The study was conducted in southwest Madagascar within the Velondriake LMMA (43°13`30 E, 22°04`22 S), 150km north of the regional capital of Toliara ( Figure 1 ). The Velondriake LMMA stretches along 40km of coastline, incorporating 35 coastal villages with a combined population of approximately 8000 people (see Gardner et al., 2020 for further information on Velondriake). For those living in the LMMA, seafood provides 99% of protein, and 87% of the adult population derive their livelihoods from small-scale fishing (Barnes-Mauthe et al., 2013).

The marine area of the Velondriake LMMA covers approximately 600 km². The reef complex in the LMMA is characterised by a fringing and barrier reef system with several patch reefs situated in a 5km wide channel. The region has relatively high annual variation in sea surface temperature (McClanahan et al., 2009) with a range of up to 11°C recorded (22-33°C in 2015; Blue Ventures, unpublished data) and the presence of large, anticyclonic eddies, caused by the interaction of the south equatorial current with the Madagascar landmass (Quartly and Srokosz, 2004). The distinctive oceanographic climate has resulted in a biogeographically unique reef assemblage that includes corals endemic to Madagascar such as Pocillopora fungiformis and Stylophora madagascarensis (Veron, 2002), and 430 species of reef fish (Nadon et al., 2008).

PITs have been identified as a time efficient survey method of monitoring disturbance over large reefs areas with non-specialist surveyors (Facon et al., 2016). To ensure data accuracy, before undertaking PITs, the Blue Ventures citizen scientists underwent training in survey and identification of corals and other benthic life forms. To be eligible to undertake surveys, participants were required to pass a computer test and an in-water identification tests with an accuracy rate of 90%. Following completion of these tests, participants undertook a PIT with a member of trained staff. On the PIT test, the participant’s data had to be in 100% agreement with the data collected by the staff member when identifying major groups (e.g., between coral, algae, substrate), and citizen scientists had to correctly identify organisms within these major groups to at least 80% accuracy (e.g., branching or massive coral growth forms, calcareous or turf algae). Quadrat surveys of the adult coral community assemblage and the recruit community assemblage (where identifying to genus level was required) were undertaken only by trained field staff.

We investigated the severity of cyclone damage in relation to three continuous cyclone parameters: (i) wind speed (average maximum wind speed when the cyclone was within 70km of the reef site in km/h), (ii) duration (the number of hours the cyclone was within 70km of the reef site) and (iii) distance (the minimum distance from centre of the cyclone to the reef site in km). Wind speed and track data for Cyclone Haruna and historic cyclones in the region were extracted from the International Best Track Archive for Climate Stewardship (Knapp et al., 2010). Seventy kilometres was selected as the distance threshold as it is within the range observed for severe damage following cyclone events of similar intensity to Cyclone Haruna (Manzello et al., 2007; Puotinen, 2007).

Generalized linear mixed model (GLMM) analysis was used to assess variation of cyclone impact on (i) hard coral cover (ii) taxonomic richness (number of coral genera per site) and (iii) and recruit density (number of colonies <10cm diameter per m²). We divided the survey data into four time periods related to the date of the cyclone. We refer to the two years before the cyclone as year -2 (22nd February 2011 - 22nd February 2012) and year -1 (23rd February 2012 - 22nd February 2013) and the two years after the cyclone as year 1 (23rd February 2013 - 22nd February 2014) and year 2 (23rd February 2013 - 22nd February 2014). GLMM analysis of hard coral cover and recruit density included survey data two years before and two years after the cyclone. For taxonomic richness, data was available for only one year before and two years after the cyclone. GLMM analysis was selected to allow for the inclusion of fixed and random effects. Fixed effects included cyclone parameters wind speed, duration and distance and environmental variables and reef type and reef depth ( Table S1 ). We also included a fixed effect (before/after) for whether a survey was conducted before or after the cyclone occurred to quantify the overall impact of the cyclone on coral cover, taxonomic richness and coral recruits. The before/after term was allowed to interact with cyclone parameters to try and understand how exactly cyclones impact on coral communities (Laird and Ware, 1982). We also included reef type and reef depth in interaction with the before/after term to determine if different reef environments were impacted differently by the cyclone. We examined Pearson correlation coefficients and Variance Inflation Factors (VIFs) to assess potential multi-collinearity between fixed effect variables. Wind and duration metrics were found to be strongly negatively correlated (r = -0.83 to - 0.85). We therefore did not include both wind and duration in the same model, and calculated VIF for the two potential subsets of predictors. The highest VIF for any fixed effect in all subsets of models was 1.15. In each model, site was included as a random effect. Observer identity was included as a random effect in models investigating taxonomic richness and recruit density.

We assessed whether model assumptions were followed using graphical procedures, evaluating homogeneity of variance by plotting residuals against each fixed effect and against fitted values. All numerical explanatory variables were scaled by standardisation to the same mean and variance in order to compare their relative effects on response variables. Taxonomic richness was modelled as a poisson distribution with a log link function. Hard coral cover and recruit density were modelled as a negative binomial distribution with a log link function (poisson models for these variables were overdispersed). All models were fit using Laplace Approximation. For each of the GLMM analyses, null models (with no fixed effects), a model including only before/after as a fixed effect and all possible combinations of fixed effects in interaction with the before/after term were tested. Model performance was compared using Akaike’s Information Criterion, corrected for small sample sizes (AICc; Akaike, 1974). The model with the smallest AICc value was considered the ‘best’ model and competing models were considered equivalent if ΔAICc < 2. In addition to AIC, we used Likelihood Ratio Tests to compare the goodness of fit and test the significance of individual fixed effects when dropped from the best model.

The analysis of similarities (ANOSIM) test is a non-parametric test that can be used to assess for significant differences between the composition of different groups. ANOSIM was used to test if there is a significant difference between coral community composition before and after Cyclone Haruna. A one-way, pairwise ANOSIM based on Bray-Curtis distance on a square root transformed compositional matrix (species by survey) was conducted to determine the significance of differences in the (i) taxonomic composition of adult corals (ii) morphological composition of adult corals and the (iii) taxonomic composition of coral recruits, one year before and one year after Cyclone Haruna. Sites included in the analysis are listed in Table S2 . The identification of which taxa and morphologies were most important at driving the differences pre and post cyclone were conducted using similarity percentage (SIMPER) analysis. Wilcoxon signed rank tests were used to assess if there was a significant difference in mean coral recruit size before and after the cyclone. Mean relative cover of coral genera and coral morphologies and the relative frequency of recruits size classes were calculated for each year as the number of colonies of each coral genera, morphology or recruit size class per site divided by the total number of observations at that site.

Analysis was conducted using the ‘lme4’ package (Bates et al., 2007) and the ‘vegan’ package (Oksanen et al., 2007) in R Version 1.0.143 (R Core Team, 2020).

Mean coral cover in the study region was 37.8% ( ± 1.2 SE) in year –2 and 36.9% ( ± 1.1 SE) in year -1. During the first year after the cyclone (year 1), mean coral cover declined to 21.1% ( ± 0.9). Coral cover declined across all surveyed sites, but the extent of coral decline varied among locations from -1.4% to -45.8%. Two years after the cyclone (year 2) mean coral cover increased slightly to 23.0% ( ± 1.1 SE; Figure 2 ). However, recovery between year 1 and 2 was not observed at all sites and changes in coral cover ranged from – 11% to + 21.9%.

The most parsimonious models based on AICc values ( Table 1 ) and Likelihood Ratio Tests ( Table S3 ) contained the cyclone parameters duration and distance and the environmental characteristic reef depth, all in interaction with the before/after term ( Table S4 ). The model results indicate that cyclone duration had a significant negative effect, with sites in the path of the cyclone for longer showing a greater decrease in coral cover. Distance and reef depth were also found to have negative effects, with sites located furthest from the cyclone track and sites at a greater depth having greater losses in coral cover after the cyclone. The difference between the marginal R² (0.21) and the conditional R² (0.56) of the best model indicates unexplained site-to-site variation in hard coral cover. Including wind speed and reef type in the models did not improve the model fit. Our best model slightly under-predicted hard coral cover where coral cover was highest, but overall we found a largely linear relationship between observed hard coral cover and that predicted by the model ( Figures S1, S2 ). When we removed those surveys with the highest coral cover from the model, the same results are obtained with respect to which fixed effects are included in the best model and the directionality of their influence on hard coral cover. An additional analysis assessing proportional change in coral cover rather than raw coral cover values gave highly similar results with respect to duration, depth and distance [see Supplementary Material ( Tables S5, S6 )].

A total of 53 coral genera were identified across all the surveys. Mean taxonomic richness was 18 (SE ±2.12) one year before (year -1) and 16 (SE ± 1.50) one year after the cyclone (year 1; Figure 2 ). The most parsimonious models (ΔAIC < 2) were the null model and one with the before/after term ( Table 1 ). The before/after term was not significant, suggesting the cyclone did not consistently change taxonomic richness up or down. Including cyclone parameters (duration, distance and wind speed) and reef type and depth did not significantly improve the model.

The taxonomic and morphological composition showed a significant change when comparing one year before (year-1) and one year after the cyclone (year 1; Genera: ANOSIM, R = 0.07, P < 0.05; Morphology: ANOSIM; R = 0.12, P <0.05). This dissimilarity was driven by changes in the abundance of the dominant coral taxa Acropora, Pocillopora, Galexea and Stylophora (SIMPER: 18.0%, 9.3%, 5.5% and 4.9%, respectively) and branching, encrusting and massive morphologies (SIMPER: 34.4%, 20.4%, 8.1%, respectively). From surveys conducted before the cyclone, the relative cover of Acropora ranged from 10% to 55% between sites. The mean relative cover of Acropora across the region decreased slightly in the year following the cyclone. Other branching genera, Pocillopora and Stylophora also decreased in relative cover following the cyclone. Massive Porites had the greatest increase in relative cover ( Figure 2 ). In the year following the cyclone, the relative cover of branching corals decreased from 34.2% (± 3.05) to 24.34% (± 3.65) and encrusting and massive corals became the dominant morphologies [relative cover of 32.21% ( ± 3.00) and 28.34% (± 2.78) respectively; Figure 2 ].

The greatest decrease in density of coral recruits was observed between the two years before the cyclone from 31.8 (± 2.08) recruits/m² in year -2 to 24.5 (± 1.93) recruits/m² in year -1 ( Figure 3 ). In year 1 after Cyclone Haruna, the density of recruits continued to decrease to a mean of 20.5 (± 1.21) recruits/m². In the second-year post-cyclone, recruit density increased to above pre-cyclone levels [28.9 (± 2.22) recruits per m^-2]. Using a negative binomial GLMM, several models were statistically equivalent (ΔAIC<2), and had limited predictive power ( Table 1 ). Likelihood ratio tests revealed adding cyclone parameters to the model did not improve the model accuracy (p > 0.05) and therefore the null model was assumed the best model.

The composition of coral recruits differed significantly one year before and after the cyclone (ANOSIM, R = 0.11, P < 0.05). Acropora, Pocillopora and Favia contributed most towards the dissimilarity (SIMPER; 9.2%, 8.3% and 7.0% respectively). Acropora was the dominant genus of recruits and after the cyclone, increased in mean relative abundance across the study region from 15.1% (± 1.5) to 17.0% ( ± 2.23). Pocillopora and Favia became less abundant and there was an increase in the relative count of massive Porites, Psammacora and Seriatopora ( Figure 3 ). No significant change was observed in the mean size class of recruits following the cyclone (Wilcoxon signed-rank; V=54, p > 0.05: Figure 3 ).

Extreme winds and waves caused by cyclones dislodge and damage coral colonies causing a decline in coral cover (McLeod et al., 2019). Cyclone Haruna caused coral cover to decline in the Velondriake LMMA by a mean 15.8% in the study region, but the spatial distribution damage was substantially patchy. A mosaic of impact severity is typical of cyclone damage (Harmelin-Vivien, 1994; Guillemot et al., 2010; Perry et al., 2014; Cooke and Marx, 2015), as coral vulnerability is influenced by factors from local (Madin et al., 2012; Madin et al., 2014) to regional scales (Young and Hardy, 1993). No reports of cyclone impact of similar strength events are available for comparison in the WIO. Cyclone damage varies substantially between cyclone events and reef regions. However, the decline observed at the present study site is comparable to the severity of damage caused by category 3 Cyclone Rona (1999) on the Great Barrier Reef (16% decline in average HCC; Cheal et al., 2002) and the average reduction of coral cover caused by hurricanes in the Caribbean (17% decline on average HCC; Gardner et al., 2005).

Previous studies have found sustained wind speed, storm duration and distance from the storm track as effective predictors of coral cover loss (Puotinen, 2007; Fabricius et al., 2008). For Cyclone Haruna, the duration of cyclone impact, the distance from the cyclone path and depth of the reef were the most effective predictors of coral cover decline. Across the study region, reef sites were within 70km of Cyclone Haruna for between 18 to 24 hours. Reef sites within the threshold distance for a longer duration, experienced more severe coral loss. The duration of cyclone impact at a reef site will impact the severity of coral loss as short exposure to high energy waves will only break or dislodge fragile colonies, while persistent exposure to high energy waves will have time to weaken and remove stronger and larger areas of reef framework (Puotinen et al., 2016). Reef depth and distance from the cyclone track modulated the effect of the cyclone on coral cover, with coral loss greater at deeper reef sites and reef sites located further away from the cyclone track. The maximum surveyed reef depth was 13 m, well within the range of which direct mechanical damage to a reef from a tropical cyclone has been observed (refs in Harmelin-Vivien, 1994). For this study, it is possible that more severe coral decline occurred on deeper reefs as they were generally located offshore, exposed to cyclone waves arriving from the open ocean. On the other hand, shallower reef sites located behind the offshore barrier may have been more sheltered as wave energy was dissipated by the barrier reef (Done, 1992; Fabricius et al., 2008). The destruction potential of waves that were generated from the cyclone as it crossed the inshore area are likely to have been limited by shallower depth and shorter fetch (Young and Hardy, 1993). Cyclone damage related to distance from cyclone path is often unpredictable as evidenced by corals located in the direct path of the cyclone remaining unchanged (Massel and Done, 1993), yet coral up to 800km away experiencing severe decline (Puotinen et al., 2020). All reef sites in the study region were located close enough to the cyclone track (between 8.7 km and 39.3 km) in which severe damage would be expected (Fabricius et al., 2008). It is likely that the pattern of coral decline was strongly influenced by localised topography producing refraction of storm waves (Harmelin-Vivien, 1994). The rate of damage would be higher at sites struck directly by cyclone-induced waves as opposed to those sheltered by an adjacent reef. As Cyclone Haruna crossed directly from offshore to inshore in the study region, the reefs were impacted by cyclone-induced waves from multiple directions, which may have produced complex patterns of damage that could not be captured in our model.

Cyclone Haruna did not significantly impact the taxonomic richness of the adult coral community in the study region. The relative difference of the cyclone impact on coral cover compared to taxonomic richness can be partially attributed to the fact that dominant branching morphologies reduced in cover but did not disappear (Rogers, 1993; Fabricius et al., 2008). Following the cyclone, the dominant coral morphology changed from branching to encrusting and massive. A loss of the more fragile coral morphologies (branching and tabular) can be expected from cyclone-induced damage, as they are more vulnerable to fragmentation or dislodgement from extreme waves (Baldock et al., 2014). At some sites, we detected an increase in the relative cover of Acropora in the year following the cyclone. This increase may be caused by fragmentation of existing Acropora colonies concurrent with the mortality of other genera. Overtime, fragments can attach and fuse with underlying coral colonies or substrate and survive (Linares et al., 2011), but due to tissue damage or partial mortality, are also more at risk of mortality from secondary impacts such as bioerosion or fouling from macroalgae (Knowlton et al., 1981). In other cyclone events, mortality from secondary impacts has caused severe mortality rates of Acropora to be recorded from 5 months to over a year after a cyclone event (Knowlton et al., 1981; Muko et al., 2013), however, this does not seem to be the case within the Velondriake LMMA following Cyclone Haruna.

Following declines in coral cover after a disturbance, the survivorship of coral recruits will play a key role in the structuring of the future adult coral assemblage (Coles and Brown, 2007). We observed the greatest decrease in recruit density in the two years before Cyclone Haruna. Interannual variability in recruit density relative to adult coral cover may occur as small individuals can be dislodged or killed outright by biotic (Doropoulos et al., 2014) or abiotic factors (Edmunds et al., 2010) that cause larger individuals only partial or no mortality (Babcock, 1991). Recruit density continued to decrease at some sites in the year following Cyclone Haruna, although we found no significant impact of the cyclone across the region. From our data, it is not possible to extrapolate the influence of Cyclone Haruna compared to other biotic and abiotic pressures on recruit density. Compared to adult coral cover, the relationship between cyclone parameters (wind speed, distance from cyclone track and duration of cyclone impact) and coral recruits may be less pronounced as small colonies will have significantly less drag than their adult counterparts (Harmelin-Vivien, 1994). It is also likely that parameters not captured in our model, such as local scale topography and reef rugosity provided localised shelter for smaller colonies, creating a mosaic of impact severity at an intra-reef scale.

Following the cyclone, observations indicate an increase in macroalgae cover (Blue Ventures, unpublished data). In other regions, algal blooms caused by severe storms have been linked to the failure of coral recruitment and survivorship, as recruits can be outcompeted by the algae for suitable settlement substrate (Baldock et al., 2014; Doropoulos et al., 2014). However, following Cyclone Haruna we found no significant difference in the relative abundance of recruit size classes (1-10cm) and two years after the cyclone recruit density increased to above pre-cyclone levels. This suggests macroalgae did not inhibit the survivorship of recruits which had settled before the cyclone or the settlement of new arrivals.

The taxonomic composition of coral recruits was significantly different following Cyclone Haruna. There was an increase in the relative abundance of sturdier genera including Porites massive, Psammacora, Montipora. This increase can likely be attributed to sturdier morphologies being less fragile to wave action during the cyclone (Baldock et al., 2014) and self- recruitment from surviving adult colonies (Graham et al., 2011). These genera may also have a competitive advantage to secondary cyclone impacts as they are relatively more tolerant to stress (Darling et al., 2013). The genus Acropora is the most abundant genus in the study region, so its recovery is crucial for the maintenance of the coral ecosystem. The ability of Acropora to rapidly colonise space following a disturbance through fragmentation or recruitment may explain the increase in the relative abundance of Acropora recruits after Cyclone Haruna (Pratchett et al., 2020). This is consistent with previous studies where Acropora has rapidly recovered in the aftermath of disturbance events, attributed to fast growth and high levels of recruitment (Sheppard et al., 2002; Linares et al., 2011).

The persistence of the coral community assemblage is dependent on the return interval and intensity of disturbance events (Pratchett et al., 2020). The reef community in southwest Madagascar has evolved with relatively frequent cyclones (Puotinen et al., 2020). Since 1913, there have been 15 cyclones recorded in the region, 13 of these have been category 1 on the Saffir-Simpson scale (Knapp et al., 2010). The remaining two cyclones were Cyclone Earnest (category 3) in 2005 and Cyclone Haruna. No other reports regarding the impact of cyclones on coral reefs in southwest Madagascar are available. However, we can assume the severity of damage caused by category 1 cyclones would be, in general, less than observed following Cyclone Haruna (Fabricius et al., 2008). It is predicted anthropogenic climate change will cause the global average intensity of tropical cyclones to increase (Knutson et al., 2020) and therefore the return interval between damaging cyclones to shorten (Puotinen et al., 2020). Predictions of changes to the size, frequency and spatial distribution of cyclones are more uncertain (Knutson et al., 2020) however, evidence suggests the poleward migration of the latitude at which cyclones reach maximum intensity is already occurring (Kossin et al., 2014). If this trend continues, tropical cyclones may become more prevalent in southwest Madagascar.

The high rates of replenishment of Acropora may be sufficient for the reef ecosystem to recover between cyclone impacts (Thompson and Dolman, 2010). However, mass coral bleaching and mortality caused by marine heatwaves (MHWs) will likely be the leading control on the composition of reef communities (Loya et al., 2001) and lead to an increase in the relative abundance of thermally tolerant species (Woesik et al., 2011). Since the late 20^th century, MHWs have increased in frequency, intensity and duration in the Mozambique Channel and southwest Madagascar has experienced the longest and most extreme MHWs in the region (Mawren et al., 2021). Under a high emission scenario (RCP 8.5), MHWs are predicted to become more intense and prolonged in the WIO (Jacobs et al., 2021). When the impact of tropical cyclones and MHWs are combined in a disturbance system of threats that also includes Acanthaster spp invasions (Bigot and Quod, 2000), coral diseases (Sheridan et al., 2014) and anthropogenic pressures, such as overfishing (Brenier et al., 2011; Gilchrist et al., 2020; Gough et al., 2020), it is likely the reef composition will change to be dominated by stress tolerant coral taxa (Loya et al., 2001; Woesik et al., 2011).

A shift in the community composition to an assemblage dominated by massive and robust stress-tolerant genera has already been observed areas of southwest Madagascar (Harris et al., 2010) and in other reef regions (Hughes, 1994; Fine et al., 2019). Dominance of less complex morphologies will alter the habitat availability and likely lead to a decline in the biodiversity and abundance of reef associated organisms (Pratchett et al., 2008). Small-bodied and juvenile reef fish are particularly vulnerable to a decreased reef complexity (Richardson et al., 2017; Fontoura et al., 2020) they rely on the coral for refuge and nutrition (Gratwicke and Speight, 2005).

In Vezo small-scale fisheries, reef fish represent 30% to 50% of target fish species (Blue Ventures, unpublished data). Therefore, a change to reef fish populations may have implications for the food security and livelihoods of Vezo communities. While it is not possible to remove the threat of cyclone damage to corals, marine management measures can help reduce the likelihood of catastrophic regime shifts. For example, protecting a diversity of coral habitats from destructive fishing gear will provide opportunity for a diversity in ecosystem responses (Nyström et al., 2008) and provide an insurance policy against losing ecosystem functions (van Nes and Scheffer, 2005). Where herbivorous fish are heavily exploited, removing fishing pressure can support the existence of diverse functional groups of herbivores and reduce the risk of a shift from coral to algal dominance after a disturbance event (Graham et al., 2015). Within the Velondriake LMMA, no-take zones permanently closed to extractive activities have been found to have significantly higher fish biomass than fished sites (786 ± 247 kg/ha vs 247 ± 33 kg/ha; Gilchrist et al., 2020), indicating their potential role in coral recovery.

In the present study, we used long term monitoring data collected by trained field staff (nationals and international) and international citizen science volunteers to assess the impact of Cyclone Haruna. This programme was not designed to monitor disturbance-related responses in coral health, and therefore variability in the spatial distribution and sample replication of surveys inhibited deeper investigation on the impact of Cyclone Haruna on reefs. However, despite challenges and limitations in data analysis, in remote and data-scarce regions, using the data that is available is critical to support local reef health and strengthen regional knowledge of ecological and societal outcomes for coral reefs and dependent communities. Engaging local stakeholders in reef monitoring is critical to foster public support for coral reef conservation and local ownership of management efforts (Aswani et al., 2015). Monitoring programmes must be accompanied by skills training and sustainable financing (Hattam et al., 2020). In locations with economic constraints, citizen science provides an opportunity ​​to support robust large scale and long-term monitoring (Chandler et al., 2017). Monitoring approaches should be designed to address localised threats to coral health and social-ecological dynamics, yet guided by global best practices (e.g. Global Coral Reef Monitoring Network; Hill and Wilkinson, 2004) to allow for the aggregation of data at a global level (Freiwald et al., 2018). As disturbance regimes increase on coral reefs, monitoring coral reef status and trends will provide critical data which can be used to inform local marine management and national and international obligations to support reef health (Obura et al., 2019).