African savanna raptors show evidence of widespread population collapse and a growing dependence on protected areas

The conversion of natural habitats to farmland is a major cause of biodiversity loss and poses the greatest extinction risk to birds worldwide. Tropical raptors are of particular concern, being relatively slow-breeding apex predators and scavengers, whose disappearance can trigger extensive cascading effects. Many of Africa’s raptors are at considerable risk from habitat conversion, prey-base depletion and persecution, driven principally by human population expansion. Here we describe multiregional trends among 42 African raptor species, 88% of which have declined over a ca. 20–40-yr period, with 69% exceeding the International Union for Conservation of Nature criteria classifying species at risk of extinction. Large raptors had experienced significantly steeper declines than smaller species, and this disparity was more pronounced on unprotected land. Declines were greater in West Africa than elsewhere, and more than twice as severe outside of protected areas (PAs) than within. Worryingly, species suffering the steepest declines had become significantly more dependent on PAs, demonstrating the importance of expanding conservation areas to cover 30% of land by 2030—a key target agreed at the UN Convention on Biological Diversity COP15. Our findings also highlight the significance of a recent African-led proposal to strengthen PA management—initiatives considered fundamental to safeguarding global biodiversity, ecosystem functioning and climate resilience.


Road transect analyses
Mean encounter rates for each combination of species, survey period and protected area (PA) status were extracted from published sources for West Africa (28 species) 1 , Kenya (22 species) 2 and northern Botswana (25 species) 3,4 .Encounter rates for northern Cameroon were calculated from a combination of published 5 and unpublished survey data (15 species), collected in 1973, 2000 and 2007-2010.The latter were made by R.B. and B.M.C., using the same routes and methods as in ref. 5 , enabling us to extend the 'recent' survey period in northern Cameroon to span 2000-2010.
In raptor road transect studies, most individual transect surveys yield few or no sightings of a given species.Count data therefore tend to follow a Poisson-like distribution, with high levels of variation between transect surveys, indicating a strong likelihood of over-dispersion, and potentially of zeroinflation.To estimate mean encounter rates during the early (1973) and recent surveys (2000-2010)  in northern Cameroon, we used generalised linear mixed effects models (GLMMs), consistent with the approach adopted in Kenya 2 and northern Botswana 4 .Specifically, we used the package glmmTMB 6 in R version 3.5.1 7 to model the relationship between survey period, PA status and the number of individuals seen during each transect survey, specifying a Poisson or a negative binomial error distribution and calculating the variance for the latter either as φµ or as μ(1+μ/k) (refs. 6,8).For each of these three models, we specified a non-zero-inflated and two zero-inflated versions, where the level of zero inflation was either assumed to be constant, or to vary in relation to survey period and PA status, after ref. 6 .We thus compared nine model variants per species.
In each model we entered the number of individuals of the target species per transect survey as the dependent variable.Since changes in abundance between survey periods were likely to differ between protected and unprotected areas, we entered 'Period' and 'PA status' (both binary) as an interaction term.'Survey year' was fitted as a random term, as some transects had been surveyed more than once in a given year.Since transects varied in length, we entered transect length (log transformed) as an offset term, following refs. 9,10, and used a log-link function throughout.
When examining the effects of period and PA status, we used the package DHARMa 11 to identify models showing an acceptable fit, and selected a final model based on minimum AICc value.We used the Anova function to calculate Chi-squared and P-values for each explanatory term, and applied the R predict function to derive the number of encounters predicted for each transect survey.We divided these values by transect length to give a predicted encounter rate (birds 100 km -1 ), and calculated the mean encounter rate for protected and unprotected areas (UPAs) in each survey period.
Mean encounter rates for raptor species surveyed in northern Botswana during 1991-1995 and 2015-2016 are given in ref. 4 , in which separate estimates for PAs and UPAs are provided for only two species, whose change rates differed significantly in relation to PA status.In order to calculate mean encounter rates for all combinations of species, period and PA status in northern Botswana, we applied the same approach as described above for northern Cameroon, and as used for Kenya 2 .In Botswana, the Degree Grid Square (DGS) through which each transect passed had also been recorded, and since some DGSs encompassed multiple transects, we entered 'DGS' as a random term, to account for the lack of independence between such transects.

Case selection
We restricted our analyses to cases in which -within the early survey period -at least five individuals were seen in PAs and five in UPAs, with a minimum of 20 individuals seen in total.These thresholds were met in 90 combinations of species (42) and study area (up to four), for which the median number of individuals encountered in both periods combined was 206 (range: 27-8,193; quartiles: 102-424; Supplementary Table 3).Within the early survey period alone, a median of 32 individuals (quartiles: 17-70) were detected in PAs and 62 (quartiles: 27-136) in UPAs.Sample sizes fell below one or both thresholds in a further 47 cases, including 10 cases in which fewer than five individuals were recorded in PAs and UPAs combined.Cases in which sample sizes failed to meet the above thresholds during the early survey period were excluded from the analysis, to ensure that trend estimates were not disproportionately influenced had 1-2 additional, or fewer, individuals been seen.Furthermore, where initial sample sizes fell below five individuals the apparent direction of change is likely to have been biased -towards recording substantial proportional increases.
Since trend estimation was much less sensitive to the effects of small sample size during the subsequent survey period, minimum sighting thresholds were not applied to recent surveys.This also ensured that cases in which a species had been extirpated, or had become effectively too rare to detect, were retained in the dataset.
In addition, we excluded six cases involving three species: African fish eagle Haliaeetus vocifer (all four studies), which is associated mainly with large water bodies and hence poorly captured by road transect studies; and common and lesser kestrel (Falco tinnunculus and F. naumanni) in West Africa, where these taxa were not always separated at species level 1 .

Survey routes and protected areas
Our estimates of the extent of land surveyed in each country were based on biome coverage 12 and annual precipitation levels recorded in areas through which survey routes had passed.Areas in which the biome and average annual precipitation matched those of the routes surveyed were included in the analysis; those in which the dominant biome or precipitation levels differed from those associated with survey routes were excluded (Extended Data Fig. 1, Supplementary Table 6).
The effectiveness of protected area management can differ markedly between PA types and countries, reflecting differences in their aims, national legislation, governance and conservation budgets.PA categories that were considered by study authors to afford little or no meaningful protection for wildlife, or where the degree of protection afforded was uncertain, were excluded from analyses and treated as unprotected.These typically included forest reserves, hunting areas, partial reserves and community conservancies.Across the six countries in which road transects were conducted, the remaining categories (Supplementary Table 4) comprised 168 PAs, of which 44 (26%) were surveyed.Together, surveyed sites accounted for 73% of the land contained within the 168 PAs.

Comparing protected and unprotected areas
Encounter rate differences between protected and unprotected land may potentially have several causes, including inherent differences in habitat suitability that pre-date site designation; a deterioration in habitat suitability outside of PAs, as human pressures intensify; or the positive effects of habitat management and species protection within PAs.
To examine the potentially confounding effects of habitat variation on PA-UPA differences, we reexamined raptor encounter rates from transects surveyed in northern Botswana.Here, the number of transect surveys conducted was large (n = 656) and the boundaries of several major PAs are completely straight for long distances (Extended Data Fig. 1), bisecting extensive tracts of land that are relatively uniform in terms of vegetation, rainfall and (exceptionally low) human population density (M.H., pers.obs.).Indeed, much of the area falls within the same biome 13 , and the habitat mix is essentially similar within PAs and adjacent UPAs (M.H., pers.obs.), minimising any confounding effects of habitat variation.
Since transects surveyed in Botswana had been assigned to their respective 1x1 (c.100x100 km) grid squares 3,4 , we compared encounter rates in three situations: A. protected areas; B. unprotected areas in squares where PAs were present; C. squares where PAs were absent.We predicted that for most raptor species, encounter rates would be highest on type 'A' and lowest on type 'C' transects.Conversely, we expected rates of decline between survey periods to be lower on type 'A' than type 'C' transects.For both measures we expected intermediate results from type 'B' transects, on the basis that these were likely to be physically closer and ecologically more similar to PA transects than were 'C' transects.Comparisons between 'A' and 'B' transects should therefore provide a more accurate measure of the effects of site protection, while partly controlling for habitat effects.
To make this comparison we modelled encounter rates for 23 species, from which at least five individuals were recorded on each of the three transect types during the early survey period (1990-1995).We estimated encounter rates using the approach described under 'Road transect analyses'.That is, we used generalised linear mixed effects models (glmmTMB in R version 3.5.1) in which the number of individuals encountered per transect survey was entered as the dependent variable.Period and PA status (now three categories instead of two) were entered as an interaction term, and grid square as a random term.Transect length (log transformed) was entered as an offset, following refs. 9,10, and a log-link function was used throughout.We used the R predict function to derive the number of encounters predicted for each transect survey, from which we calculated encounter rates (birds 100 km -1 ).
During the early survey period (1990-1995), encounter rates differed significantly between the three transect types (Kruskal-Wallis:  2 2 = 6.755,P = 0.034, n = 23 species).The median encounter rate from type 'A' transects (within PAs) was 1.6 times that of type 'B' transects (in squares with PAs; P = 0.356), and 3.2 times that of type 'C' transects (in squares lacking PAs; P <0.001; Bonferroni adjustment applied) (Supplementary Table 7, Fig. 3).In contrast, there was no significant difference in encounter rates overall during the second survey (Kruskal-Wallis  2 2 = 2.641, P = 0.267), in which the median encounter rate from type 'A' transects was 1.4 times that of type 'B' (P = 0.634) and 3.4 times that of type 'C' transects (P <0.001).
Between survey periods, 22% of species showed an increase in encounter rates on type 'A' transects, compared with 35% and 17% on transect types 'B' and 'C' ( 2 2 = 0.363, P = 0.834).The median percentage change was negative in all three categories ('A': -37%; 'B': -33%; 'C': -48%; Supplementary Table 7), but there was no significant effect of transect type overall (Kruskal-Wallis:  2 2 = 0.639, P = 0.726), or in pairwise comparisons In conclusion, raptor encounter rates on type 'A' transects were c. 1.4-1.6 times that of type 'B' transects, and 3.2-3.4times that of type 'C' transects, suggesting that raptor abundance levels within PAs in northern Botswana were approximately 1.5-3.2times higher than on unprotected land.Similar spatial patterns of variation in raptor numbers in Botswana have been described by Herremans and Herremans-Tonnoeyr 3 , based on point counts conducted during 1990-1995.These showed that core areas (>30 km within PA boundaries) supported higher densities than more peripheral protected land.Encounter rates on unprotected land up to 15 km from PA boundaries were just 65-70% of those at the core, dropping to c. 40% at a distance of 15-30 km.
These patterns could reflect variation in habitat suitability in relation to distance from PA boundaries, but are also likely to have been influenced by the dispersal of individual raptors between PAs and adjacent UPAs.A more targeted study design would be required to separate the direct effects of site protection from those of habitat variation and dispersal, to gauge the effectiveness of PA management on its own.However, we caution that the range and intensity of anthropogenic pressures influencing raptor trends in Botswana are likely to be markedly atypical of Africa as a whole.While the impacts of cattle grazing on vegetation cover -and hence on raptor prey populations -are substantial, the country has an exceptionally low human population density (c.10% of the continental average; Supplementary Table 8), whose direct effects on raptor populations, through persecution, poisoning, infrastructure development and disturbance, are thought to be small, relative to those in most other African countries 3 .

SABAP2 analysis
To determine the direction of change in the abundance of raptor species in South Africa, we examined variation in reporting rates during the second Southern African Bird Atlas Project, spanning 2008-2021 14 .That is, we measured change in the proportion of atlas survey visits during which at least one individual of the target species was recorded.Since longer visits are more likely to yield at least one sighting of a target species, we first investigated the relationship between visit duration and reporting rate for each of the 30 raptor species meeting the selection criteria described in the Methods.
While the mean reporting rate increased sharply between visits of 1 and 2 hours duration, little change was evident for visits of 2-5 hours, after which the mean rate increased gradually (Supplementary Fig. 2).We therefore limited our analysis of SABAP2 data to survey visits lasting 2-5 hours, inclusive.

Anthropogenic pressures
Declines in raptor encounter rates in West Africa (Burkina Faso, Niger and Mali combined) were significantly greater than those recorded in the three remaining regions (Fig. 4a), perhaps reflecting regional variation in human population growth and its attendant pressures.To examine this possibility, we assessed patterns of change in three anthropogenic factors over the lifespan of each study: human population density 15 , livestock density 16 and the proportion of land used for agricultural production 17 .Here, we summarise variation in these potential drivers among the countries surveyed, while noting that a full assessment of environmental change in each region lies beyond the scope of this study.
While the human population rose during the course of each study, the annual rate of change recorded in West Africa (+2.7%) lay between the extremes represented by South Africa (+1.3%) and Kenya (+3.2%), and was similar to that of Africa as a whole (+2.6%) (Supplementary Table 8).By 2005, human population densities within the five study areas ranged from 3.2 km -2 to 60.2 km -2 (Botswana and Kenya, respectively), and were lower in West Africa (14.8 km -2 ) than the average across Africa (30.5 km -2 ).This partly reflects the very extensive areas of Sahara-Sindian biome encompassed by Mali and Niger, where the human population is sparse, contrasting with the Sahel and Sudan-Guinea Savanna biomes further to the south, where raptor surveys were conducted.Hence, the average figure for West Africa may have under-estimated densities within the Sahel and Sudan-Guinea Savanna biomes.
Changes in livestock density showed a broadly similar pattern; the annual rate of change recorded in West Africa (+1.9%) lay close to that of Africa as a whole (+1.7%), and between the extremes evident in Botswana (-1.8%) and Cameroon (+2.9%).Similarly, by 2005 the mean livestock density in West Africa (6.9 livestock units km -2 ) was close to the mean for Africa (6.8), but midway between that of Botswana (2.2) and Kenya (15.9) (Supplementary Table 8).Here again, the relatively low average value recorded for West Africa likely reflects the sparsity of livestock within the Sahara-Sindian biome of northern Mali and Niger.Notwithstanding this effect, West Africa was exceptional in terms of agricultural expansion during the 1970s-2000s.The annual rate of expansion in agricultural landcomprising arable, permanent crops, other cultivated land and pasture -was more than three times that of Africa as a whole, and almost twice that of any other study area.Although the proportion of agricultural land remained comparatively low in 2005, West Africa's rapid agricultural expansion may partly explain the steep declines evident among its raptor populations.Note, however, that differences in cultural values, belief-based consumption and attitudes to wildlife use will also have contributed to regional differences in raptor population change.

Detectability
Patterns of change in species' encounter rates could have been influenced by changes in their detectability, caused by variation in the height or density of roadside woody vegetation over time.Since vegetation structure was not monitored in the four road transect studies, we were unable to test whether changes in woody cover had occurred on a scale likely to influence detection rates.Instead, we examined evidence from land cover studies, which confirmed that continental-scale changes in forest cover and woody plant encroachment had occurred throughout Africa since the 1970s.
In sub-Saharan Africa, the area of forest and natural non-forest vegetation contracted by 16.3% and 4.7% respectively, during 1975-2000 18 , coinciding approximately with the timespan of road transect studies conducted in West Africa, northern Cameroon and Kenya.In contrast, several studies have reported a marked, widespread increase in woody plant encroachment within non-forest natural habitats in parts of Africa 19,20 , including an 8% increase within sub-Saharan Africa in 1986-2016 21 .Thus, while the extent of natural non-forest vegetation contracted during 1975-2000 18 , woody plant encroachment within this land cover type increased 21 .
These changes have been attributed to a range of factors, including overgrazing by livestock, loss of browsing by wild megaherbivores, fire suppression and CO2 enrichment 19,20,21 .Venter et al. 21show that the direction and scale of change in woody plant cover has varied markedly between the seven countries included in this study.While woody plant cover declined in Niger and Kenya, it increased in the remaining countries (Supplementary Information Table 9).To gauge whether these changes are likely to have influenced raptor detection rates, we assumed that the annual rates of change in woody plant cover calculated for 1986-2016 21 had occurred throughout the timespan of each raptor study.Had this been the case, woody plant cover in Niger and Kenya would have declined by 3% and 4%, respectively, while rising by 18% in Mali, and by 19% in Botswana (since 1994) (Supplementary Information Table 9).Although woody cover change was highest in Cameroon as a whole (+40%), this figure is unlikely to have been representative of northern Cameroon -where the raptor surveys were conducted -due to its much lower rainfall.
Woody encroachment within savanna habitats may thus have influenced species detectability over the timespan of each survey.However, it is unlikely that a given increase in woody plant cover across the landscape has led to an equivalent reduction in raptor detectability.Thus, the overall 8% increase in woody plant cover recorded within non-forest natural habitats in sub-Saharan Africa during 1986-2016 21 is likely to have been partly offset by the 4.7% contraction in the area of non-forest natural habitats recorded during 1975-2000 18 , and was small in comparison with most of the species declines reported here.
Supplementary Table 1| The timing and distances covered by road transect surveys from which data were drawn.The four studies covered a combined distance of 94,151 km, of which 28% lay within protected areas.


Supplementary Table 5| Comparison of change estimates derived from two methods.We estimated each species' rate of change from two alternative scenarios, in which its encounter rates within unsurveyed PAs were assumed to have been the same as in surveyed PAs, or the same as in UPAs.We calculated median, Q1 and Q3 change rates, over three generations, based on these extrapolations (Table 1).Below, we compare the results from this approach ('Extrapolation') with those from an alternative, in which unsurveyed PAs were excluded from the analysis ('No extrapolation').Change estimates from the two approaches typically differed by just 1-2 percentage points (median: 1.0; range: 0.1-7.4;n = 42).Disparities between the two approaches are illustrated in Supplementary Fig. 1.  15 .

No extrapolation Extrapolation
c The combined density of livestock units for cattle, goats, sheep and camels ref. 16 .d Land that is either arable, under permanent crops, meadows or pastures, or cultivated and natural growing ref. 17  e Annual rate of change across the time span of each study.The start and end values were taken as the average of three consecutive years, e.g.1994, for 1993-1995.
f The area of agricultural land expressed as a percentage of the land area of the country or region. .
Supplementary Table2|Body mass, diet and generation lengths of study species.Species are listed in descending order of mass.Size classes were based on mass and diet, with the majority of those species assigned to the 'large' category (≥ 1300 g) being dependent on medium-sized mammals, birds or reptiles, or on carrion.
a Time interval separating the mid-points of the 'early' and 'recent' survey periods.aEndemic or near-endemic to the African continent.
27counter rates and total numbers of individuals recorded in each road transect study.Encounter rates are shown in relation to protected area status and survey period, and were derived from 53,209 sightings of the 42 study species.Supplementary Table4|Designation types included as protected areas in the four road transect studies.Designation names were taken from the World Database on Protected Areas27.
a The midpoint of each survey period.
Supplementary Table8|Changes in human and livestock population densities, and agricultural land area throughout each study period.Annual rates of change in density are shown for each country or region surveyed for raptors.Values for the African continent(1970-2005)are provided for context.Since survey periods differed between raptor studies, densities are given for a single year (2005), which falls within or near to the end of each study period, to illustrate the degree of variation between study areas.The mid-point of each survey period (N.Botswana, N. Cameroon, Kenya, West Africa).In South Africa, although SABAP2 data spanned 2008-2021, livestock density data were available only up to 2020.b Change in human population estimates, extracted from ref. a