Philippa Kaur

Quantifying heating of coastal ecosystemsA variety of metrics have been developed to quantify MHWs, but to date, they have mainly focused on surface heating evident from SST. Surface MHWs have been defined based on exceedance of 90% confidence intervals calculated seasonally from historical SST10,11. Defined in this way, the threshold temperature for quantifying a MHW is likely to vary seasonally from, for example, winter to summer. However, the physiological and ecological relevance of seasonally changing thresholds remains unclear, especially for tropical biota such as corals that typically inhabit a relatively narrow range of temperatures close to their physiological thermal limits12,13,14. Assessing heating in the specific context of temperature-induced coral bleaching has instead focused on calculating cumulative degree heating, sensitive to both the magnitude and duration of heating, above a fixed, putative ‘bleaching threshold’ defined by the local Maximum Monthly Mean SST (MMM); i.e., the mean summer-time peak temperature predicted to initiate coral stress and bleaching. Such heat accumulation has most commonly been expressed as Degree Heating Weeks (DHW in °C-weeks) accumulated over a 12-week period15,16, or sometimes using monthly SST data to compute Degree Heating Months (DHM in °C-months) over 3 months17,18. Numerous studies have documented coral bleaching in shallow water linked to periods of anomalously high SST and accumulated DHW, especially during El Niño events1,2,4,19,20,21. However, DHW generally only explain a limited proportion of the observed variation in bleaching, even for communities in very shallow water (e.g., 50% of bleaching variation at 2 m depth7). A number of studies have documented bleaching that was less than that predicted based on contemporaneous SST and DHW. This discrepancy has sometimes been hypothesised to reflect ongoing coral acclimatisation to increasing temperatures and/or shifts in community composition towards more heat-tolerant genotypes and species22,23,24,25,26. Bleaching rates higher than predicted by SST, under limited DHW, have also been documented, and can be species-specific and more pronounced in heat-intolerant cryptic species27. It is unclear to what extent the relatively limited power of SST metrics such as DHW to predict coral bleaching results from an incomplete description of environmental conditions, a lack of nuance in describing biological thresholds and organisms’ reactions to elevated temperatures, genetic variation and cryptic species27,28, or, most likely, a combination of factors.Estimates of surface MHW severity are also highly sensitive to both the spatial and temporal scales of SST data considered. Present-day satellite products allow degree heating to be calculated at relatively fine spatial and temporal resolution using SST data measured daily over pixels of ~25 km2 size (e.g., NOAA Coral Reef Watch29). While many heating assessments in the context of coral bleaching continue to focus on long-term heating based on temperature anomalies accumulated across months (12 weeks or 3 months for DHW or DHM, respectively)18,20,30, here we focus on higher-resolution heating calculated as Degree Heating Days (DHD in °C-days) using daily SST data. Calculation of DHD over 12-day windows is analogous to the more coarse resolution DHW (i.e., weekly data over 12 weeks)31 but with different units, finer temporal resolution, and shorter time lags between the actual elevation of environmental temperatures and the resulting accumulated heating metric (see the DHD and DHW comparison in Fig. S5).To investigate the role of spatial scale in characterising MHWs, we analysed SST at a range of scales around Moorea between 1985 and 2019: 2° × 2° (~50,000 km2), 1° × 1° (~12,300 km2) and 0.1° × 0.1° (~100 km2) (see Fig. S1). The importance of spatial scales in heating severity apparent at the surface is reinforced by the regional heterogeneity in SST (see Fig. 2 and data animations in the Supplementary Information). Considering heating at local scales and finer temporal resolutions maximises the potential to detect, characterise and compare heating events at the surface in a way that is relevant to in situ MHW conditions (see also Guo et al.32. for the importance of scales in MHW assessments). For instance, when assessed using NOAA’s ‘Regional Bleaching Heat Stress Gauges’ for the Society Islands, DHW reached 4.54 °C-weeks in April 2016 and 5.35 °C-weeks in May 201933, indicating moderate likelihood of bleaching during both events, although inherent to calculations of accumulated DHW, the maximum heating was centred multiple weeks after the in situ heating events actually occurred (e.g., see Fig. S5e, f). Heat accumulation using higher resolution DHD across 12-day windows around Moorea itself (2° × 2°) was more closely aligned temporally to the heating events, suggesting a much higher bleaching risk in 2016 (6.83 °C-days) than 2019 (2.10 °C-days; Table S3). This likely reflects the localised heterogeneity in SST during the 2019 MHW, when hotter surface conditions prevailed north of Moorea (Fig. 1b). At a local scale, assessed using SST within a ~10 × 10 km area north of Moorea (see Fig. S1), heating was similar to regional estimates in 2016 (5.73 °C-days), but in 2019 revealed an intense, localised heatwave over Moorea’s north shore (15.4 °C-days; see Fig. 1e, f; Table S1). Because of these demonstrated advantages of higher temporal-resolution analysis, we rely on DHD rather than DHW for our analysis of MHW patterns among years at Moorea.Fig. 2: Regional sea-surface temperature (SST) variability during the peak of the 2016 and 2019 surface marine heatwaves around Moorea suggested hotter conditions during 2016.Panels show SST over a, b 20° × 20° and c, d 10° × 10° during 2016 and 2019, respectively, focusing on the date of peak SST observed over Moorea’s north shore (8 April 2016 and 4 April 2019). Dashed squares in (a, b) show extent of (c, d) and those in (c, d) the extent of data shown in Fig. 1a, b (2° × 2°). Coastlines based on Wessel and Smith77.Full size imageContrasting surface and subsurface heatingComparing surface and subsurface MHWs is challenging for many coastal ecosystems, even once SST data of appropriately fine spatial and temporal scale are obtained, due to a lack of long-term, in-situ temperature data through which to assess mean climatological patterns below the sea surface. Moorea represents one of the few coral reef systems with consistent in-situ observations over timescales (decades) and depths (sea surface to 40 m) relevant to understanding the oceanographic drivers of subsurface heating and their impacts on coral bleaching. Our analysis of long-term SST records indicates there have been 16 local-scale (0.1° × 0.1°) surface MHWs over the north shore of Moorea relative to a MMM-based bleaching threshold of 29.8 °C (Table S4), compared to 14 regional-scale (2° × 2°) events (Table S3). Localised heating over the north shore was often greater than regional SST would suggest, with the hottest event recorded in 2003 reaching 17.7 °C-days locally (Table S4), compared to 15.5 °C-days regionally (Table S3). Further details on historical events can be found in ‘Surface MHW history around Moorea’ in the Supplementary Information, which provides context for the six more recent surface MHWs that have occurred over the north shore since continuous in-situ, reef-level observations began at Moorea in 2005 (MHWs in 2007, 2012, 2015, 2016, 2017, and 2019; Table S4). Two recent, contrasting events in 2016 and 2019 demonstrate the extent to which thermal environments at depth, and the associated severity of coral bleaching, can vary substantially from predictions based on sea-surface conditions.MHW severity based only on SST can miss important information on the conditions experienced by organisms at depths greater than the surface skin layer quantified through remote sensing, which may only be few millimetres thick34. For example, although the localised peak in sea-surface temperatures were essentially identical between 2016 (30.1 °C; Fig. 1a) and 2019 (30.2 °C; Fig. 1b), and regional surface heating metrics and warnings were similar33, markedly different heat accumulation occurred due to the different duration that temperatures remained above the putative coral bleaching threshold (MMM + 1 = 29.8 °C; up to 2 days in April 2016 compared to up to 11 days in April 2019; Table S1). Yet these results from local SST—of similar SST maximums in 2016 and 2019, but longer durations above the threshold in 2019—only capture some of the significant differences that led to constating MHW severity and ecological outcomes between years and across depths.Daily average temperatures measured in situ at reef level in water depths of 10–40 m over ~15 years are well correlated with daily SST (r2 = 0.94–0.78 at 10–40 m). However, the strength of the relationship between SSTs and in situ temperatures declines with increasing water depth, even when in situ temperatures are averaged to a daily resolution31,35. The potential for subsurface attenuation of heating over coral reefs has previously been demonstrated using high-resolution in-situ water temperature data in the context of both regional upwelling and local internal-wave climates31,36,37, with observations of periodic transport of deeper, cooler water onto reef habitats at a large number of reefs globally31,35,38. The propagation of internal-wave energy is associated with significant vertical displacements of density isopycnals and isotherms; e.g., ~60 m displacements along the Hawaiian Ridge39. Upon encountering a sloping bottom, internal-wave dynamics become complex, and, for habitats on the fore reef slope, typically result in rapid, periodic cooling (rather than oscillations around a mean temperature) as water masses associated with e.g., 24–27 °C isotherms are vertically advected onto the reef and recede again31,40,41. In deeper reef habitats there may also be periodic heating associated with exposure to warmer surface water masses when internal waves lead to downward displacement of isotherms31, but the overall magnitude of any resulting net heating is small across the depth range considered here (i.e., no average heating at depths of 40 m and less; Fig. S4).To separate the effects of low- and high-frequency processes driving heating across the reef slope, we used a filtering approach specifically designed and validated by Wyatt et al.31. to estimate coral reef thermal regimes without internal waves. In situ temperatures were filtered to isolate variability at frequencies higher and lower than the local inertial period (~40.0 h), effectively removing the effects of internal waves from lower frequency processes (i.e., multi-day weather patterns and seasonal effects; see Fig. S3). Contrasting the observed and filtered in-situ temperature variations (black and white lines, respectively, in Fig. 3) highlights differences in the processes driving the 2016 and 2019 subsurface MHWs. During the 2019 MHW around Moorea, the filtered, or ‘non-internal wave’ (NIW), temperatures closely resembled the observed temperatures (Fig. 3e–h, m–p) implying limited internal-wave cooling (IWC). Consistent warming across the water column was evident in 2019 and temperatures remained above the coral bleaching threshold for multiple days during early to late April (Table S1). By contrast, the 2016 MHW was characterized by temperatures remaining generally below the bleaching threshold and significant high-frequency variability indicative of IWC across depths, such that temperatures only exceed the predicted bleaching threshold for hours or less at a time (Fig. 3a–d, i–l; Table S1). The high-resolution temperature observations show that IWC was greatly reduced during 2019 (Fig. 3e–h). The power spectral density of observed temperatures, concentrated at semi-diurnal frequencies and consistent with internal-wave forcing at this location41, was significantly higher in 2016 and lower in 2019 than the average across years at 10 m depth (Fig. 4a; see inset). In deeper water at 20–40 m depths, temperature variance within the semi-diurnal frequency band increased relative to shallow depths and became more similar between the two events, such that at 40 m the semi-diurnal variability was equivalent in 2016 and 2019 (Fig. 4b–d; see insets). However, this similarity in temperature variance does not indicate an equivalent magnitude of IWC, since variability during 2019 (Fig. 3p) was around a warmer background temperature closer to the coral bleaching threshold. Extending the comparison of 2016 and 2019 to other recent local MHWs demonstrates two distinct types of events: greater IWC across reef depths during 2012, 2015, and 2016 MHWs, versus reduced IWC during the 2007, 2017 and 2019 MHWs (Fig. 5).Fig. 3: Contrasting reef-level temperature variations across depths on Moorea’s north shore during the 2016 and 2019 marine heatwaves.Panels on the left show the observed high-frequency water temperature variations (black lines, measured at 2-min intervals) during the hottest months (Apr–May) in a–d 2016 and e–h 2019 at a, e 10, b, f 20, c, g 30 and d, h 40 m depths. Right panels focus on relative variation during the heatwave peaks across the same depths: i–l* 06–12 Apr 2016 and m–p 11–17 Apr 2019. Non-internal-wave temperature variations are shown based on observed temperatures filtered to remove the high-frequency influence of internal waves (white lines). The satellite-derived sea-surface temperatures (SST; grey line) are shown for comparison to in situ temperatures. The horizontal dashed line shows the ‘bleaching threshold’ (maximum monthly mean + 1 °C) and the background shading provides a reference relative to temperatures above (red), equal (yellow) and below (blue) this threshold. The red dashed squares denote the axis limits in the right panels. *Note: 40 m logger during 2016 incorrectly recorded at 2-h interval.Full size imageFig. 4: Reduced semi-diurnal temperature variability during the summer of 2019 in shallower water on the north shore of Moorea.Power spectral density (PSD) plots (logarithmic scale) were computed within a 12-day window at a 10 m, b 20 m, c 30 m, and d 40 m water depths during the summer months (Dec–May) in 2016 (blue), 2019 (red), and 2004–2018 (black; excluding 2016 and 2019). Shading shows the 95% confidence intervals for each PSD. The tidal constituents (dotted lines) show variance consistent with semi-diurnal (M2) forcing across depths, with diurnal (K1) forcing in 10 m of water along with some variability consistent with the shallow water lunar overtide (M4). Insets show details of semi-diurnal differences at each depth.Full size imageFig. 5: Comparison of internal-wave cooling (IWC) across depths during recent surface marine heatwaves (MHWs) around Moorea.Based on average a daily temperature variance (in °C2) and b IWC ((overline{{{{{{rm{IWC}}}}}}}) in °C) during the six recent local MHWs (see Table S4 for dates) that can be grouped into: (1) high IWC events during 2012 (blue), 2015 (green) and 2016 (purple); and, (2) low IWC events that coincided with bleaching events in 2007 (red) and 2019 (orange), along with early 2017 (yellow). The daily variance and (overline{{{{{{rm{IWC}}}}}}}) during Dec–Apr across all years (2005–2019) is shown for reference (black dashed lines). Contours of c–f heat accumulation as degree heating days (DHD in °C days) and g–j degree cooling days due to internal waves (DCDIW in °C days) across depths are shown for 2007, 2012, 2016 and 2019. Due to data gaps in the in situ records, contours are not shown for the 2015 or 2017 events.Full size imageThe magnitude of temperature fluctuations produced by IWC, i.e., occurring at the semi-diurnal frequency, are of similar magnitude to long-term ocean warming and climate change threatening coral reefs globally. The average IWC ((overline{{{{{{rm{IWC}}}}}}})) during the high-IWC MHWs (2012, 2015 and 2016) was between 0.14 and 0.60 °C (Fig. 3a; Table S4) and comparable to the overall SST increase measured over tropical coral reefs during the last four decades (~0.65 °C18). As a result of the subsurface cooling caused by internal waves, the 2016 MHW, which was moderate at the surface (5.7 °C-days), was mild at 10 m ( More

Field approachOur searches uncovered 301 papers reporting field studies. After screening the titles abstracts, and full texts, we kept 130 articles for the analysis (Supplementary Fig. 1), from which we obtained 1342 observations regarding 57 Cx. mosquito species from 28 countries and 135 localities (Fig. 1A). Of these 1342 observations, 733 (54.61%) were classified as high quality, (i.e., the number of individuals tested was specified) (Supplementary Tables 1 and 2). The best represented countries were the USA (64.7%, number of observations = 869), Italy (9.3%, n = 125), and Iran (2.9%, n = 39). Based on mosquito field surveillance and individuals testing positive, we concluded that JES is distributed mainly in the Nearctic, Palearctic and Oriental regions (Fig. 1A).Figure 1(A) Weighted Minimum Infection Rates and (B) Weighted Transmission Efficiency of mosquito populations for JES. High-quality data. The size of circles represents the magnitude of the estimates. Map was generated using R software version 4.1.2 with the packages mapdata, maps and tydiverse (https://www.r-project.org) and edited with Inkscape (https://inkscape.org/es/).Full size imageWest Nile virusWNV was detected mainly in the USA (76.5%, number of observations = 826), Italy (4.9%, n = 53) and Iran (3.6%, n = 39) (Fig. 1A). We also recorded 23 species (57%, 41 species) interacting with this virus (Supplementary Tables 1 and 3). Cx. quinquefasciatus became naturally infected in North America [Infection Frequency (IF) = 2.33] (Table 1). We recorded WNV interacting with Cx. tritaeniorhynchus in Asia (IF = 1.02), with Cx. pipiens in Europe (IF = 1.74) and with Cx. antennatus, Cx. neavei, Cx. perexiguus, Cx. perfuscus, Cx. poicilipes, Cx. quinquefasciatus and Cx. tritaeniorhynchus in Africa (IF = 1) (Supplementary Table 3).Table 1 Description of the variables.Full size tableThe highest infection rates were found in North America in Cx. restuans [Standardized minimum infection rate (SMIR) = 56.01], and in Africa and Europe in Cx. pipiens (SMIR = 20.45 and 29.25, respectively). No positive SMIR values were reported in Asian mosquitoes, and Oceanic mosquitoes were not sampled for this virus (Fig. 2A and Supplementary Table 3). The highest infection risk or potential was recorded in species from the USA, such as Cx. restuans (Infection Risk (IR) = 69.50), Cx. pipiens (IR = 55) and Cx. tarsalis (IR = 52.16) (Fig. 3A and Supplementary Table 3). Finally, WNV lineage 1 was detected in Algeria, Turkey, Portugal, Mexico, Tunisia, Iran, Spain and Italy, lineage 2 in Italy, Bulgaria, Greece and the Czech Republic, and lineage 5 in India (Supplementary Table 2).Figure 2(A) Box plots for the Weighted Minimum Infection Rates and (B) Weighted Transmission Efficiency Rates for JES. Boxes indicate 2nd and 3rd quartiles, vertical lines upper and lower quartiles, and horizontal lines the median. Points indicate outliers. The Y axis was transformed to Sqrt (Square root).Full size imageFigure 3JES (A) Infection Risk and (B) Transmission Risk by mosquito species.Full size imageJapanese encephalitis virusJEV was detected mainly in Taiwan (21.6%, n = 24), Korea (18%, n = 20) and Australia (15.3%, n = 17) (Fig. 1A). We found 23 mosquito species interacting with JEV: Cx. vishnui was the one most frequently found to be positive (IF = 1.20), followed by Cx. tritaeniorhynchus (IF = 1.17), Cx. pipiens and Cx. annulus (IF = 0.98) in Asia, while the most susceptible species in Oceania were Cx. sitiens and Cx. gelidus (IF = 0.71) (Supplementary Table 3).The highest SMIR values were recorded in Asia in Cx. rubithoracis (62.38), Cx. annulus (47.68) and Cx. tritaeniorhynchus (28.16) (Fig. 2A and Supplementary Table 3), and Cx. fuscocephala had the highest estimated natural IR (Fig. 3A, Supplementary Table 3). Three genotypes were recorded: genotype I (strain VNKT/479/2007, VNKT/486/2007, and JEV Ishikawa12), genotype III (Tibet-Culex-JEV1-5), and genotype V (K12YJ1174). These were isolated in China, Vietnam, and Japan (Genotype I), Italy, China (Genotype III), and Korea (Genotype V) (Supplementary Table 2).Usutu virusField studies on USUV have been conducted in Europe, Africa, and Asia, most of them in Italy (66.6%, n = 72), Czechoslovakia (11.1%, n = 12) and Slovakia (7.4%, n = 8) (Fig. 1A). Six species were reported to be susceptible to natural infection. Cx. perexiguus had the highest IF and SMIR (1.30) (Supplementary Table 3). In Africa, Cx. antennatus (IF = 1), and in Asia Cx. pipiens (IF = 1) were the most likely to be positive, while Cx. pipiens had the highest IR (5.19) (Fig. 3A, Supplementary Table 3). The recorded strains were USU181_09/USU090-10/USU173_09 (Italy) and USU/Croatia/Zagreb-102/2018 (Italy).St. Louis encephalitis virusThe field studies on SLEV focused on North America (97.7%, n = 43) and Brazil (2.2%, n = 1). Three species were recorded interacting with this virus. Cx. erraticus had the highest IF (2.06), SMIR (2.06) and IR, followed by Cx. quinquefasciatus (North America) (IF = 0.73, SMIR = 1.97) (Fig. 2A, Supplementary Table 3).The highest estimated IR of JES was for Cx. pipiens (Europe), which can be naturally infected with WNV and USUV, followed by Cx. quinquefasciatus (North America), which can be infected with WNV and SLEV (Fig. 3A).Experimental approachExperimental studies were reported in 481 articles. After screening the titles, abstracts, and full texts, as well as opportunistic records, 95 articles remained for the analysis (Supplementary Fig. 2). From these we obtained 189 high quality observations of the TE of JES in 11 countries, 40 localities, and 21 species (Fig. 1B, Supplementary Table 1). The USA was the best represented country (54.4%, n = 103), followed by Germany (13.2%, n = 25) and Australia (12.6%, n = 24). There was, however, a notable lack of information on the vector competence of Cx. mosquitos for JES in many regions of the world, such as Central and South America, and Africa (Fig. 1B).The most common means of infection was oral (94.8%, 395 observations), while the rest were intrathoracic. Intrathoracic infection bypasses the midgut barrier so is not considered natural infection. We therefore carried out the subsequent analyses using only the data on oral infection (Supplementary Table 4).We used a generalised linear model (GLM) for the statistical analysis, which was conducted only on the WNV dataset (strain NY99), the only one with sufficient observations for the purpose (n = 63). We did not find a significant effect of viral titre, temperature, or days post infection on TE. However, more data with a wide range of values is necessary to confirm these observations. On the other hand, we found that the Extrinsic Incubation Period (as DPI) was shorter at higher temperatures (Fig. 4 and Supplementary Table 5).Figure 4Relationship between temperature and Days Post Infection for WNV strain NY99.Full size imageWest Nile virusMosquito populations from many locations on all continents have been studied for their vector competence for this virus, particularly in the USA (60.3% of observations, n = 96), Germany (15.7%, n = 25) and Australia (6.9%) (Fig. 1B). Our bibliographic research revealed 21 species of Cx. with the ability to transmit WNV under laboratory conditions (Supplementary Table 6). Cx. pipiens (North America) and Cx. tarsalis were the most frequently studied species and were the most efficient in transmitting the virus (Transmission Frequency (TF) = 2.33) (Table1). Cx. quinquefasciatus had the highest TF (1.70) in Africa, Cx. modestus in Europe (TF = 1.32), and Cx. annulirostris and Cx. quinquefasciatus in Oceania (TF = 1.48) (Supplementary Table 6).Concerning Standardized Transmission Rates (STE) estimates, Cx. quinquefasciatus had the highest values in the USA (STE = 1.63), Cx. pipiens in Europe (0.90), Cx. tritaeniorhynchus in Asia (1.8), Cx. neavei in Africa (0.17) and Cx. annulirostris in Oceania (2.45) (Fig. 2B, Supplementary Table 6). We found 20 different strains of WNV tested. The TE of the various WNV strains vary considerably, but lineage 1 was more efficient than lineage 2. There were also more studies on the lineage 1 strains (n = 11), which exhibited high variation (Fig. 5).Figure 5Box plots for WNV (A) lineages and (B) strains used to measure Weighted Transmission Rates.Full size imageJapanese encephalitis virusJEV has been studied mainly in mosquito populations from France (45%, n = 20) and Australia (34%, n = 15), but also the United Kingdom, India, Taiwan, New Zealand, and the USA (Fig. 1B). Six mosquito species are capable of transmitting JEV. Cx. pipiens (Europe) had the highest TF (1.85), while Cx. gelidus had high values of STE (1.73) (Fig. 3B and Supplementary Table 6).St. Louis encephalitis virusVector competence for SLEV has been studied in two countries: the USA (93.3%, n = 42) and Argentina (6.6%, n = 3), and 7 mosquito species have been investigated. Cx. nigripalpus was the most efficient in transmitting the virus (TF = 1.60), while Cx. pipiens had the highest STE (0.68) (Figs. 2B, 3B and Supplementary Table 6).Usutu virusStudies have also been conducted on the Usutu virus in mosquito populations in the USA (28.57%, n = 4), the United Kingdom (42.8%, n = 6) and Senegal (25%, n = 4), in particular on Cx. neavei, Cx. pipiens and Cx. quinquefasciatus (TF = 1). Cx. neavei also had the highest STE (0.79) (Fig. 3B).We found reports of JES transmission under laboratory conditions in 22 Cx. species, and natural infections in 32 species (55.1% of the total sample) in the field. Cx. pipiens complex (biotypes quinquefasciatus, pipiens, molestus and pallens) was the most common vector accounting for 36.9% (n = 660) of the experimental observations and 25.7% (n = 1342) of the field observations. With both approaches, WNV was the most common flavivirus, accounting for 80.4% of the field observations and 86.7% of the experimental data (Fig. 1A,B). Only WNV, therefore, had enough observations to make comparison between the experimental and field data possible. We were able to compare 16 mosquito species and found a high positive correlation between TF and IF (R = 0.57, p = 0.02) (Fig. 7).In summary, we found that the species with the highest infection-transmission risk (IRT) for WNV was Cx. restuans, for USUV it was Cx. pipiens (Europe), for SLEV Cx. quinquefasciatus (North America), and for JEV Cx. gelidus (Oceania) (Fig. 6 and Supplementary Tables 2 and 6).Figure 6JES infection-transmission risk by continent and flavivirus.Full size image More

Feit, B. et al. Landscape complexity promotes resilience of biological pest control to climate change. Proc. R. Soc. B. 288, 20210547 (2021).Article 

Google Scholar 
Hall-Spencer, J. M. & Harvey, B. P. Ocean acidification impacts on coastal ecosystem services due to habitat degradation. Emerg. Top. Life Sci. 3, 197–206 (2019).Article 
CAS 

Google Scholar 
Loke, L. H. L. & Todd, P. A. Structural complexity and component type increase intertidal biodiversity independently of area. Ecology 97, 383–393 (2016).Article 

Google Scholar 
Oliver, T. H. et al. Biodiversity and resilience of ecosystem functions. Trends in Ecol. Evol. 30, 673–684 (2015).Article 

Google Scholar 
Bullock, J. M. et al. Future restoration should enhance ecological complexity and emergent properties at multiple scales. Ecography ecog. 4, 05780 (2022).Ortega, J. C. G., Thomaz, S. M. & Bini, L. M. Experiments reveal that environmental heterogeneity increases species richness, but they are rarely designed to detect the underlying mechanisms. Oecologia 188, 11–22 (2018).Article 

Google Scholar 
Griffin, J. N., Byrnes, J. E. K. & Cardinale, B. J. Effects of predator richness on prey suppression: a meta-analysis. Ecology 94, 2180–2187 (2013).Article 

Google Scholar 
Katano, I., Doi, H., Eriksson, B. K. & Hillebrand, H. A cross-system meta-analysis reveals coupled predation effects on prey biomass and diversity. Oikos 124, 1427–1435 (2015).Article 

Google Scholar 
Loke, L. H. L., Ladle, R. J., Bouma, T. J. & Todd, P. A. Creating complex habitats for restoration and reconciliation. Ecol. Eng. 77, 307–313 (2015).Article 

Google Scholar 
Torres-Pulliza, D. et al. A geometric basis for surface habitat complexity and biodiversity. Nat. Ecol. Evol. 4, 1495–1501 (2020).Article 

Google Scholar 
Chesson, P. Mechanisms of maintenance of species diversity. Annu. Rev. Ecol. Syst. 31, 343–366 (2000).Article 

Google Scholar 
Chesson, P. & Kuang, J. J. The interaction between predation and competition. Nature 456, 235–238 (2008).Article 
CAS 

Google Scholar 
Terborgh, J. W. Toward a trophic theory of species diversity. Proc. Natl. Acad. Sci. USA 112, 11415–11422 (2015).Article 
CAS 

Google Scholar 
Pringle, R. M. et al. Predator-induced collapse of niche structure and species coexistence. Nature 570, 58–64 (2019).Article 
CAS 

Google Scholar 
Sandom, C. et al. Mammal predator and prey species richness are strongly linked at macroscales. Ecology 94, 1112–1122 (2013).Article 

Google Scholar 
Grabowski, J. H. Habitat complexity disrupts predator-prey interactions but not the trophic cascade on oyster reefs. Ecology 85, 995–1004 (2004).Article 

Google Scholar 
Crowder, L. B. & Cooper, W. E. Habitat structural complexity and the interaction between bluegills and their prey. Ecology 63, 1802 (1982).Article 

Google Scholar 
Almany, G. R. Does increased habitat complexity reduce predation and competition in coral reef fish assemblages? Oikos 106, 275–284 (2004).Article 

Google Scholar 
Anderson, T. L. & Semlitsch, R. D. Top predators and habitat complexity alter an intraguild predation module in pond communities. J. Anim. Ecol. 85, 548–558 (2016).Article 

Google Scholar 
Brothers, C. A. & Blakeslee, A. M. H. Alien vs predator play hide and seek: How habitat complexity alters parasite mediated host survival. J. Exp. Mar. Biol. Ecol. 535, 151488 (2021).Article 

Google Scholar 
Horinouchi, M. et al. Seagrass habitat complexity does not always decrease foraging efficiencies of piscivorous fishes. Mar. Ecol. Prog. Ser. 377, 43–49 (2009).Article 

Google Scholar 
Ryer, C., Stoner, A. & Titgen, R. Behavioral mechanisms underlying the refuge value of benthic habitat structure for two flatfishes with differing anti-predator strategies. Mar. Ecol. Prog. Ser. 268, 231–243 (2004).Article 

Google Scholar 
Flynn, A. J. & Ritz, D. A. Effect of habitat complexity and predatory style on the capture success of fish feeding on aggregated prey. J. Mar. Biol. Ass. 79, 487–494 (1999).Article 

Google Scholar 
Klecka, J. & Boukal, D. S. The effect of habitat structure on prey mortality depends on predator and prey microhabitat use. Oecologia 176, 183–191 (2014).Article 

Google Scholar 
James, P. L. & Heck, K. L. The effects of habitat complexity and light intensity on ambush predation within a simulated seagrass habitat. J. Exp. Mar. Biol. Ecol. 176, 187–200 (1994).Article 

Google Scholar 
Michel, M. J. & Adams, M. M. Differential effects of structural complexity on predator foraging behavior. Behav. Ecol. 20, 313–317 (2009).Article 

Google Scholar 
Preisser, E. L., Bolnick, D. I. & Benard, M. F. Scared to death? The effects of intimidation and consumption in predator-prey interactions. Ecology 86, 501–509 (2005).Article 

Google Scholar 
Preisser, E. L., Orrock, J. L. & Schmitz, O. J. Predator hunting mode and habitat domain alter nonconsumptive effects in predator-prey interactions. Ecology 88, 2744–2751 (2007).Article 

Google Scholar 
Rypstra, A. L., Schmidt, J. M., Reif, B. D., DeVito, J. & Persons, M. H. Tradeoffs involved in site selection and foraging in a wolf spider: effects of substrate structure and predation risk. Oikos 116, 853–863 (2007).Article 

Google Scholar 
Janssen, A., Sabelis, M. W., Magalhães, S., Montserrat, M. & van der Hammen, T. Habitat structure affects intraguild predation. Ecology 88, 2713–2719 (2007).Article 

Google Scholar 
Grabowski, J. H., Hughes, A. R. & Kimbro, D. L. Habitat complexity influences cascading effects of multiple predators. Ecology 89, 3413–3422 (2008).Article 

Google Scholar 
Hughes, A. R. & Grabowski, J. H. Habitat context influences predator interference interactions and the strength of resource partitioning. Oecologia 149, 256–264 (2006).Article 

Google Scholar 
Bonett, D. G. Meta-analytic interval estimation for standardized and unstandardized mean differences. Psychol. Methods 14, 225–238 (2009).Article 

Google Scholar 
Huey, R. B. & Pianka, E. R. Ecological consequences of foraging mode. Ecology 62, 991–999 (1981).Article 

Google Scholar 
Egger, M., Smith, G. D., Schneider, M. & Minder, C. Bias in meta-analysis detected by a simple, graphical test. BMJ 315, 629–634 (1997).Article 
CAS 

Google Scholar 
Ritchie, E. G. & Johnson, C. N. Predator interactions, mesopredator release and biodiversity conservation. Ecol. Lett. 12, 982–998 (2009).Article 

Google Scholar 
Chaplin-Kramer, R., O’Rourke, M. E., Blitzer, E. J. & Kremen, C. A meta-analysis of crop pest and natural enemy response to landscape complexity: pest and natural enemy response to landscape complexity. Ecol. Lett. 14, 922–932 (2011).Article 

Google Scholar 
Paxton, A. B. et al. Meta-analysis reveals artificial reefs can be effective tools for fish community enhancement but are not one-size-fits-all. Front. Mar. Sci. 7, 282 (2020).Article 

Google Scholar 
Eggleston, D. B., Lipcius, R. N., Miller, D. L. & Coba-Cetina, L. Shelter scaling regulates survival of juvenile Caribbean spiny lobster Panulirus argus. Mar. Ecol. Prog. Ser. 62, 79–88 (1990).Rogers, A., Blanchard, J. L. & Mumby, P. J. Fisheries productivity under progressive coral reef degradation. J. Appl. Ecol. 55, 1041–1049 (2018).Article 

Google Scholar 
Gontijo, L. M. Engineering natural enemy shelters to enhance conservation biological control in field crops. Biol. Control 130, 155–163 (2019).Article 

Google Scholar  More

MAWF. National Rangeland Management Policy & Strategy. Restoring Namibia’s Rangelands (2012).SAIEA. Strategic environmental assessment of large-scale bush thinning and value-addition activities in Namibia. (2016).de Klerk, J. N. Bush Encroachment in Namibia. Report on Phase 1 of the Bush Encroachment Research, Monitoring and Management Project. (2004).Muntifering, J. R. et al. Managing the matrix for large carnivores: A novel approach and perspective from cheetah (Acinonyx jubatus) habitat suitability modelling. Anim. Conserv. 9, 103–112 (2006).Article 

Google Scholar 
Marker, L. L., Dickman, A. J., Mills, M. G. L., Jeo, R. M. & Macdonald, D. W. Spatial ecology of cheetahs on north-central Namibian farmlands. J. Zool. 274, 226–238 (2008).Article 

Google Scholar 
Wang, J. et al. Impacts of juniper woody plant encroachment into grasslands on local climate. Agric. For. Meteorol. 307, 108508 (2021).Article 
ADS 

Google Scholar 
Shen, X. et al. Effect of shrub encroachment on land surface temperature in semi-arid areas of temperate regions of the Northern Hemisphere. Agric. For. Meteorol. 320, 108943 (2022).Article 
ADS 

Google Scholar 
Shen, X. et al. Vegetation greening, extended growing seasons, and temperature feedbacks in warming temperate grasslands of China. J. Clim. 35, 5103–5117 (2022).Article 
ADS 

Google Scholar 
Martins, A. R. O. & Shackleton, C. M. Population structure and harvesting selection of two palm species ( Hyphaene coriacea and Phoenix reclinata ) in Zitundo area, southern Mozambique. For. Ecol. Manage. 398, 64–74 (2017).Article 

Google Scholar 
Brown, G. W., Murphy, A., Fanson, B. & Tolsma, A. The influence of different restoration thinning treatments on tree growth in a depleted forest system. For. Ecol. Manage. 437, 10–16 (2019).Article 

Google Scholar 
Belsky, A. Influences of trees on savanna productivity: Tests of shade, nutrients and grass–tree competition. Ecology 75, 922–932 (1994).Article 

Google Scholar 
Hagos, M. G. & Smit, G. N. Soil enrichment by Acacia mellifera subsp. detinens on nutrient poor sandy soil in a semi-arid southern African savanna. J. Arid Environ. 61, 47–59 (2005).Ludwig, F., Kroon, H. D., Berendse, F. & Prins, H. H. T. The influence of savanna trees on nutrient, water and light availability and the understorey vegetation. Plant Ecol. 170, 93–105 (2004).Article 

Google Scholar 
Ridolfi, L., Laio, F., D’Odorico, P. & D’Odorico, P. Fertility island formation and evolution in dryland ecosystems. Ecol. Soc. 13, 13 (2008).Article 
MATH 

Google Scholar 
Wiegand, K., Ward, D. & Saltz, D. Multi-scale patterns and bush encroachment in an arid savanna with a shallow soil layer. J. Veg. Sci. 16, 311–320 (2005).Article 

Google Scholar 
Burke, A. Savanna trees in Namibia – Factors controlling their distribution at the arid end of the spectrum. Flora Morphol. Distrib. Funct. Ecol. Plants 201, 181–201 (2006).
Google Scholar 
Buyer, J. S., Schmidt-Küntzel, A., Nghikembua, M., Maul, J. E. & Marker, L. Soil microbial communities following bush removal in a Namibian savanna. Soil 2, 101–110 (2016).Article 
ADS 
CAS 

Google Scholar 
Dwivedi, V. & Soni, P. A review on the role of soil microbial biomass in eco-restoration of degraded ecosystem with special reference to mining areas. J. Appl. Nat. Sci. 151–158 (2011). https://doi.org/10.31018/jans.v3i1.173.Smit, G. N. An approach to tree thinning to structure southern African savannas’ for long-term restoration from bush encroachment. J. Environ. Manage. 71, 179–191 (2004).Article 
CAS 

Google Scholar 
MAWF. Forestry and environmental authorisation process for bush harvesting projects. 34 (2017).Smit, G. N., de Klerk, J. N., Schneider, M. B. & van Eck, J. Detailed assessment of the biomass resource and potential yield in a selected bush encroached area of Namibia. 141 (2015).Marker, L. et al. and Distribution. in Cheetahs: Biology and Conservation: Biodiversity of the World: Conservation from Genes to Landscapes (eds. Marker, L., Boast, L. k. & Schmidt-Küntzel, A.) 9–20 (John Fedor, 2018). https://doi.org/10.1016/B978-0-12-804088-1.00004-6.NAPHA. Namibia Professional Hunting Association. http://www.napha-namibia.com/conservation/huntable-species/carnivora/ (2015).SWA. Nature Conservation Ordinance, 1975 (No. 4 of 1975). vol. 1975 (1975).Marker, L. et al. The status of Key pre species and the Consequences of Prey Loss for Cheetah Conservation in North and West Africa. in Cheetahs: Biology and Conservation: Biodiversity of the World: Conservation from Genes to Landscapes. (eds. Marker, L., Boast, L. k. & Schmidt-Küntzel, A.) 151–161 (John Fedor, 2018). https://doi.org/10.1016/B978-0-12-804088-1.00004-6.Kiruki, H. M., van der Zanden, E. H., Gikuma-Njuru, P. & Verburg, P. H. The effect of charcoal production and other land uses on diversity, structure and regeneration of woodlands in a semi-arid area in Kenya. For. Ecol. Manage. 391, 282–295 (2017).Article 

Google Scholar 
Harmse, C. J., Kellner, K. & Dreber, N. Restoring productive rangelands: A comparative assessment of selective and non-selective chemical bush control in a semi-arid Kalahari savanna. J. Arid Environ. 135, 39–49 (2016).Article 
ADS 

Google Scholar 
Nghikembua, M. T. et al. Response of wildlife to bush thinning on the north central freehold farmlands of Namibia. For. Ecol. Manage. 473, 118330 (2020).Article 

Google Scholar 
Soto-Shoender, J. R., McCleery, R. A., Monadjem, A. & Gwinn, D. C. The importance of grass cover for mammalian diversity and habitat associations in a bush encroached savanna. Biol. Conserv. 221, 127–136 (2018).Article 

Google Scholar 
Strohbach, B. J. Environmental information service, Namibia for the Ministry of Environment and Tourism, the Namibian Chamber of Environment and the Namibia University of Contribution to the knowledge of southern African Lepismatidae. Namibian J. Environ. 8327, 14–33 (2017).
Google Scholar 
Smit, N. BECVOL 3: An expansion of the aboveground biomass quantification model for trees and shrubs to include the wood component. Afr. J. Range Forage Sci. 31, 179–186 (2014).Article 

Google Scholar 
Zimmerman, I. Causes and Consequences of Fenceline Contrasts in Namibia. (Free State, Bloemfontein, South Africa, 2009).Dwyer, J. M. & Mason, R. Plant community responses to thinning in densely regenerating Acacia harpophylla forest. Restor. Ecol. 26, 97–105 (2018).Article 

Google Scholar 
Thomas, S. C. & Martin, A. R. Carbon content of tree tissues: A synthesis. Forests 3, 332–352 (2012).Article 

Google Scholar 
Chave, J. et al. Improved allometric models to estimate the aboveground biomass of tropical trees. Glob. Chang. Biol. 20, 3177–3190 (2014).Article 
ADS 

Google Scholar 
Djomo, A. N. & Chimi, C. D. Tree allometric equations for estimation of above, below and total biomass in a tropical moist forest: Case study with application to remote sensing. For. Ecol. Manage. 391, 184–193 (2017).Article 

Google Scholar 
Ganamé, M., Bayen, P., Dimobe, K., Ouédraogo, I. & Thiombiano, A. Aboveground biomass allocation, additive biomass and carbon sequestration models for Pterocarpus erinaceus Poir. in Burkina Faso. Heliyon 6, (2020).Picard, N., Saint-André, L. & Henry, M. Manual for building tree volume and biomass allometric equations: From field measurement to prediction. Food and Agricultural Organization of the United Nations, Rome, and Centre de Coopération Internationale en Recherche Agronomique pour le Développement. (2012).Feyisa, K. et al. Allometric equations for predicting above-ground biomass of selected woody species to estimate carbon in East African rangelands. Agrofor. Syst. 92, 599–621 (2018).Article 

Google Scholar 
Boys, J. M. & Smit, N. G. Development of an Excel Based Bush Biomass Quantification Tool. (2020).Ngomanda, A. et al. Site-specific versus pantropical allometric equations: Which option to estimate the biomass of a moist central African forest?. For. Ecol. Manage. 312, 1–9 (2014).Article 

Google Scholar 
CCF. Cheetah Conservation Fund Bush PTY (Ltd). https://bushblok.com/management-plan (2019).Wykstra, M. et al. Improved and Alternative Livelihoods: Links Between Poverty Alleviation, Biodiversity, and Cheetah Conservation. in Cheetahs: Biology and Conservation: Biodiversity of the World: Conservation from Genes to Landscapes (eds. Marker, L., Boast, L. k. & Schmidt-Küntzel, A.) 223–237 (John Fedor, 2018).Zimmermann, I. et al. The influence of two levels of debushing in Namibia’s Thornbush Savanna on overall soil fertility, measured through bioassays. Namibian J. Environ. 1, 52–59 (2017).
Google Scholar 
Nghikembua, M. T. et al. Restoration thinning reduces bush encroachment on freehold farmlands in north-central Namibia. For. An Int. J. For. Res. 1–14 (2021). https://doi.org/10.1093/forestry/cpab009.Fick, S. E. & Hijmans, R. J. WorldClim 2: new 1-km spatial resolution climate surfaces for global land areas. Int. J. Climatol. 37, 4302–4315 (2017).Article 

Google Scholar 
Mendelsohn, J., Jarvis, A., Roberts, C. & Robertson, T. Atlas of Namibia: A portrait of the land and its people. (2003).Curtis, B. A. & Mannheimer, C. A. Tree Atlas of Namibia; National Botanical Research Institute (NBRI). (National Botanical Research Institute, 2005).Mannheimer, C. & Curtis, B. Le Roux and Muller’s Field Guide to the Trees & Shrubs of Namibia. (Macmillan Education Namibia (PTY) LTD, 2009).Honsbein, D., Shiningavamwe, K., Iikela, J. & de la Puerta Fernandez, Maria, L. Animal Feed from Namibian Encroacher Bush. (2017).Coates Palgrave, K. Trees of Southern Africa. (Struik publishers, 1993).Nghikembua, M., Harris, J., Tregenza, T. & Marker, L. Spatial and temporal habitat use by GPS collared male cheetahs in modified bushland habitat. Open J. For. 06, 269–280 (2016).
Google Scholar 
Venables, W. N. & Ripley, B. D. Modern Applied Statistics with S. (Springer, 2002).Bates, D., Mächler, M., Bolker, B. & Walker, S. Fitting linear mixed-effects models using lme4. J. Stat. Softw. 67, 1–48 (2015).Article 

Google Scholar 
Mehtatalo, L. & Lappi, J. Forest biometrics with examples in R. (Taylor & Francis Inc, 2020).R Core Team. R: A language and environment for statistical computing (2021).Birch, C. & Middleton, A. Economics of Land Degradation Related To Bush Encroachment in Namibia. (2017).Neke, K. S., Owen-Smith, N. & Witkowski, E. T. F. Comparative resprouting response of Savanna woody plant species following harvesting: the value of persistence. For. Ecol. Manage. 232, 114–123 (2006).Article 

Google Scholar 
Smit, N. Response of Colophospermum mopane to different intensities of tree thinning in the Mopane Bushveld of southern Africa. Afr. J. Range Forage Sci. 31, 173–177 (2014).Article 

Google Scholar 
DAS. Bush Control Manual. (John Meinert Printing, 2017).Leinonen, A. Wood chip production technology and costs for fuel in Namibia. VTT Tiedotteita – Valtion Teknillinen Tutkimuskeskus (2007).Chakanga, M. A preliminary analysis of the Economic Plots Research Data of the CCF Bush Project. (2003).West, P. W. Thinning. in Growing Plantation Forests vol. 9783319018 115–129 (Springer International Publishing, 2014).Dwyer, J. M., Fensham, R. & Buckley, Y. M. Restoration thinning accelerates structural development and carbon sequestration in an endangered Australian ecosystem. J. Appl. Ecol. 47, 681–691 (2010).Article 

Google Scholar 
Groengroeft, A., de Blécourt, M., Classen, N., Landschreiber, L. & Eschenbach, A. Acacia trees modify soil water dynamics and the potential groundwater recharge in savanna ecosystems. in Climate change and adaptive land management in southern Africa – assessments, changes, challenges, and solutions (ed. (eds. Revermann, R. et al.) 177–186 (Klaus Hess Publishers, 2018).Richter, C. G. F., Snyman, H. A. & Smit, G. N. The influence of tree density on the grass layer of three semi-arid savanna types of southern africa. Afr. J. Range Forage Sci. 18, 103–109 (2001).Article 

Google Scholar 
Stafford, W. et al. The economics of landscape restoration: Benefits of controlling bush encroachment and invasive plant species in South Africa and Namibia. Ecosyst. Serv. 27, 193–202 (2017).Article 

Google Scholar 
Joubert, D. F., Smit, G. N. & Hoffman, M. T. The influence of rainfall, competition and predation on seed production, germination and establishment of an encroaching Acacia in an arid Namibian savanna. J. Arid Environ. 91, 7–13 (2013).Article 
ADS 

Google Scholar  More

Identification of windthrow eventsLandsat images from January 1st 2018 to December 31st 2019 were filtered on 20% or less of cloud coverage, and only the least cloudy image at each location was selected to make an image composite covering the entire Amazon region. In total, 395 least cloudy Landsat 8 images within the Amazon boundary during 2018–2019 were selected and displayed in false color (red: shortwave infrared band, green: near-infrared band, blue: red band) on Google Earth Engine for windthrow events identification (Supplementary Fig. 6). Hollow regions on Supplementary Fig. 6 (2.8% of the total area of the Amazon region) indicated that no clear images with 1 year before the identification were displayed in bright green colors (due to reflectance in near-infrared band from the pioneer species). “Old” windthrows account for ~80% of total identified windthrows, and they were verified using historical Landsat images that can go as far as 1984 (when Landsat 5 was launched). “Old” windthrows were validated once they were found with clear shape and more distinguish color on the historical Landsat images (Supplementary Fig. 7c). 10–15% of “old” windthrows without fan-shape were eliminated from this study because it was hard to identify if they were windthrows or other types of forest disturbance. The minimum size of windthrows identified in this study was 25,000 m2. This process generated the location and rough size of 1012 visible (both “old” and “new”) windthrow scars with fan-shaped patch, scattered small disturbance pixels tails, and an area of over 25,000 m2 (Supplementary Fig. 8). Based on a gap-size probability distribution function that simulates the entire disturbance gradient from all sizes of windthrows19, the proportion of total tree mortality represented by large windthrows ( >25,000 m2) identified in this study is 0.5–1.1%.Among 1012 visual identified windthrows, the occurrence year of 125 windthrows were identified using Landsat 5,7,8, MODIS, and TRMM dataset (Supplementary Table 2), and 38 windthrows from these 125 windthrows had clear remote sensing evidence to validate their occurring date (Supplementary Table 3). It is difficult to get the accurate year and date of occurrence of all identified windthrows. Previous studies showed that windthrows in the northwestern Amazon took ~20 years to recover to 90% of “pre-disturbance” biomass from all damage classes while forests in the central Amazon took ~40 years to recover40. The biomass recovery depends on the windthrow severity and time since disturbance33. Based on the recovery time (20–40 years) and the time of windthrow identification (2018–2019), we estimated that these 1012 windthrows most likely occurred within 30 years (between 1990 and 2019), and the estimated occurrence period was validated using the range of the occurrence year (1986–2019) of 125 windthrow cases.Windthrow density dataThe windthrow density shown in Fig. 1b was generated using 1012 windthrow points in QGIS45. We created a 2.5° by 2.5° grid map, and the windthrow density was calculated by counting the number of windthrows in each grid. These values were then converted to a density with units of counts of windthrows per 10,000 km2. We chose 2.5 degrees to aggregate the data to make sure that over 50% of grids have at least 1 windthrow event while still preserving the spatial distribution of mean afternoon CAPE over the Amazon. The contour lines displayed in Fig. 1c were generated using the “Contour” function on the windthrow density map in QGIS.Meteorological dataTo derive the correlation between windthrow density and meteorological variables, we used ERA 5 global reanalysis hourly CAPE on single levels from 1979 to present at 0.25° × 0.25° resolution provided by the European Center for Medium-Range Weather Forecasts. ERA 5 CAPE was computed by considering parcels of air departing at different pressure levels below the 35 kPa level, with maximum–unstable algorithm under a pseudo-adiabatic assumption46. Afternoon mean CAPE map was calculated as the average of hourly CAPE data from 17:00–23:00 UTC (13:00–19:00 local time in Amazon) over all the months between 1990 and 2019. We chose to average CAPE over 30 years because these windthrow events occurred in these 30 years and calculating the average can help capture the overall spatial pattern of CAPE and minimize the influence of interannual climate variability on windthrow events.To project future windthrow density in the Amazon for the end of the 21st century, we analyzed meteorological output from 10 ESMs that participated in CMIP6 (https://www.wcrp-climate.org/wgcm-cmip/wgcm-cmip6). The models used in this research were listed in Supplementary Table 1. We extracted daily surface temperature (tas), specific humidity (huss), surface pressure (ps), temperature (ta) from these models to calculate daily nondilute, near-surface-based, adiabatic CAPE. CMIP6 CAPE was calculated by considering the buoyancy of a near-surface parcel lifted adiabatically to a series of discrete pressure levels (100 kPa to 10 kPa in increments of 10 kPa). CMIP6 CAPE is calculated as follows:$${CAPE}=mathop{sum }limits_{i=1}^{10}{{{{{rm{d}}}}}}p{{{{{rm{H}}}}}}({b}_{i}){b}_{i}$$
(1)
Where ({{{{{rm{d}}}}}}p) = 10 kPa, H is the Heaviside unit step function, and ({b}_{i}=frac{1}{{rho }_{i}}-frac{1}{{rho }_{e,i}}), with ({rho }_{i}) being the parcel density at pressure level i and ({rho }_{e,i}) being the environmental density at pressure level i.The future projections in our analysis were based on SSP585, a high-emission scenario with high radiative forcing by the end of the century. We calculated mean daily CAPE over 1990–2015 as current CMIP6 CAPE and mean daily CAPE over 2070–2099 as future CMIP6 CAPE. Since different approaches were used to calculate ERA 5 CAPE and CMIP6 CAPE47, the absolute CAPE values of the two datasets are not comparable. Therefore, for each ESM model, we scaled future CMIP6 CAPE by multiplying, grid-wise, the delta CAPE generated from an individual model in CMIP6 with the ERA 5 current mean afternoon CAPE (Fig. 1c) as follows:$${delta},{CAPE}=(CAPE_{CMIP6_{,future}},-CAPE_{CMIP6_current})/CAPE_{CMIP6_current}$$
(2)
$$CAP{E}_{scaled_CMIP6_,future}=(1+delta,CAPE)times CAP{E}_{ERA5}{_}_{current}$$
(3)
The delta CAPE indicated the projected increase in CAPE from 1990–2015 to the end of the 21st century. In this way, a scaled CMIP6 future CAPE map was generated for each model, and an ensemble-mean scaled CMIP6 CAPE map over 10 ESM models can be found in Supplementary Fig. 5b. The scaled CMIP6 future CAPE values were within plausible range compared to the ERA 5 current mean afternoon CAPE values, and both current and future CAPE maps were used to produce the increase in area with high CAPE values ( >1023 J kg−1) in Table 1. However, it is worth noting that the scaling with relative changes in delta CAPE (%) is more sensitive to CMIP historical baseline conditions than absolute changes of CAPE (J kg−1), which will likely introduce a larger scaled spread (min/max CAPE changes).The increase in area with storm-favorable environments was calculated as follows:$$Increase=(are{a}_{future}-area_{current})/are{a}_{current}$$
(4)
Where areacurrent is the area of CAPE  > 1023 J kg−1 for current ERA 5 CAPE, and areafuture is the area of CAPE  > 1023 J kg−1 for the scaled CMIP6 future CAPE.A model of windthrow densityWe developed a model based on the relationship between satellite-derived windthrow density and mean afternoon CAPE from the ERA 5 reanalysis over 1990–2019. The non-parametric model provides a look-up table of windthrow density as a function of CAPE within the range of observations. Counts of observed windthrow events and Amazon’s area were separately binned by CAPE using the same bins, producing two histograms of CAPE. The ratio of the former to the latter gives the density of windthrow events (windthrow events per area) as a function of CAPE. To avoid noise at the tails of the histograms, the six CAPE bins were chosen such that each bin would have about the same number of windthrow events (either 168 or 169). The total number of windthrow events is given by the sum over bins of the product of windthrow density and area. The minimum and maximum of current ERA 5 mean afternoon CAPE was 42 and 1549. The minimum CAPE value of the first bin was extended to 0 and the maximum CAPE value of the last bin was extended to infinity under the assumption that the windthrow density is similar for neighboring values. Based on the windthrow density and CAPE relationship used in the model, it is the increase in the area with high CAPE that then leads to an increase in the number of windthrow events.It is worth noting that the future windthrow density produced by models may be underestimated because the windthrow observations within regions with high CAPE were incomplete due to high cloud coverage. Moreover, the non-parametric model makes the conservative assumption that the windthrow density does not increase at higher, as-yet unobserved values of mean afternoon CAPE.Future projections of windthrow densityWe combined the non-parametric relationship (Fig. 2a) with the future CAPE map generated from the ten CMIP6 ESMs (adjusted by ERA 5 mean CAPE values) to estimate the changes in windthrow density at the end of the 21st century. We estimated uncertainties for windthrow density projections by combining information about model-to-model differences. The analysis yielded a set of 10 estimates. The overall windthrow density increase and uncertainty were estimated using the mean increase and one standard deviation from the ensemble of the 10 models. More

Berner, L. T. et al. Summer warming explains widespread but not uniform greening in the Arctic tundra biome. Nat. Commun. 11, 4621 (2020).Article 
ADS 
CAS 

Google Scholar 
Elmendorf, S. C. et al. Plot-scale evidence of tundra vegetation change and links to recent summer warming. Nat. Clim. Change 2, 453–457 (2012).Article 
ADS 

Google Scholar 
Overland, J. E. et al. Surface air temperature. In Arctic Report Card: Update for 2019 (eds Richter-Menge, J. et al.) (U.S. National Park Service, 2020).
Google Scholar 
Post, E., Steinman, B. A. & Mann, M. E. Acceleration of phenological advance and warming with latitude over the past century. Sci. Rep. 8, 3927 (2018).Article 
ADS 

Google Scholar 
Diepstraten, R. A. E., Jessen, T. D., Fauvelle, C. M. D. & Musiani, M. M. Does climate change and plant phenology research neglect the Arctic tundra?. Ecosphere 9, e02362 (2018).Article 

Google Scholar 
Flynn, D. F. B. & Wolkovich, E. M. Temperature and photoperiod drive spring phenology across all species in a temperate forest community. New Phytol. 219, 1353–1362 (2018).Article 
CAS 

Google Scholar 
Billings, W. D. & Bliss, L. C. An alpine snowbank environment and its effects on vegetation, plant development, and productivity. Ecology 40, 388–397 (1959).Article 

Google Scholar 
Billings, W. D. & Mooney, H. A. The ecology of arctic and alpine plants. Biol. Rev. 43, 481–529 (1968).Article 

Google Scholar 
Sørensen, T. Temperature relations and phenology of the northeast Greenland flowering plants. Meddr Gronland 1–305 (1941).Barrett, R. T. & Hollister, R. D. Arctic plants are capable of sustained responses to long-term warming. Polar Res. 35, 25405 (2016).Article 

Google Scholar 
Julitta, T. et al. Using digital camera images to analyse snowmelt and phenology of a subalpine grassland. Agric. For. Meteorol. 198–199, 116–125 (2014).Article 
ADS 

Google Scholar 
Petraglia, A. et al. Responses of flowering phenology of snowbed plants to an experimentally imposed extreme advanced snowmelt. Plant Ecol. 215, 759–768 (2014).Article 

Google Scholar 
Semenchuk, P. R. et al. High Arctic plant phenology is determined by snowmelt patterns but duration of phenological periods is fixed: An example of periodicity. Environ. Res. Lett. 11, 125006 (2016).Article 
ADS 

Google Scholar 
Hollister, R. D., Webber, P. J. & Bay, C. Plant response to temperature in northern Alaska: Implications for predicting vegetation change. Ecology 86, 1562–1570 (2005).Article 

Google Scholar 
Oberbauer, S. et al. Phenological response of tundra plants to background climate variation tested using the International Tundra Experiment. Philos. Trans. R. Soc. B Biol. Sci. 368, 20120481 (2013).Article 
CAS 

Google Scholar 
Tieszen, L. L. Photosynthesis in the principal Barrow, Alaska, species: A summary of field and laboratory responses. In Vegetation and Production Ecology of an Alaskan Arctic Tundra (ed. Tieszen, L. L.) 241–268 (Springer, 1978).Chapter 

Google Scholar 
Körner, Ch. CO2 exchange in the alpine sedge Carex curvula as influenced by canopy structure, light and temperature. Oecologia 53, 98–104 (1982).Article 
ADS 

Google Scholar 
Tieszen, L. L. Photosynthesis and respiration in arctic tundra grasses: Field light intensity and temperature responses. Arct. Alp. Res. 5, 239–251 (1973).Article 
CAS 

Google Scholar 
Huang, M. et al. Air temperature optima of vegetation productivity across global biomes. Nat. Ecol. Evol. 3, 772–779 (2019).Article 

Google Scholar 
Marchand, F. L., Mertens, S., Kockelbergh, F., Beyens, L. & Nijs, I. Performance of high arctic tundra plants improved during but deteriorated after exposure to a simulated extreme temperature event. Glob. Change Biol. 11, 2078–2089 (2005).Article 
ADS 

Google Scholar 
Yan, W. An equation for modelling the temperature response of plants using only the cardinal temperatures. Ann. Bot. 84, 607–614 (1999).Article 

Google Scholar 
Zhou, G. & Wang, Q. A new nonlinear method for calculating growing degree days. Sci. Rep. 8, 10149 (2018).Article 
ADS 

Google Scholar 
Kramer, K. Selecting a model to predict the onset of growth of Fagus sylvatica. J. Appl. Ecol. 31, 172 (1994).Article 

Google Scholar 
Nakano, Y., Higuchi, Y., Sumitomo, K. & Hisamatsu, T. Flowering retardation by high temperature in chrysanthemums: Involvement of FLOWERING LOCUS T-like 3 gene repression. J. Exp. Bot. 64, 909–920 (2013).Article 
CAS 

Google Scholar 
del Olmo, I., Poza-Viejo, L., Piñeiro, M., Jarillo, J. A. & Crevillén, P. High ambient temperature leads to reduced FT expression and delayed flowering in Brassica rapa via a mechanism associated with H2A.Z dynamics. Plant J. 100, 343–356 (2019).Article 

Google Scholar 
Wolkovich, E. M. et al. Warming experiments underpredict plant phenological responses to climate change. Nature 485, 494 (2012).Article 
ADS 
CAS 

Google Scholar 
Hollister, R. D. et al. A review of open top chamber (OTC) performance across the ITEX Network. Arct. Sci. https://doi.org/10.1139/AS-2022-0030 (2022).Article 

Google Scholar 
Bütikofer, L. et al. The problem of scale in predicting biological responses to climate. Glob. Change Biol. 26, 6657–6666 (2020).Article 
ADS 

Google Scholar 
Gu, S. Growing degree hours—A simple, accurate, and precise protocol to approximate growing heat summation for grapevines. Int. J. Biometeorol. 60, 1123–1134 (2016).Article 
ADS 
CAS 

Google Scholar 
Roltsch, W. J., Zalom, F. G., Strawn, A. J., Strand, J. F. & Pitcairn, M. J. Evaluation of several degree-day estimation methods in California climates. Int. J. Biometeorol. 42, 169–176 (1999).Article 
ADS 

Google Scholar 
Richardson, A. D. et al. Ecosystem warming extends vegetation activity but heightens vulnerability to cold temperatures. Nature 560, 368–371 (2018).Article 
ADS 
CAS 

Google Scholar 
Ettinger, A. K., Buonaiuto, D. M., Chamberlain, C. J., Morales-Castilla, I. & Wolkovich, E. M. Spatial and temporal shifts in photoperiod with climate change. New Phytol. 230, 462–474 (2021).Article 
CAS 

Google Scholar 
Seyednasrollah, B., Swenson, J. J., Domec, J.-C. & Clark, J. S. Leaf phenology paradox: Why warming matters most where it is already warm. Remote Sens. Environ. 209, 446–455 (2018).Article 
ADS 

Google Scholar 
Breshears, D. D. et al. Underappreciated plant vulnerabilities to heat waves. New Phytol. 231, 32–39 (2021).Article 

Google Scholar 
Chaudhry, S. & Sidhu, G. P. S. Climate change regulated abiotic stress mechanisms in plants: A comprehensive review. Plant Cell Rep. 41, 1–31 (2022).Article 
CAS 

Google Scholar 
Sun, X. et al. Global diurnal temperature range (DTR) changes since 1901. Clim. Dyn. 52, 3343–3356 (2019).Article 

Google Scholar 
Ballinger, T. J. NOAA Arctic Report Card 2021: Surface Air Temperature. https://doi.org/10.25923/53XD-9K68 (2021).Jagadish, S. V. K., Way, D. A. & Sharkey, T. D. Plant heat stress: Concepts directing future research. Plant Cell Environ. 44, 1992–2005 (2021).Article 
CAS 

Google Scholar 
Gilmore, E. C. Jr. & Rogers, J. S. Heat units as a method of measuring maturity in corn. Agron. J. 50, 611–615 (1958).Article 

Google Scholar 
Sánchez, B., Rasmussen, A. & Porter, J. R. Temperatures and the growth and development of maize and rice: A review. Glob. Change Biol. 20, 408–417 (2014).Article 
ADS 

Google Scholar 
Molitor, D., Junk, J., Evers, D., Hoffmann, L. & Beyer, M. A high-resolution cumulative degree day-based model to simulate phenological development of grapevine. Am. J. Enol. Vitic. 65, 72–80 (2014).Article 

Google Scholar 
CaraDonna, P. J., Iler, A. M. & Inouye, D. W. Shifts in flowering phenology reshape a subalpine plant community. Proc. Natl. Acad. Sci. 111, 4916–4921 (2014).Article 
ADS 
CAS 

Google Scholar 
Inouye, B. D., Ehrlén, J. & Underwood, N. Phenology as a process rather than an event: From individual reaction norms to community metrics. Ecol. Monogr. 89, e01352 (2019).Article 

Google Scholar 
Miles, W. T. S. et al. Quantifying full phenological event distributions reveals simultaneous advances, temporal stability and delays in spring and autumn migration timing in long-distance migratory birds. Glob. Change Biol. 23, 1400–1414 (2017).Article 
ADS 

Google Scholar 
Moussus, J.-P., Julliard, R. & Jiguet, F. Featuring 10 phenological estimators using simulated data. Methods Ecol. Evol. 1, 140–150 (2010).Article 

Google Scholar 
Dowle, M. & Srinivasan, A. data.table: Extension of ‘data.frame’ (2019).Auguie, B. egg: Extensions for ‘ggplot2’: Custom Geom, Custom Themes, Plot Alignment, Labelled Panels, Symmetric Scales, and Fixed Panel Size (2019).Wood, S. & Scheipl, F. gamm4: Generalized Additive Mixed Models using ‘mgcv’ and ‘lme4’ (2020).Auguie, B. gridExtra: Miscellaneous Functions for ‘Grid’ Graphics (2017).Hamner, B. & Frasco, M. Metrics: Evaluation Metrics for Machine Learning (2018).Gilli, M., Maringer, D. & Schumann, E. Numerical Methods and Optimization in Finance (Elsevier/Academic Press, 2019).MATH 

Google Scholar 
Garnier, S. viridis: Default Color Maps from ‘matplotlib’ (2018).Wickham, H. et al. Welcome to the Tidyverse. J. Open Source Softw. 4, 1686 (2019).Article 
ADS 

Google Scholar  More

Systematic paleontologySauropterygia Owen, 186038.Eosauropterygia Rieppel, 199439.Pachypleurosauroidea Huene, 195640.Pachypleurosauridae Nopcsa, 192841.Luopingosaurus imparilis gen. et sp. nov.EtymologyThe genus name refers to the Luoping County, at which the fossil site is located. Species epithet imparilis (Latin) means peculiar and unusual.HolotypeA ventrally exposed skeleton with a posterior part of the caudal missing, IVPP V19049.Locality and horizonLuoping, Yunnan, China; Second (Upper) Member of Guanling Formation, Pelsonian (~ 244 Ma), Anisian, Middle Triassic37.DiagnosisA pachypleurosaurid distinguishable from other members of this family by the following combination of features (those unique among pachypleurosaurids identified with an asterisk): snout (preorbital portion) long and anteriorly pointed, 55.0% of skull length (*); orbital length about one quarter of skull length; internal naris retracted, without contribution from premaxilla; nasal ending at level of anterior margin of prefrontal; dentary length 71.7% of mandibular length; hyoid length 9.7% of mandibular length; presence of entepicondylar foramen in humerus; 21 cervical and 27 dorsal vertebrae (*); distinct expansions of distal heads of posterior two sacral ribs; six pairs of caudal ribs; phalangeal formula 2–3-5–5-3 for manus and 2–3-4–6-4 for pes (*); Metatarsal I short and stout with expanded proximal end, 56.4% of Metatarsal V in length (*); and Metatarsal IV being longest phalange in pes.Comparative descriptionThe holotype and only currently known specimen of Luopingosaurus has a preserved length of 46.2 cm from the rostral tip to the 30th caudal vertebra (for measurements, see Table 1). The estimated total length of the body may have reached 64 cm, assuming similar tail proportions of pachypleurosaurids. As such, Luopingosaurus is longer than most of other pachypleurosauroids that are small-sized with a maximum total length rarely exceeding 50 cm4,9,10,11,12,14,15,16,18,23,25, although some pachypleurosaurids are notably larger (e.g., 88 cm in Diandongosaurus cf. acutidentatus22, ~ 120 cm in Neusticosaurus edwardsii8, and ~ 130 cm in Wumengosaurus delicatomandibularis13).Table 1 Measurements (in mm) of the holotype (IVPP V19049) of Luopingosaurus imparilis gen. et sp. nov. R, right.Full size tableThe pre-orbital portion, distinctly longer than the postorbital region, measures 55% of the total skull length (the premaxillary symphysis to the occipital condyle) and 51% of the mandibular length. The paired premaxillae form most of the snout anterior to the naris with a pointed anterior tip, similar to the conditions in Wumengosaurus13,30 and Honghesaurus23. By contrast, other pachypleurosauroids uniformly have a blunt rostrum4,6,7,8,9,10,11,12,14,15,16,18,22,25. The premaxilla bears a long posteromedial process inserting between the anterior parts of the elongate nasals (Fig. 3). The premaxillary teeth are homodont with a tall peduncle and a short, conical crown, but the tooth number is hard to estimate because of occlusion of jaws. The posterior parts of nasals contact each other medially, and posteriorly, they contact the frontals in an interdigitating suture at the level of the anterior margin of the prefrontal. In Honghesaurus23, Wumengosaurus30, Neusticosaurus8 and Serpianosaurus9, the even longer nasal extends posteriorly beyond this level and ends at the anterior portion of the orbit.Figure 3Skull and mandible of Luopingosaurus imparilis gen. et sp. nov., IVPP V19049. Head before (a) and after (b) dusted with ammonium chloride. (c), Line- drawing. (d, e), two computed laminography scanning slices. (f), reconstruction in ventral view. ac, acetabulum; an, angular; ar, articular; ax, axis; c, cervical vertebra; den, dentary; eo, exoccipital; f, frontal; hy, hyoid; in, internal naris; j, jugal; m, maxilla; n, nasal; p, parietal; par, prearticular; pof, postfrontal; prf, prefrontal; pt, pterygoid; q, quadrate; qj, quadratojugal; sa, surangular; sp, splenial; sq, squamosal; stf, supratemporal fossa; v, vomer.Full size imageThe orbit is oval and large, measuring 24.8% of the skull length (Fig. 3). The lateral margin of the frontal contacts the prefrontal anteriorly and the postfrontal posteriorly, and defines most of the medial border of the orbit. The L-shaped jugal, together with the posterolateral process of the maxilla, forms the lateral border of the orbit. No distinct lacrimal is discernable; the bone is probably absent as in other sauropterygians. The postfrontal contacts the dorsal process of the triradiate postorbital ventrally, and both bones define the posterior border of the orbit. Additionally, the posterior process of the postorbital contacts the anterior process of the squamosal, forming the bar between the supratemporal fossa and the ventrally open infratemporal fenestra. The jugal extends beyond the ventral margin of the postorbital and also contacts the anterior process of the squamosal, resembling the conditions in Wumengosaurus30, Honghesaurus23 and Diandongosaurus15. This contact is absent in other pachypleurosauroids4,6,7,8,9,10,11,12.A pair of vomers and pterygoids and a right palatine are discernable in the palate (Fig. 3a–c). The vomer is elongate and slender, extending anteriorly well beyond the nasal. The internal naris, partly covered by the detached splenial, is longitudinally retracted, corresponding to a retracted external naris (Fig. 3d–f). The medial margin of the naris is defined by the nasal, without contribution from the premaxilla. A retracted naris is otherwise present in Wumengosaurus13,30, Qianxisaurus16 and Honghesaurus23. Similar to the condition in Honghesaurus23, the retracted naris of Luopingosaurus is relatively short, having a longitudinal diameter distinctly less than half of the longitudinal diameter of the orbit. By contrast, other pachypleurosauroids4,6,7,8,9,10,11,12,25 generally have an oval-shaped naris. The elongate palatine has a slightly convex medial margin suturing with the pterygoid. Because of the coverage of the detached splenial, the lateral portion of the palatine is unexposed, and it is hard to know whether an ectopterygoid is present or not. The pterygoid is the largest and longest element of the palate, measuring 55.2% of the mandibular length. It has an anterior projection that contacts the vomer anteromedially, and does not participate in the margin of the internal naris. At the level of the posterior orbital margin, the pterygoid has a triangular lateral extension, which was termed as the ectopterygoid process of the pterygoid in Neusticosaurus8. The pterygoid extends back to the occipital condyle, and covers the basicranium and parietals in ventral view. Additionally, the bone has a broad posterolateral process that is set off from the palatal surface by a distinct ridge, resembling the conditions in Serpianosaurus9 and Neusticosaurus8. Posteriorly, the basioccipital is exposed in ventral view, showing the area for attachment to the right exoccipital.The left quadrate is exposed in lateral view with its dorsal process extending underneath the base of the descending process of the squamosal. The posterior margin of the quadrate is excavated, as in many other pachypleurosaurids (e.g., Serpianosaurus9 and Honghesaurus23). The quadratojugal is narrow and splint-like, flanking the anterior margin of the quadrate. A pair of hyoids are ossified. They are rod-like, slightly expanded at both ends. The dentary is wedge-shaped, being 71.7% of the mandibular length. Laterally, it bears a longitudinal series of pores and grooves parallel to the oral margin of the bone (Fig. 3a). The elongate angular tapers at both ends, contacting the dentary anterodorsally and the surangular dorsally in ventral view. The surangular, slightly shorter than the angular, contacts the articular posterodorsally, with a pointed anterior tip wedging into the notched posterior margin of the dentary. The retroarticular process of the articular is very short with a rounded posterior margin. Medially, the splenial and prearticular form most of the inner wall of the mandible. The splenial tapers at both ends and enters the mandibular symphysis anteriorly, having a length similar to the dentary. The relatively slender prearticular contacts the splenial anterodorsally, extends posteriorly and abuts the articular dorsally, measuring 41.1% of the mandibular length.The whole series of 21 cervical vertebrae (including the atlas-axis complex) is well exposed ventrally. The atlas centrum is oval, much smaller than the axis centrum (Fig. 3c). From the axis, the cervical vertebrae increase gradually in size toward the trunk vertebrae posteriorly. The bicephalous cervical ribs have typical free anterior and posterior processes as in other pachypleurosauroids8,9. The trunk is relatively long, including 27 dorsal vertebrae. The posterior dorsal ribs show certain pachyostosis (Fig. S1). Each gastralium consists of five elements (a short and more massive median element and two slender rods in line towards each side; Figs. 3, 4a, b, S1), similar to the conditions in most of other pachypleurosauroids9,11,18,25 (except Neusticosaurus8). Three sacral ribs are clearly revealed by X-ray computed microtomography (Fig. 4c–f). They are relatively short and stout, with the posterior twos bearing a distinct expansion on their distal heads. The distal expansion of the sacral rib is also present in Keichousaurus11, Prosantosaurus25, Qianxisaurus16 and Wumengosaurus13, but it is not pronounced in other pachypleurosauroids4,6,7,8,9,10. The caudal ribs are relatively few, six pairs in number. Additionally, several chevron bones are visible in the proximal caudal region, and they are gradually reduced in length posteriorly (Fig. 4d).Figure 4Girdles, limbs and vertebrae of Luopingosaurus imparilis gen. et sp. nov., IVPP V19049. Photo (a) and line-drawing (b) of pectoral girdle, forelimbs and anterior dorsal vertebrae. Photo (c), line-drawing (d) and two computed laminography scanning slices (e, f) of pelvic girdle, hind limbs and posterior vertebrae. as, astragalus; ca, caudal vertebra; cal, calcaneum; car, caudal rib; co, coracoid; d, dorsal vertebra; dltp, deltopectoral crest; enf, entepicondylar foramen; fe, femur; fi, fibula; h, humerus; il, ilium; int, intermedium; is, ischium; mc, metacarpal; mt, metatarsal; pu, pubis; s, sacral vertebra; sc, scapula; sr, sacral rib; ti, tibia; ul, ulna; uln, ulnare.Full size imageThe paired clavicles and the median interclavicle form a transverse bar at the 20th cervical vertebrae (Fig. 4a, b). The blade-like clavicle tapers posterolaterally with its distal projection overlapped by the scapula in ventral view. The left clavicle contacts the right one anterodorsally to the interclavicle. The interclavicle tapers laterally to a point at each end. The anterior margin of the interclavicle is convex and its posterior margin is slightly concave without a midline projection (contra the condition in Anarosaurus42). The scapula consists of a broad ventral portion and a relatively narrow and elongate dorsal wing that varies little through its length. The coracoid is hourglass-shaped with a slightly concave posterolateral margin and a conspicuously concave anteromedial margin. The medial margin is straight, along which the coracoids would articulate each other in the midline. The humerus is constricted at the middle portion with a nearly straight preaxial margin and a concave postaxial margin. A slit in the expanded distal portion of this bone indicates the possible presence of an entepicondylar foramen (Fig. 4a, b). The radius, slightly longer than the ulna, is more expanded proximally than distally. The ulna is straight with a slightly constricted shaft. In each forelimb, there is two nearly rounded carpals, ulnare and intermedium; the former is half the width of the latter. Five metacarpals are rod-like, slightly expanded at both ends. Among them, Metacarpal I is the shortest, 48% of the length of Metacarpal II. Metacarpal III is slightly shorter than Metacarpal IV, which is the longest. Metacarpal V is 71% of the length of Metacarpal IV. The phalangeal formula is 2–3–5–5–3 for the manus, indicating presence of hyperphalangy in Luopingosaurus (see Discussion below).In the pelvic girdle, the ilia, pubes and ischia are well exposed (Fig. 4c–f). The ilium is nearly triangular with a relatively long and tapering posterior process. The plate-like pubis is well constricted at its middle portion, with the medial portion nearly equal to the lateral portion. The obturator foramen is slit-like, located at the posterolateral corner of this bone (Fig. 4e). The ischium is also plate-like, having a relatively narrow lateral portion and an expanded medial portion that is notably longer than the medial portion of the pubis. The posterolateral ischial margin is highly concave. The posterior pubic margin and anterior ischial margin are moderately concave, and both together would enclose the thyroid fenestra. The femur is slightly longer than the humerus, with a constricted shaft and equally expanded ends (Fig. 4d). No internal trochanter is developed. The tibia is nearly equal to the fibula in length; the former is straight and thicker than the slightly curved latter. Two ossified tarsals, calcaneum and astragalus, are nearly rounded; the latter is significantly larger than the former. As in Honghesaurus23, the astragalus lacks a proximal concavity. The right metatarsals are well-preserved. Metatarsal I is the shortest and stoutest phalange with an expanded proximal end, and Metatarsal IV is the longest. Metatarsal II is nearly twice the length of Metatarsal I. Metatarsal III is slightly shorter than Metatarsal IV, and Metatarsal V is 76% of the length of Metatarsal IV. The phalangeal count is 2–3–4–6–4, which is complete judging from the appearance of the distal phalanges in the right pes (Fig. 4c). More

Fully terrestrial eDNA sampling approaches offer a potentially powerful addition to biodiversity monitoring efforts23,24. However, protocols for using eDNA-based methods to characterize terrestrial biodiversity, and vertebrate communities in particular, are still nascent27,28,30. In this study, we show for the first time that an eDNA metabarcoding approach can be used to broadly characterize tree-dwelling mammal communities by sampling tree trunks and surrounding soil. Our findings add to recent work (e.g., for reptiles22,24) showing that surface eDNA collection methods, which are relatively untested compared with soil-based eDNA methods, can also be effective at detecting terrestrial vertebrates. Further, we demonstrate that supplementing metabarcoding detection with qPCR-based methods can greatly improve sensitivity, a potentially important consideration for monitoring schemes focused on rare taxa (e.g., Refs.11,12). Together, our results have significant implications for global biodiversity conservation as the broader guild of arboreal vertebrates includes highly threatened5,48,49, as well as invasive alien species50, that are often cryptic, inhabit inaccessible locations, and are therefore challenging to monitor.Our methods captured over 60% of the mammalian diversity expected at the sites, and a similar fraction of the subset of arboreal species, despite sampling only 21 trees. Species accumulation curves suggest that more species would likely have been added with increased sampling effort. These results broadly agree with those of Leempoel et al.23 who found that soil eDNA metabarcoding well characterized mammal communities in California chaparral. However, in both our study and that of Leempoel et al.23, some conspicuous absences were evident. Bats comprised all of the arboreal species that we expected but failed to detect at our sites using metabarcoding (Fig. 1A). Leempoel et al.23 also noted a lack of bat detections (2 of 14 possible taxa detected), which they suggested could be due to low efficiency of either the 12S primer set or of their soil sampling methods for that order. While both reasons could also apply to the lack of bats detected in our study (discussed further below), the performance of the 12S primer set very likely contributed to our lack of American black bear detections as MiMammal-U primers are known to be ineffective at amplifying bear DNA38. These challenges highlight the reality that false negatives and varying detectability among species are common issues to all survey approaches, including eDNA metabarcoding. Our study represents a rare example among metabarcoding studies in that it uses repeated sampling and community occupancy models to quantify false negative rates. This quantitative approach, coupled with continued experimentation with different molecular techniques and survey methods (e.g., Refs.23,27), will be vital to helping researchers decide how eDNA metabarcoding methods will fit into existing biodiversity monitoring efforts moving forward.Although our results suggest that sampling for tree-roosting bats using eDNA metabarcoding still requires further research and optimization, our approach likely has application to characterizing communities in a much broader range of arboreal species globally. Geographic regions with multiple elusive arboreal mammals of management interest—for example, gliders and tree kangaroos in Australasia, or primates in the global tropics—may be particularly suited for a metabarcoding approach for community-level assessments4,8,9,49. It may be especially useful for rapid biodiversity assessments (RBAs51) in remote forested environments, where the ability to collect multiple samples relatively rapidly without regard to time of day would be a key advantage27. Existing survey methods to monitor arboreal mammals tend to be optimized for particular groups of species, often segregated by body size and behavior, with no suitable single method available to characterize all members of the guild4,8,9,16,49. Diurnal and nocturnal species, for example, often require separate survey methods or timing8. While camera traps capture both diurnal and nocturnal species, they typically miss smaller species16,23. The need for multiple methods to survey for nocturnal and diurnal, or large and small, species separately raises the cost of sampling and can result in datasets that are difficult to compare across sites because of inherent sampling biases8. Excluding bats, we found encouraging results for both diurnal and nocturnal arboreal species of a broad range of body sizes, detecting all seven expected species (Fig. 1A). While more work is needed to assemble robust genetic reference libraries before global arboreal mammal monitoring with eDNA metabarcoding will be broadly feasible, a clear advantage of the technique remains the power to detect a broad swath of species, with widely varying morphologies and behaviors, with a single method23,27,28,51.The promise of eDNA metabarcoding approaches for at least some arboreal guilds is well illustrated by our results for southern flying squirrel, Glaucomys volans. Like other flying squirrels (Tribe: Pteromyini), this species is strictly nocturnal, highly arboreal, and tends to get injured in live traps, making it difficult to directly observe and monitor48,52. Yet G. volans eDNA was readily detectable using metabarcoding in our study, occurring in 19–26% of soil samples and 47–52% of roller samples across both sites. Our similarly encouraging results for detecting other squirrels (Sciuridae) also bode well for management applications. For example, the methods would enable fine-scale mapping of habitat use in places such as the United Kingdom where native red squirrels (Sciurus vulgaris) are outcompeted by eastern gray squirrels, or the Delmarva peninsula (USA) to support the conservation efforts for the Delmarva fox squirrel (Sciurus niger cinereus)53. Further research is needed to determine the extent to which our results for squirrels generalize to other taxa with similar active tree-climbing lifestyles (e.g., gliders4, primates49).Our finding that soil samples revealed fewer species, had lower detection probability, and had lower read counts than roller samples, even for some non-arboreal species like white-tailed deer, likely reflects multiple factors. First, soil and tree bark represent markedly different biological and chemical environments that likely differ in eDNA quantity by species, eDNA persistence rates54,55, and microorganism abundance. The latter may be especially pertinent to our study as we observed a relatively large drop in the number of reads after removal of microorganism reads, especially for soil samples. This suggests that performing additional purification steps prior to sequencing could boost the ability of both methods, and especially soil eDNA, to detect target species by increasing mammalian sequencing depth. Other in-lab factors, such as method of extraction23 or choice of primers, similarly have the potential to influence the recovery and amplification of target species’ DNA and should be the focus of future research.Next, our focal trees were not chosen to occur near any special attractants or areas of multi-species use, such as saltlicks or water sources, which has proven successful in other vertebrate eDNA studies18,25,31,32,56. It is possible that adding a broader range of soil sampling sites, including some targeted towards other guilds (e.g., burrow users32,56), would have yielded a more complete inventory. Nevertheless, both soil and surface methods have advantages over the much more commonly-used metabarcoding approaches that rely on natural water bodies for assessing mammal communities16,17,27,28,29,30,38 as they are not limited to where these features occur. Our study is the first to suggest that surface eDNA metabarcoding methods can be a powerful supplement to established soil-based methods of characterizing mammal communities, especially for arboreal species.As noted, bats were especially lacking from our eDNA metabarcoding results, with only two of six likely species detected. Notably, our metabarcoding species list lacked two of the bat species that our sampling scheme was designed around (eastern red bat and northern long-eared bat) and for which we had confirmed recent presence at the sites (Table 1). The lack of northern long-eared bat detections may directly relate to recent precipitous population declines (~ 99%) caused by white-nose syndrome57. However, the lack of eastern red bat detections was especially surprising as roosting of this species was suspected based on telemetry in 17 of our 21 target trees. Reasons for this omission may relate to the fact that eastern red bats roost singly on small twigs and in leaf clusters, and therefore may not leave much DNA on tree trunks. Another possibility is low efficiency of the 12S primer set for bats, although we were unable to find information about this in the literature. It is notable that Leempoel et al.23 had a similarly poor representation of bats with comparable soil-based methods. However, our metabarcoding results did indicate that we are capable of detecting even uncommon, or at least unexpected, bat species with our methods. Eastern small-footed bat, which is typically viewed as a rock-roosting species and is considered endangered by the International Union for Conservation of Nature (IUCN)58, was detected in both soil and surface eDNA samples from Morristown National Historic Park. This species was not otherwise confirmed as present at the site until a year later, in spring 2022, when it was caught in a mist net (BM, unpublished data). Our results with respect to bat detections, along with those of others23, underscore the need for further research to adapt eDNA metabarcoding methods to this vulnerable group, which could contribute much needed demographic and distribution information. This is especially urgent as 18% of bat species are listed as “data deficient” by the IUCN, while 57% lack basic population trend information5,58.Our comparison of qPCR to metabarcoding detection methods for big brown bat represents a hopeful result for the use of eDNA to monitor rare vertebrates that are of particular conservation interest. It is well-known that qPCR-based eDNA surveys targeted towards individual species return higher detection probabilities and have greater power at low abundance, than metabarcoding approaches59. Our results agree, showing for the first time that adding a qPCR step in the analysis of surface and soil eDNA samples can be effective for detecting bats in forested environments. The addition of a qPCR step opens the door for developing species-specific assays to increase detection power for endangered or elusive bat species, or other cryptic arboreal mammals49. Emerging molecular detection approaches such as droplet digital PCR have the potential to increase this sensitivity even further59. Like other eDNA-based tools and survey tools in general, careful consideration of sampling effort, the natural history of target species, and the configuration of different field and molecular methods will be key to optimizing our approach to characterize mammal communities, or to target a particular species, in different regions.Although eDNA surveys are not inexpensive given the need for both fieldwork and molecular analyses, they can be cheaper than conventional approaches, especially if such approaches require many hours of fieldwork or expensive equipment60. Thus, the relative cost-effectiveness of surface or soil eDNA surveys will depend heavily on the mammal communities of interest, the mix of methods that must be employed to effectively sample them, and the purpose of the sampling efforts. However, even if costs are increased, eDNA surveys can reduce field time to the extent that they can improve detection rates, either by replacing or supplementing conventional sampling methods (e.g., as a supplement to visual observations). With higher detection rates, fewer visits are required to achieve the same results. This operational efficiency would be especially advantageous when field conditions present safety risks, are intrusive to sensitive habitats, or are challenging to access. For example, adding surface eDNA sampling to existing visual surveys of eastern wood rats (Neotoma floridana), a cryptic mammal that inhabits steep, rocky slopes in the eastern US, could likely increase detection power, thereby reducing the need for additional risky and costly sampling visits. More studies involving direct comparisons among methods (e.g., Refs.23,24,30,60), in a variety of ecoregions, are needed to determine the extent to which incorporating our methods into existing vertebrate monitoring workflows would increase efficiency.Finally, we detected other vertebrates, including seven birds and one salamander, in soil and surface eDNA samples, despite our use of a mammal-specific primer set. This is similar to results from Leempoel et al.23 in California using the same primer set, in which six bird species were detected. We found that surface eDNA detected more bird species than soil, perhaps for the same reasons as for mammals (above). Our results provide evidence that surface eDNA surveys, with taxon-specific primers, could be used to survey bird communities, or used to target particularly rare species in forested ecosystems (e.g., Ref.61). Our detection of a salamander, coupled with recent promising research into reptile detection using surface eDNA methods22,24 suggests a broader potential for applications with other vertebrates as well. Finally, both surface and soil eDNA metabarcoding can be expanded beyond forests, providing insight into their effectiveness in other habitats (e.g., caves17 or talus slopes). Our study and others highlight that the potential of coupling surface and soil eDNA methods for detecting and monitoring mammalian biodiversity, and terrestrial organisms generally, has yet to be fully realized. More