Modelling the impact of climate change on Tanzanian forests

Climate change is pressing extra strain on the already degraded forest ecosystem in Tanzania. However, it is mostly unknown how climate change will affect the distribution of forests in the future. We aimed to model the impacts of climate change on natural forests to help inform national‐level conservation and mitigation strategies.


| INTRODUC TI ON
Tropical forests form the most abundant terrestrial reservoir of carbon storage and biodiversity (Newmark, 2006), but have experienced climate change impact, deforestation and habitat fragmentation (Bonan, 2008;Gibbs et al., 2010). The projected increase in global mean temperature of 4.3 ± 0.7 0 C by 2,100 for RCP8.5 is likely to affect further the geographical distribution, composition and productivity of tropical forest ecosystems (IPCC, 2014) adversely affecting vital ecosystem services. Sub-Saharan Africa has been identified as one of the most vulnerable parts of the world to the effects of climate change (Chidumayo, Okali, Kowero, & Larwanou, 2011;Serdeczny et al., 2016). Climate change is predicted to increase hazards such as flood and fire hazard, disease, food insecurity and habitat degradation (Serdeczny et al., 2016).
The effects of climate change on African tropical forest habitats mostly results from changes in precipitation patterns (particularly the influence of the El Niño-Southern Oscillation (ENSO)) (Butt et al., 2015) and the Subtropical Indian Ocean Dipole (SIOD)) and subsequent effects on soils and groundwater availability (Müller, Waha, Bondeau, & Heinke, 2014), alongside increases in atmospheric availability of CO 2 concentration and nitrogen deposition (Serdeczny et al., 2016). Even though the effect of climate changes has already been felt, its impact on the tropical forests remains relatively understudied (Delire, Ngomanda, & Jolly, 2008;Markham, 1998;Pacifici et al., 2015). This is mainly in resource-poor sub-Saharan Africa, where data are scarce, creating a barrier to incorporating climate change scenarios into land management and conservation planning (Lee & Jetz, 2008).
Increasingly, global initiatives and commitments are considering African tropical forests as critical components of climate change mitigation strategies such as the Bonn Challenge on Forest landscape restoration (FLR) , the United Nations Framework Convention on Climate Change (UNFCCC) on reducing emissions from deforestation and forest degradation (REDD+) (Romijn, Herold, Kooistra, Murdiyarso, & Verchot, 2012), the Rio + 20 land degradation neutrality (Grainger, 2015), Aichi Target 15 on the restoration of degraded ecosystems (Tobón et al., 2017) and the 2030 agenda of the United Nations for Sustainable development goals (SDGs) 13 and 15 (Swamy, Drazen, Johnson, & Bukoski, 2017). To ensure that these strategies are successful and enable effective conservation, it is essential to establish a baseline in terms of forest habitat extent and resilience to climate change pressures (Clark, Gelfand, Woodall, & Zhu, 2014;Verdone & Seidl, 2017). This should be determined in a scalable and tractable manner, including modelling projections of future distributions to bridge the gap of data deficiency regarding sub-Saharan forests (Montagnini & Jordan, 2005).
Habitat suitability modelling (hereafter referred to as HSM) or species distribution modelling is widely applied in estimating changes in habitat suitability and counteract negative impacts of climate change (Edenius & Mikusiński, 2006;Lim et al., 2018;Title & Bemmels, 2017). It represents a valuable tool for informing policy-makers about the effects of climate change on forest community . HSM focuses at identifying both the most influential environmental and climatic variables describing presence/ absences, abundances or even growing conditions of forest species and the optimal relationships between their distributions and these explanatory variables (Jiménez-Alfaro et al., 2018). The provision of environmental and climatic variables from globally, often freely, available Earth Observation (EO) datasets enables simulations and subsequent information to be determined over scales that are suitable for national, regional and even global decision-making (Edenius & Mikusiński, 2006).
Maximum entropy (MaxEnt) modelling (Phillips, Anderson, & Schapire, 2006;Renner & Warton, 2013) has been used to successfully predict forest species habitat suitability under current and future climate scenarios for a range of sites across the world. For example, climate change impacts on forest habitat suitability and diversity in the Korean Peninsula (Lim et al., 2018); how much does climate change threatens European forest tree species distributions (Dyderski, Paź, Frelich, & Jagodziński, 2017); climate change impact on the distribution of Dipterocarp trees in Asia (Deb, Phinn, Butt, & McAlpine, 2017); induced range shift in miombo woodland due to climate change in Southern Africa (Pienaar, Thompson, Erasmus, Hill, & Witkowski, 2015). However, most of these studies are limited to a single-tree species, lacking multiple tree species (such as Edenius & Mikusiński, 2006;Jiménez-Alfaro et al., 2018;Rondinini, Stuart, & Boitani, 2005). In contrast, modelling multiple tree species tend to yield better results (Edenius & Mikusiński, 2006) as this approach relies on detecting the shared pattern of the environment response for sparsely recorded species, thereby simplifying intricate species-specific patterns. It also enables direct interpretation by decision-makers ) that would typically be at the community level, except for studies into particular threatened species (Brummitt et al., 2015).
Approaches like MaxEnt rely on the availability of species or habitat presence data, typically based on field observations of a particular species or habitat. In the United Republic of Tanzania, the National Forest Inventory (NFI) provides a comprehensive dataset that includes over 19 thousand observations of forest type with over 50 thousand points for dominant tree species over all ecological zones across the country (Minunno et al., 2019;Storch, Dormann, & Bauhus, 2018;Tomppo et al., 2014) providing an exciting opportunity to provide baseline maps of forests and woodlands extent and the subsequent influence of climate change (Lim et al., 2018).

| Study area
This study focused on the mainland United Republic of Tanzania (hereafter referred to as Tanzania) in East Africa ( Figure 1). Tanzania's precipitation is characterized by bimodal rainfall distribution patterns ranging below 400 mm and over 2000 mm per year (Hardy et al., 2013). The maximum mean temperature ranges of 26.6-33.1°C and minimum at 5.3-18.3°C (NBS, 2017). The diverse geomorphological landscape results in a variety of climatic conditions, giving rise to a different set of forest communities (Table 1) from lowland rainforests in the north-west of the country to montane forests scattered across upland areas associated with the Eastern Arc Mountains (Burgess et al., 2004).
In Tanzania, forests ( Figure 2 and Table 1) occupy an estimated 48.1 million hectares of land, equivalent to 55% of the total area (MNRT, 2015). Forests in Tanzania are rich in biodiversity and placed among the 36 global biodiversity hotspots (Hrdina & Romportl, 2017;Myers, Mittermeier, Mittermeier, Da Fonseca, & Kent, 2000). However, it is among the countries with the highest reported forest loss (Hansen et al., 2013) and high vulnerability to the effects of climate change (Montade et al., 2018;Platts, McClean, Lovett, & Marchant, 2008). Forests contribute significantly to rural life in Tanzania for food security, woody biomass for energy supply and household subsistence. Ecologically, forests help to conserve soil and water resources, harbouring genetic, functional and taxonomic diversity. Approximately 35% of forests are protected through forest reserves, national parks and game controlled areas.
However, 75% of forests are found on unprotected, general-use land and are therefore vulnerable to degradation or deforestation, mainly due to growing need for agriculture and biofuel production, as well as extensive uncontrolled firewood collection, charcoal production, as well as the effects of forest fires (MNRT, 2015).

| Forest occurrence data
Forest occurrence records were acquired from the National Forest   (Dyderski et al., 2017). The presenceonly records were chosen based on abundance from the plot measurements for each forest type ( Figure 2 and Table 1). The selection included both percentage frequency (occurrence) and abundance (proportional of individuals). This implies that only the most frequent and abundant species from each forest type were selected.
The plots consisted of 1, 5, 10 and 15 m radius concentric nested circular sub-plots, collected over a series of study clusters-L-shaped transects, consisting of six to ten plots with 250 m spacing between plots. The study clusters were distributed based on a double sampling for stratification approach (see Tomppo et al. (2014)) for full details of this procedure). The timberline species were excluded from the analysis as their habitat changes are mainly the result of different non-climatic anthropogenic drivers such as land management decisions (e.g. Bodin et al., 2013).

| Spatial rarefaction
Geographical bias in the habitat or species occurrence data is likely to result in model over-fitting and artificial inflation of model performance (Boria, Olson, Goodman, & Anderson, 2014;Veloz, 2009).
Therefore, the original 59,208 forest type points underwent a stepwise spatial rarefication process, based on the random selection of a single location within grids of increasing size (Brown, 2014).
TA B L E 2 Summary statistical information for major predictor variables of forest types based on the occurrence data used in this study. Bio1: mean annual temperature; Bio12: mean annual rainfall; Bio14: rainfall driest month; Elv: elevation; Tri: terrain ruggedness index Specifically, we created a 5 x 5 km fishnet grid over the entire extent, to produce a single distribution point selected in each grid, with at least the distribution points be at 5 km apart. It was performed for each forest category separately to avoid eliminating too many observations from less extensive forest types, such as mangrove and montane forests. This procedure resulted in the selection of 1,307 occurrence points (n = 103 montane, n = 276 lowland, n = 168 mangrove, n = 378 closed woodland, n = 301 open woodland and n = 81 thicket) that were considered to be spatially independent.

| Environmental variables
The selection of environmental variables was based on a conceptual model that encompasses factors deemed to control the presence, or in some cases, absence, of a particular species (Jiménez-Alfaro et al., 2018). In this instance, we based our variable selection on the parameters that control the physically based forest growth model 3-PG (Physiological Principle in Predicting Growth) (Landsberg & Waring, 1997;White, Scott, Hirsch, & Running, 2006). 3-PG includes a large number of parameters, but we limited our selection to those parameters listed in Appendix 1 as Table A2.
Future climate data were ensemble mean downscaled to the resolutions (~ 1 km) using 18 pairwise combinations of five regional climate models (RCMs) driven by 10 general circulation models 1-arc second elevation data were obtained from USGS Earth Explorer to generate a terrain ruggedness index, a proxy measure of topographic heterogeneity (Riley, DeGloria, & Elliot, 1999). Soil characteristic variables were obtained from the World Soil Information (ISRIC) (https://www.isric.org) included soil types (see Appendix 1 as Table A1) (Hengl et al., 2015). A pairwise Pearson correlation (r) was used to test for collinearity between predicting variables, taking a relationship r > 0.7 or < −0.7 as highly correlated (Braunisch et al., 2013;Dormann et al., 2012) (see Appendix 1 as Table A3).

| Forest modelling
The modelling process focused on forest types (Table 1, Figure 2) based only on dominant tree species (Table 2). The inventory data adequately presented the distribution of forest types at different compositional gradients to predict suitable habitats for both current and future climate. The approach involves modelling forest types independently and then ensemble the results .   (Merow, Smith, & Silander, 2013). The loglog (clog log) output format was selected based on a sampling design that typically reflects the presence of localities and abundance of each forest type per quadrant at the presence probability of 0.63 (Phillips, 2005) and the location of occurrence is well estimated (Phillips et al., 2017). Jackknife resampling was used to examine the importance of each variable contribution to the potential distribution of vegetation types (Olivier, van Aarde, & Lombard, 2013).

| Construction of baseline and change maps
The final habitat suitability maps were generated by transforming the continuous probability values, ranged from 0 to 1 representing low and high probability, respectively, to discrete values of being either suitable or not suitable for the baseline. Following Spiers, Oatham, Rostant, and Farrell, (2018), the 10th-percentile training presence threshold was used to define suitable and unsuitable habitat for current and future projections. The future predicted habitat is calculated, for each forest type and taken as the difference between the baseline model and the future models to generate change maps (Maharaj & New, 2013) at RCP4.5 and RCP8.5, respectively, and presented with four predicted habitats of unsuitable, suitable, expansion and contraction (Table 3).

| Model performance evaluation
The models were evaluated using the qualitative statistic for the area under the curve (AUC) of the receiver operating characteristic (ROC) curves of the test data for the predicted mean accuracy model output for each forest type (Fielding & Bell, 1997;Merow et al., 2013).

| Baseline model accuracy assessment
AUC values have received criticism as they are vulnerable to overinflation of model performance where spatial autocorrelation exists within the model variables and where a modelled habitat niche is small relative to the extent of the modelled area (Williams Cross, Crump, Drost, & Thomas, 2015). To alleviate these issues, we conducted an independent measure of model accuracy using the forest tree species data, removed during spatial rarefication process, including a total of 57,901 points. The agreement was quantified using three metrics: 1) overall % accuracy and associated confidence inter-

| Model performance and habitat suitability estimation
Mean test AUC score demonstrated a high degree of accuracy (AUC > 0.9) for modelling the suitability of montane, lowland, man-

| Variables importance to each model
Precipitation and temperature (mean annual precipitation, rainfall driest month and mean annual temperature) (Table 6)

| Predicted forests habitat change
We assessed the impact of climate change on forests extent in Tanzania using national forest inventory data (Tomppo et al., 2014).
Our results indicate that climate change will affect all forest habitats suitability across Tanzania. The results reveal that climate change will threaten forests at various scales: forests with a narrow geographical range occurring at high altitude (i.e. montane forests) will experience more loss of their current habitat in the future. This may be associated with fragmented strips of montane forests, and particularly high endemism has increased a great sensitivity to climate change (Foster, 2001). Moreover, future climate change will extensively threaten microhabitat forests (i.e. thickets) occurring in a semi-arid climate (Moncrieff, Scheiter, Slingsby, & Higgins, 2015).
These projections indicate that climate change, especially temperature rise, will accelerate habitat loss of already vulnerable forests such as thickets (Chidumayo et al., 2011). Mangrove forests are predicted to expand their current range as a response to climate change (Godoy & Lacerda, 2015), although the future extent shift is more likely to be driven by sea-level rise, which was not factored, into the present study (Alongi, 2008).

| Potential suitable habitat impacted
The loss of suitable habitat for the montane forest is projected to be extensive, with losses exceeding 40% even under the optimistic RCP4.5 scenario by 2055 (Tables 7 and 8). This predicted loss is particularly pronounced in the high biodiversity areas of the Eastern Arc Mountains, a foothill of Rungwe and Livingstone mountain range along Lake Nyasa (Figure 6a). A projected reduction in rainfall results in a contraction of montane forests to higher elevations, illustrated by the projected loss of montane forest communities at lower elevations around Mount Kilimanjaro. The isolated nature of these montane habitats, sometimes termed "forest islands" (Fjeldså, 1999), form essential refugia for several species including 15 mammal species identified as vulnerable or high-risk status within the Udzungwa Mountains (Rovero et al., 2006). Forest loss in montane regions has severe implications for wildlife migration as these forests provide vital corridors linking reserves in Ruaha to the Selous Game Reserve via the montane forests of the Udzungwa Mountains (Jones et al., 2012). Additionally, loss of suitable habitat for forests in these regions is likely to increase sediment supply within the Rufiji basin, affecting downstream wetland dynamics and water resources (Ochieng, 2002).
Rising temperatures and reduced rainfall during the dry season are projected to result in losses of suitable lowland forest habitat above 10% by 2085 (Table 8) (Medley & Hughes, 1996;Sharam, Sinclair, & Turkington, 2006 (Continues) carbon sinks), capturing CO 2 from the atmosphere and store it in their biomass than terrestrial trees (Alongi, 2012;Ray & Jana, 2017).
Under projected climate change scenarios, habitats suitable for mangrove forests are predicted to expand their range by 40% (Tables 7 and 8) at both low and high emissions (Figure 8). It is chiefly due to rising temperatures and subsequent evaporation, coupled with reduced annual rainfall totals leading to increased salinity, a favourable condition for mangrove ecosystem (Alongi, 2015). Therefore, an increase in temperature would be positive to the mangrove ecosystem as more accelerated growth, changes in community composition, diversity and latitudinal expansion (Alongi, 2015;Hanebuth, Kudrass, Linstädter, Islam, & Zander, 2013). Similarly, a rise in sea level influenced by future climate change is expected to alter mangrove forests significantly as they are susceptible to any shift in sea level (Alongi, 2008;Crase, Vesk, Liedloff, & Wintle, 2015). The relative sea-level rise may cause landward retreat in mangrove forests supported by sediment composition on the upland habitat (Godoy & Lacerda, 2015).

| Implications for forests conservation planning
A dramatic decline in the projected extent of Tanzanian forests over the next 50 years is expected to be driven by regional and national climatic factors. Our study, therefore, identifies a tractable method of using existing forest inventory data to predict the distribution of The habitat modelling procedure demonstrated that climate has a substantial control on the distribution of Tanzanian forest communities. As a result, even under an optimistic climate change scenario (RCP4.5), forest communities in Tanzania are projected to decrease in an immense range. Notably, montane forests of Tanzania are globally significant in terms of biodiversity (Fjeldså, 1999;Jones et al., 2012;Rovero et al., 2006), yet they are projected to halve in extent by 2085. Although forest communities like closed, open woodland, and mangrove forest may expand into other regions in response to climate change, montane forests are constrained by elevation and therefore show particular vulnerability to changes in temperature. As such, montane species may well act as a barometer for regional climate change (e.g. Kimball & Weihrauch, 2000). Focusing on monitoring efforts in these regions may be vital in identifying changes in forest composition and biodiversity in response to climate change, in the hope that this can steer policy before we reach a crucial tipping point. For instance, through efforts like the African Forest Landscape Restoration initiatives (FLR) with a target of restoring 100 million hectares of deforested and degraded landscape across Africa by 2030 (Mills et al., 2015).
Other more direct anthropogenic factors compound the threat from climate change as these forest communities undergo extensive F I G U R E 6 Predicted spatial changes in the potential habitat distribution area based on the thresholds provided in Table 4 for ( felling for building material and charcoal production, as well as increasing frequency of forest fires (Sharam et al., 2006). These forest habitats extend across approximately half of Tanzania, and habitat degradation or loss of this magnitude can have serious implications, particularly in terms of loss of carbon sink (Makundi & Okiting'ati, 1995) and their role in wildlife migratory patterns: projected losses coincide with wildlife corridors with regional significance such as the Selous-Niassa, Udzungwa-Ruaha and Muhezi-Swagaswaga migratory routes (Hofer et al., 2004;Medley & Hughes, 1996;Sharam et al., 2006).

| Limitations of the study
This study adopted a widely accepted methodology (e.g. Elith et al., 2006;Lim et al., 2018;Merow et al., 2013) that facilitates mapping of forests habitat suitability and their alteration due to climate change; however, it suffers from the same limitations associated with known uncertainties of the data and climate models (e.g. Watling, Brandt, Mazzotti, & Romañach, 2013). Similarly, the forest habitats prediction focused at a county level, and therefore, our results should be interpreted at the national scale rather than a regional or small local scale.

F I G U R E 7
Predicted spatial changes in the potential habitat distribution area based on the thresholds provided in Table 4 for (

| Future research perspectives
Future simulations should consider using the information on the spatial pattern of change, such as proximity (distance rasters) to urban centres and road networks, and density rasters of projected population growth (population data surface). Construction of road networks across forests is likely to trigger increased forest degradation and fire incidences that in turn are expected to alter regional climate (Fonseca et al., 2019;Nepstad et al., 2001). Future work also should explicitly consider the impact of sea-level rise and geomorphology on Tanzanian mangroves to fully understand how these essential habitats might change as a result of climate change.

| Conclusions
Climate change will alter Tanzanian forests by accelerating habitat loss, fragmentation and hence reducing ecological connectivity. The effect of forest fragmentation will compromise the potential plant pollinators' movement and seed dispersal.
The induced fragmentation is especially severe when essential wildlife corridors, such as riparian zones that connect different areas of the landscape, are impacted. The optimal management solution in this regard is to increase ecological connectivity in current forest planning and management. Ecological connectivity should be maintained in habitats that are predicted not to change and expand under future climate change by preserving native forests and, where possible, protect the remaining forest areas from other anthropogenic disturbances. Improving ecological connectivity would significantly enhance not only sustainable forest management but also improve the design and implementation of forest projects and programmes. For example, ecological connectivity in forests will improve wildlife movement. This is more prominent for the dispersed population of large mammals (e.g. elephants) (Ntongani et al., 2010), when enclosed, increase the destructions of the highly diverse forest habitats (Ripple et al., 2015). Therefore, increasing forest connectivity will enhance the natural resilience of the remaining forests to the predicted effects of climate change. Consequently, the findings call for conservation planning in different dimensions: improve management of the existing protected areas which can absorb the impact of climate change, but also expanding to newly suitable areas with effective land use planning, conservation and land reclamation.

ACK N OWLED G EM ENTS
The research was supported by the Commonwealth Scholarship Commission (CSC) in the UK with Award number TZCS-2017-721.
We are indebted to the Tanzania Forest Service (TFS) Agency, of the Ministry of Natural Resources and Tourism for providing forest inventory data and permit to conduct drone-based fieldwork on the forest areas.

PE E R R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/ddi.13152.

DATA AVA I L A B I L I T Y S TAT E M E N T
The datasets for this study are available from the corresponding author upon request.

Our research at the Earth Observation and Ecosystems
Dynamics Laboratory focuses on the integration of ground, airborne and space-borne remote sensing data for better un-

TA B L E A 3
Correlation matrix between environmental predictors and variables with high correlation are shown with r > 0.7 or < −0.7 in bold at a significance level of .01 (see Table A2