7a). However, the simulated CO2 effects over the last decades are only minor in similar arid environments (Hickler et al. 2005). The strong effects since the beginning of the last century are consistent with other global estimates. Ciais et al. (2012) estimated a pre-industrial GPP of 80 Pg year−1, which is substantially lower than the present estimate of about 120 Pg (Ciais et al., 2013; Friedlingstein et al., 2020). Using another independent method, Campbell et al. (2017) estimated a GPP increase of about 31% since the early last century. GPP and NPP are strongly related, NPP being about half of GPP (Ciais et al., 2013; Yu et al., 2018), which also applies to LPJ-GUESS (Ardö 2015). The CO2 effect is also in line with a global greening trend attribution study using different leaf area index (LAI) estimates and ten DGVMs (Zhu et al. 2016). These authors concluded that CO2 fertilisation explains 70% of the observed greening trend between 1982 and 2009, particularly in the tropics. Their analyses, however, suggested other major drivers for most of southern Africa, namely climate change in the north-eastern part of our study region and land cover change in central South Africa (Zhu et al. 2016; Fig. 26.3c). In our study region, CO2 effects have been found to be minor under extreme drought stress when meristem growth is strongly limited by leaf turgor (Xu et al., 2016). Moreover, future CO2 fertilisation effects might be smaller as the CO2 limitation of photosynthesis increasingly Saturates and because of progressive nutrient limitation (Hickler et al. 2015; Wang et al. 2020).

5 Conclusion

The two rather independent approaches to estimate spatial patterns and recent past trends in NPP we used (Earth-observation-based estimations from MODIS and the LPJ-GUESS DGVM) provided similar results. This suggests that the spatio-temporal results presented here are robust. As the MODIS estimate is more constrained by observed data, this result may also be considered as an evaluation of the LPJ-GUESS model. Thus, the model should be suitable for future projections in southern Africa in order to inform climate adaptation and mitigation measures.

Precipitation change is the most important driver of NPP dynamics in the study region, particularly of the inter-annual variations and spatial distribution. However, increasing atmospheric CO2 has had a large effect since the early twentieth century and might continue to play a significant role as anthropogenic CO2 emissions continue. Solar radiation and temperature were not found to be significant drivers of NPP in the study region.

Given the importance of precipitation in southern Africa and its projected decline, it is likely that ecosystems and agriculture will be under increasing pressure as their foundational building block, NPP, might decrease in areas where precipitation strongly decreases. The findings presented here alert us of the importance of mitigation and adaptation strategies going into the future and the challenges global change will bring. Both simulation models and Earth observation data can play an important role in meeting these challenges.