Rare ground data confirm significant warming and drying in western equatorial Africa.

Background The humid tropical forests of Central Africa influence weather worldwide and play a major role in the global carbon cycle. However, they are also an ecological anomaly, with evergreen forests dominating the western equatorial region despite less than 2,000 mm total annual rainfall. Meteorological data for Central Africa are notoriously sparse and incomplete and there are substantial issues with satellite-derived data because of persistent cloudiness and inability to ground-truth estimates. Long-term climate observations are urgently needed to verify regional climate and vegetation models, shed light on the mechanisms that drive climatic variability and assess the viability of evergreen forests under future climate scenarios. Methods We have the rare opportunity to analyse a 34 year dataset of rainfall and temperature (and shorter periods of absolute humidity, wind speed, solar radiation and aerosol optical depth) from Lopé National Park, a long-term ecological research site in Gabon, western equatorial Africa. We used (generalized) linear mixed models and spectral analyses to assess seasonal and inter-annual variation, long-term trends and oceanic influences on local weather patterns. Results Lopé’s weather is characterised by a cool, light-deficient, long dry season. Long-term climatic means have changed significantly over the last 34 years, with warming occurring at a rate of +0.25 °C per decade (minimum daily temperature) and drying at a rate of −75 mm per decade (total annual rainfall). Inter-annual climatic variability at Lopé is highly influenced by global weather patterns. Sea surface temperatures of the Pacific and Atlantic oceans have strong coherence with Lopé temperature and rainfall on multi-annual scales. Conclusions The Lopé long-term weather record has not previously been made public and is of high value in such a data poor region. Our results support regional analyses of climatic seasonality, long-term warming and the influences of the oceans on temperature and rainfall variability. However, warming has occurred more rapidly than the regional products suggest and while there remains much uncertainty in the wider region, rainfall has declined over the last three decades at Lopé. The association between rainfall and the Atlantic cold tongue at Lopé lends some support for the ‘dry’ models of climate change for the region. In the context of a rapidly warming and drying climate, urgent research is needed into the sensitivity of dry season clouds to ocean temperatures and the viability of humid evergreen forests in this dry region should the clouds disappear.

248 another type of automated unit (TinyTag Plus 2, Gemini Data Loggers 249 https://www.geminidataloggers.com/data-loggers/tinytag-plus-2, some of which record both 250 temperature and relative humidity). TinyTags were deployed in the forest from 2007 and in the 251 savanna from 2008 and used until the present (with a gap at the forest site from mid-2015 to mid-252 2016 and intermittent recording throughout 2017 partly due to equipment malfunctions caused 253 by termite infestation). Two weather stations were installed in the savanna (sited near the 254 research station, on a rock 4m from the ground) and collected data between 2012 and 2016. A 255 Davis VantagePro2 (https://www.davisinstruments.com/solution/vantage-pro2/) was installed in 256 January 2012 and recorded rainfall, temperature, relative humidity, pressure, wind speed and 257 direction, UV index and solar radiation every 30 minutes for two years until the equipment was 258 struck by lightning in January 2014. A SKYE MINIMET weather station 259 (https://www.skyeinstruments.com/minimet-automatic-weather-station/) was installed at the 260 same location in 2013 and collected temperature, relative humidity, wind speed and direction and 261 solar radiation (but not rainfall as the gauge was defective). The SKYE unit ran intermittently 262 until 2016 when the equipment was also damaged by lightning: data records between January 263 2014 and November 2014 were also lost. Finally, a sun photometer was installed at the research 264 station in April 2014 and used to record aerosol optical depth up to the present as part of the 265 NASA Aerosol Robotic Network (Aeronet; https://aeronet.gsfc.nasa.gov/; Holben et al. 1998). 266 Despite sustained effort, the remote and challenging environment at Lopé has led to a patchy 267 weather data record. This situation has been exacerbated since the introduction of automated 268 loggers, due to unreliable performance combined with difficulties and time delays in replacing or 269 repairing malfunctioning equipment and respecting annual calibration schedules with 270 manufacturers based in Europe or the USA. New equipment was often introduced out of 271 necessity when previous equipment failed, precluding the opportunity of collecting simultaneous 272 data for standardisation. Such problems have been experienced at many other field stations 273 across Africa (Maidment et al. 2017) and homogenisation is necessary in most long-term 274 instrumental climatic data sets (Peterson et al. 1998). It was therefore necessary to select and 275 standardise the Lopé data to reduce systematic biases between recording equipment. We 276 summarise the data selection steps we undertook below and provide further detail in the 277 accompanying Supplemental Information (Article S1 and Code S1). All Lopé data can be 278 downloaded from the University of Stirling's DataSTORRE (http://hdl.handle.net/11667/133).
279 Data cleaning and preparation 280 We constructed a long-term record of daily rainfall totals (1984-2018) by calibrating the two 281 sources of data (manual rain gauge and Vantage Pro weather station) using a simple linear model 282 on simultaneous records and taking the mean value for days with multiple observations (resulting 283 in a dataset of 12,050 complete daily observations out of a possible 12419 over 34 years). Where 284 possible we interpolated missing daily values using the ten-day running mean for the time series 285 (resulting in a dataset of 12111 interpolated daily observations), however 11 months spread over 286 three calendar years remained incomplete. We used these interpolated daily data to calculate total 287 monthly and annual rainfall for the months and years with complete data (397 complete monthly 288 observations out of a possible 408 and 31 complete years out of a possible 34). 289 Temperature data were recorded using six different types of equipment across two sites 290 (recorded in the forest from 1984 to 2018 and in the savanna from 2002 to 2018). Where there 291 were multiple observations from overlapping data records we calculated mean daily maximum 292 and minimum values for each site and day in the time series, and used this dataset to demonstrate 293 temperature seasonality at each site (resulting in a dataset of 7058 daily observations out of a 294 possible 12419 over 34 years at the forest and 4878 daily observations out of a possible 5844 295 over 16 years at the savanna). To create continuous time series for periodicity analyses we 296 calculated mean monthly maximum and minimum daily temperatures for each month in the time 297 series with more than five observations (resulting in a dataset of 327 monthly observations out of 298 a possible 408 from the forest site and 166 monthly observations out of a possible 192 at the 299 savanna site). Minimum daily temperatures are recorded during the night and thus avoid errors 300 associated with direct solar radiation (which we found to vary between our equipment, Article 301 S1). Because of this we chose to use minimum daily temperatures to assess long-term trends and 302 inter-annual variation. We constructed a long-term daily record by calculating mean daily 303 minimum temperature using data from both sites combined (8217 daily observations out of a 304 possible 12419 over 34 years). We summarized these data to a monthly mean time series for 305 months with more than five observations (372 monthly observations out of a possible 408 over 306 34 years). 307 Finally, we used the shorter (and/or patchier) periods of data available for relative humidity 308 (2002-2018), solar radiation (2012-2016), wind speed (2012-2016) and aerosol optical depth 309 (2014-2017) to assess seasonality and periodicity for these climate variables. We used night-time 310 relative humidity records (6pm-6am) to avoid errors associated with direct solar radiation and 311 converted to absolute humidity (g/m 3 ) using simultaneous temperature records within the R 312 package humidity (Cai 2008). We extracted aerosol optical depth data at wavelengths relevant for 313 photosynthetic activity (440, 500 and 675nm). 314 Gridded regional temperature datasets 315 Because of missing data and lack of simultaneous recording between temperature equipment at 316 Lopé we also downloaded two widely used gridded regional data products with which to 317 compare the Lopé data: daily minimum air temperature from the Gridded Berkeley Earth Surface 318 Temperature Anomaly Field (1° resolution; Rohde et al. 2013)  To characterise the seasonality of each weather variable we calculated mean values from 339 empirical daily data at three different scales: the mean value for each day of the calendar year 340 (DOY, fine-scale), the ten-day running mean of DOY (medium-scale) and the mean value for 341 each calendar month (coarse-scale). To formally assess the periodicity of each variable we used 342 Fourier analysis. The Fourier transform is a form of spectral analysis used to calculate the 343 relative strength of all possible regular cycles in time series data (Bush et al. 2017). We created 344 standardized, complete time series by filling missing values in monthly time series using the 345 mean value for the corresponding calendar month and standardizing the data by subtracting the 346 mean and dividing by its standard deviation. We then computed the Fourier transform for each 347 time series using the spectrum function from the R Stats package (R Core Team, 2019) and 348 inspected the spectra plots for peaks that represent strong regular cycles in the data (Bush et al. 349 2017).
350 Long-term trends 351 We used a linear regression framework to test whether rainfall and minimum temperature had 352 changed over the observation period (1984-2018) using non-interpolated daily data. We fitted 353 compound Poisson generalized linear mixed models (CPGLMM) for daily rainfall and linear 354 mixed models (LMM) for minimum daily temperature to account for their respective data 355 distributions. CPGLMMs are exponential dispersion models based on the Tweedie distribution 356 and are recommended for daily or monthly rainfall data which is positive and continuous with 357 many exact zeros (Hasan and Dunn 2010). We fit CPGLMMs using the cplm R package (Zhang 358 2013) and LMMs using the lme4 R package (Bates et al. 2015). DOY was included as a random 359 intercept in all models to account for seasonality and the hierarchical structure of the data. We 360 fitted initial models with Year (continuous, rescaled) as the predictor (representing long-term 361 change) and compared these to intercept-only models (representing no long-term change) 362 preferring simple models (few parameters) with lowest AIC (significantly different if delta AIC 363 >2). See R-style model notation below with ε representing residual error not accounted for by the 364 predictors of the model. (2) Daily Rainfall ~ 1 + (1|DOY) + ε 368 (3) Minimum Daily Temperature ~ Year + (1|DOY) + ε 369 (4) Minimum Daily Temperature ~ 1 + (1|DOY) + ε 370 371 We repeated the same procedure for gridded temperature data for Lopé from the daily Berkeley 372 and monthly CRU datasets. DOY was included as a random intercept within the models with 373 daily response data and Month was included as a random intercept within the models with 374 monthly response data.  Fig. 2A). We included Year (continuous, rescaled), 387 Season (factor with four levels as above) and their interaction as predictors in initial models to 388 represent long-term change varying by season. We fitted subsequent models without the 389 interaction term to represent long-term change not varying by season and compared the models 390 using AIC values. DOY was included as a random intercept in all models, as before. To estimate the magnitude of the trend in each season, rather than comparing to the global 398 intercept, we modified the best models by temporarily removing the global intercept. For all 399 models described above we inspected the residuals to check for temporal autocorrelation using 400 the R package itsadug (van Rij 2017). None of the median autocorrelation functions 401 (autocorrelation calculated for each DOY or Month respectively) showed significant temporal 402 autocorrelation. 481 change in daily rainfall in DJF and ON and significant decline in JJAS (-0.07 mm per day per 482 decade, equating to -6.35% of mean JJAS daily rainfall). 483 Minimum daily temperature at Lopé increased at a rate of +0.25°C per decade, equivalent to 484 +1.1% relative to mean minimum temperature for the time period (LMM, Estimate = 0.24; SE = 485 0.01; T = 24.84; 95% Confidence Interval = 0.22: 0.26; Table 3 and Fig. 3B). The rate of 486 warming also varied by season (Tables 4 and 5 Fig. 3C). Over time, the signal of the biannual rainfall cycle appeared to 500 decrease while the annual cycle strengthened (Fig. 3E). The annual cycle for minimum 501 temperature was, on average, three times as powerful as the biannual component and 23 times as 502 powerful as the multi-annual component (Fig. 3F). The signal of the annual cycle remained 503 dominant throughout most of the time period with patches of low power at the end of the 1980s 504 and between 2007 and 2010 (Fig. 3D). There were patches of high power in the multiannual 505 component around 2000. The signal of both the annual and semi-annual components appear to 506 have been increasing in strength over time (Fig. 3F).
507 Oceanic influences 508 Wavelet coherence analyses showed that the ENSO index (MEI) had the strongest coherence 509 with both rainfall and temperature at Lopé over the last three decades at multi-annual scales (2-4 510 years; Fig. 4 and Fig. S3). However, the influence of ENSO has been patchy through time; 511 Coherence between ENSO and rainfall was particularly strong pre-1990 and between 2007 and 512 2012 (Fig. 4A) while coherence between ENSO and minimum temperature was fairly consistent 513 up to 2000 and has become weaker since (Fig. 4B). SSTs of the southern tropical Atlantic 514 showed strong coherence with Lopé rainfall pre-2000 while SSTs of the northern tropical 515 Atlantic showed strong coherence with Lopé rainfall post-2000 at multi-annual scales (4-8 years; 516 Fig. 4C and E and Fig. S3). SATL cycled in phase with Lopé rainfall (arrows point to the right) 517 while NATL cycles in anti-phase during the 2005 to 2010 period (arrows point to the left; 518 Figures 4C and E). Within the reliable region of the wavelet coherence plots (away from edge 519 effects) the IOD does not appear to have had a particularly strong or consistent relationship with 520 either rainfall or temperature at Lopé (Fig. 4G and H and Fig. S3).

Our results
523 Lopé weather has changed significantly over the last three decades, warming at a rate of +0.25°C 524 per decade (minimum daily temperature) and drying at a rate of -75 mm per decade (total annual 525 rainfall; Figure 3A and 3B). Both trends are seasonally dependent (Table 4); with significant 526 warming occurring in all seasons, being most pronounced from October to February (see model 527 estimates in Table 5). The rainfall decline occurred predominately between March and 528 September, incorporating both the long rainy season and the long dry season (see model 529 estimates in Table 5). The drying trend at Lopé supports observations of reduced Ogooué river 530 flow from March to September (Mahe et al. 2013) and precipitation declines evident from 531 gridded gauge-data for the Gabon/Cameroon region (-1% total annual rainfall, 1968-1998; Malhi 532 & Wright 2004). However, the Lopé total annual rainfall decline of -5.5% per decade exceeds 533 the trend estimated from the regional gauge-data. While the strength of the biannual cycle in 534 rainfall appears to be declining at Lopé along with the overall long-term trend, the annual 535 component is getting more powerful. Declines in rainfall in the long dry season (June-536 September) but not the short dry season (December-February) are likely to be contributing to an 537 increased contrast between the two dry seasons and enhancing the overall annual rainfall cycle 538 (Table 5). 539 The warming trend recorded at Lopé is greater than that estimated for the location over the same 540 time period using the Berkeley and CRU gridded datasets (+0.16°C and +0.19°C respectively) 541 and that identified using satellite data for mean annual temperature for all tropical Africa 542 (+0.15°C , 1979-2010; Collins 2011). However, it is lower than the change estimated from 543 gridded observational data (CRU) for mean annual temperature specifically for African tropical 544 forests (+0.29°C per decade, 1976-1998; Malhi & Wright 2004). While there remain issues with 545 the Lopé temperature data record (lack of simultaneous recording to calibrate data recorded 546 using different equipment), there is good evidence from supporting datasets and the literature 547 that the warming trend observed at the site since 1984 is real. The slower warming trend in the 548 already cool, long dry season is likely to account for the apparent increase in the power of the 549 annual cycle for Lopé minimum temperature. 550 Our analysis of seasonality at Lopé further serves to emphasise the ecological importance of the 551 long dry season in western equatorial Africa; three to four months of dry (almost no rainfall for 552 90 consecutive days), cool (mean maximum daily temperature is 2.5°C lower in July compared 553 to April) and windy conditions with low humidity and limited light availability (Figure 2). Such 554 a defined dry season poses specific constraints to the biota and is likely to act as a temporal 555 marker for ecological events, similar to a winter event in temperate regions, while the response 556 of the plant community to recurrent and predictable seasonal drought during the long dry season 557 could be used to estimate the long-term response to drying over multi-annual time scales (Detto 558 et al. 2018). 559 Reduced light availability during the long dry season in the Gabon region is most strongly 560 associated with seasonal low-level cloud cover (Philippon et al. 2019). Aerosol load may also 561 have a seasonal influence on light availability as aerosol optical depth and solar radiation appear 562 to cycle in anti-phase although we are not able to tease apart their relative importance in this 563 analysis (Figure 2). Low direct solar radiation and cool temperatures will reduce water demand 564 during these months (e.g. potential evapotranspiration is less than 2.3mm per day during the long 565 dry season in SW Gabon; Philippon et al. 2019) and are likely contributors to the forest's ability 566 to maintain an evergreen canopy despite seasonal drought. Unsurprisingly, the savanna and 567 forest experience different microclimates because the forest canopy creates a more humid, cooler 568 climate throughout the year with a reduced range between daytime and night-time temperatures 569 ( Figure 2). It is possible that the forest may also directly enhance water supply for plants during 570 periods of low precipitation / high cloud cover due to foliar interception of low-lying clouds. At 571 another tropical forest site (~1000m above sea level), foliar interception has shown to contribute 572 an additional 40% of moisture compared to rainfall (Hutley et al. 1997) meaning that rain gauge 573 data does not always accurately represent the water balance of the forest ecosystem (Philippon et 574 al. 2019). While we do not have information on foliar interception of clouds at our study site 575 (~280m above sea level), the hydroclimatic conditions of the region do not predict occurrence of 576 cloud-affected forest here (Oliveira et al. 2014). We can assume that the impact of cloud 577 interception on water supply is negligible, although it may occur on forested hills above 600m 578 (e.g. the hill local to the study station known as The Camel which reaches 678m) and a dedicated 579 research agenda would be needed to assess the any direct contribution of clouds to moisture 580 availability, especially during the cloudy dry seasons. 581 582 We have also shown that variability in temperature and rainfall at our site is strongly influenced 583 by global weather patterns. The most important influence on Lopé temperature is the Pacific 584 ENSO index, with our analysis showing strong coherence between these two datasets on multi-585 annual scales, especially pre-2000 ( Figure 4). This result is supported by a continent-wide study 586 showing warming throughout Africa in El Nino years . None of the other oceanic 587 indices appeared to influence Lopé temperature in a consistent way ( Figure 4). As for Lopé 588 rainfall, the most important influence appears to be the tropical Atlantic. Rainfall cycled in phase 589 with southern tropical Atlantic SSTs pre-2000 and in anti-phase with northern tropical Atlantic 590 SSTs post-2000 on multi-annual scales (Figure 4). The phase relationships between these data 591 series indicate that higher than average rainfall at Lopé coincides with warm conditions in the 592 south tropical Atlantic and cool conditions in the north tropical Atlantic. This     Manuscript to be reviewed