Combining field and remote sensing data to estimate forest canopy damage and recovery following tropical cyclones across tropical regions

As tropical forests cycle the most water and carbon, it is crucial to understand the short- and long-term effects of intensifying cyclones on these ecosystems. Soil nutrient status has been shown to moderate forest cyclone responses using field litterfall measurements, but litterfall is one of the multiple cyclone impact metrtics, which may or may not be correlated with one another or with site nutrients. We used remotely sensed vegetation indices to quantify immediate damage and two-year recovery for 42 cases across nine tropical forests in Hawaii, Puerto Rico, Mexico, Australia, and Taiwan affected by 12 cyclones between 2004 and 2017. We tested whether changes in leaf area index (LAI) and enhanced vegetation index (EVI) correlated with changes in litterfall observations and how changes varied with total soil phosphorus (P) concentrations across regions. We compared cyclone-induced changes and recovery of LAI and EVI to litterfall observations compiled in a pantropical meta-analysis. We found large variation in changes in LAI and EVI across forests, with the greatest reductions in LAI (−77%) and EVI (−77%) in Mexico (Jalisco) and Puerto Rico, respectively. LAI (r = −0.52) and EVI (r = −0.60) changes correlated with those in litterfall across cases. Post-cyclone data showed recovery of LAI by four months, EVI by two months, and litterfall by ten months. We detected larger changes in LAI and EVI in forests with higher soil P, but these relationships were not significant when accounting for cyclone and site as random effects. Principal component analyses indicated a regional clustering of cases related to their contrasting cyclone regimes, with the frequency and intensity of cyclone events negatively correlated. Overall, remote sensing observations complement but do not substitute for ground observations that reveal cyclone damage and post-cyclone recovery in tropical forests, and soil phosphorus moderates some but not all metrics of stability in response to cyclones.


Introduction
Tropical cyclones are natural disturbances, events that are discrete in time and space that shape the structure, composition, function, resource availability, physical environment and dynamics of coastal forest ecosystems worldwide (White and Pickett 1985, Zimmerman et al 1996, Syvitski et al 2005, Lugo 2008).Cyclone-prone forests experience a range of cyclone regimes characterized by cyclone intensity, frequency or return interval, duration, and severity or degree of damage (White and Pickett 1985), which are being altered under climate change (Hsiang 2010, Silver et al 2014, Feng et al 2020, Kassambara and Mundt 2020, Mandal and Hosaka 2020, Reed et al 2020, Hosannah et al 2021).Understanding the impact of altered cyclone regimes on tropical forests is central to quantifying further changes in tropical forest carbon cycling under climate change.
The immediate impacts of cyclones on tropical forests are defoliation and canopy damage (Mitchell 2013), resulting in cyclone-associated peaks and fluctuations in litterfall (Lodge et al 1991, Zimmerman et al 1996, Lugo 2008, Silver et al 2014, Bomfim et al 2022).Thus, litterfall can potentially be used as an indicator of tropical cyclone severity (Chambers et al 2007a, 2007b, Bomfim et al 2022).However, litterfall sampling is labor-intensive and time-consuming and usually occurs within a limited field sampling area (Clark et al 2001).Field sampling density is also sparse in space and time, with uncertainty in anticipating the time and location of cyclone landfalls, further increasing sampling difficulty.These space, time, and disturbance predictability limitations of field litterfall sampling are due to the fundamental limitations of field sampling.Therefore, the development of remotely sensed vegetation metrics that correlate with litterfall would expand the spatial and temporal range of new investigations on how forests respond and recover from intensifying cyclone disturbances under climate change, as satellites are capturing imagery of potential study sites regardless of disturbance, time, or space (Hu and Huang 2019).
Vegetation greenness time series (e.g. the normalized difference vegetation index-NDVI) derived from optical (350-2500 nm) remotely sensed data can observe vegetation dynamics and ecosystem productivity as they capture the state of an ecosystem through time and space (Chambers et al 2007a, Ostertag et al 2008), without limitations of field sampling.Previous studies have used remote sensing vegetation greenness metrics, such as NDVI (Delaporte et al 2022), leaf area index (LAI) (Wang et al 2010), enhanced vegetation index (EVI) (Rogan et al 2011), and non-photosynthetic vegetation (NPV) (Negrón-Juárez et al 2014, Feng et al 2020) to measure the impact of cyclones on forests, since these vegetative indices (VIs) are able to capture changes to forest state and health.For instance, Lee et al (2008) used remote sensing to calculate NDVI as a proxy for canopy damage from typhoon Herb in a tropical forest in Taiwan.Peereman et al (2020) compared five vegetation indices and recommended normalized difference infrared index and EVI for cyclone disturbance observations due to their sensitivity in measuring the changes in vegetation characteristics.The impacts of hurricanes Katrina, Rita, Yasi, and Maria were quantified using the difference in NPV or ∆NPV (Chambers et al 2007a, 2007b, Negron-Juarez et al 2010, Feng et al 2020).Further, studies have shown that EVI can explain spatial variability in annual litter production in the summer growing season (Hu and Huang 2019) and that there is a linear relationship between satellite-derived ∆NPV and field-measured forest mortality (Chambers et al 2007b).However, these studies are site-specific, and tropical forest structure and composition (Phillips et al 2019), function (Aguirre-Gutiérrez et al 2021), and demographic rates (Needham et al 2022) can vary substantially across sites.Therefore, a multi-site assessment of the relationship between remote sensing metrics and field data is necessary before conducting a globally robust remote sensing analysis of cyclone impacts.
Because pantropical forest litterfall fluxes respond variably to tropical cyclones (Bomfim et al 2022), understanding the drivers of this variation will improve regional and pantropical extrapolation of cyclone impacts.Field studies have quantified changes in vegetation and vegetation indices in response to tropical cyclones (e.g.Everham and Brokaw 1996, Walker et al 1996, Lugo 2008, Ibanez et al 2019, Hogan et al 2020, Lin et al 2020) and related these changes to soil resource availability (Herbert et al 1999, Gleason et al 2008).Herbert et al (1999) reported that phosphorus (P)-fertilized forest plots showed increased field-measured LAI and, consequently, higher defoliation severity relative to co-located unfertilized plots following cyclone Iniki in Hawaii.Evidence from Australia suggests that plant species growing on high-P basalt soils had greater changes in LAI and higher branch breakage immediately after a cyclone than those growing on low-P schist soils (Gleason et al 2008).A recent pantropical meta-analysis of field-observed observations indicated that across 26 tropical forest sites affected by 22 tropical cyclones, cyclone-induced litterfall increased by 33%-38% for each 100 mg kg −1 increase in total soil P concentration (Bomfim et al 2022).However, it is not known whether the patterns seen across sites in litterfall can be detected remotely.
To elucidate the role of soil P in mediating tropical forest canopy responses to tropical cyclones, we first explore several remote sensing vegetation indices to (i) identify those with strong correlations with litterfall and (ii) quantify canopy damage and recovery following cyclones.Using pre-and post-cyclone vegetation indices, we ask: do remote-sensing-based estimates of canopy damage relate to soil P? What other cyclone and environmental factors explain the variation in cyclone responses?Does the time to canopy recovery detected with remote sensing track field litterfall recovery to pre-cyclone values?We expect forests on low-P soils to be more resistant to cyclone disturbance than forests on high-P soils, and that this P effect is modulated by cyclone regime factors, mainly intensity.

Cases and site-level variable selection
We selected 42 cases from a meta-analysis of cyclone-induced litterfall mass flux responses within tropical latitudes (23.5 • N-23.5 • S) including sites in Hawaii, Puerto Rico, Mexico, Australia, and Taiwan (Bomfim Note: A Cyclone frequency represents the number of storms listed in IBTrACS (Knapp et al 2010) over the 1955-2020 period for the 1 • latitude × 1 • longitude grid cell (where the site is located) divided by the number of years between 1955 and 2020.While not a precise measure, storm frequencies are autocorrelated in space across scales (Lugo et al 2000).
et al 2022).Each case represented a unique combination of site and cyclone disturbance for litterfall mass flux.Field data collection for each case used a field basket (litterfall trap) collection method, where baskets averaging 43 cm × 43 cm (1849 cm 2 ) attached to 1 m high poles were randomly distributed in forest sites (Silver et al 2014) to estimate litterfall mass flux variations.MODIS (moderate resolution imaging spectroradiometer) LAI 500 m and MODIS EVI 250 m data were available for 21 of these cases, representing nine sites affected by 12 tropical cyclone events between 2004 and 2017 (table 1).Our analysis focuses on these 21 cases.As shown in table 1, some sites were affected more than once by a cyclone event, such as those in Taiwan.Bomfim et al (2022) also compiled soil, lithology, climate, biogeographic, and cyclone regime variables for each case, including site-specific total soil P concentrations.The case data compiled by Bomfim et al (2022) allowed for comparisons between the total litterfall mass flux, remote sensing-derived LAI and EVI, and site-specific soil, environmental, and cyclone regime variables (figure 1).

Remote sensing
We used Google Earth Engine (GEE) to extract remote sensing data from five satellites yielding seven VIs at different resolutions (figure 1, table S1).VIs were either extracted directly from data products containing pre-calculated VI values or calculated within GEE using specific bands needed in VI equations (see supplementary methods).The MODIS LAI at 500 m and EVI at 250 m resolution were extracted from the associated data products.We evaluated LAI (500 m resolution), EVI (90 m and 250 m), NDVI (90 m, 250 m and 5 km), and kNDVI, or kernel NDVI (90 m), and determined that EVI (90 m), kNDVI (90 m), and NDVI (all resolutions) were unsuitable for our study because they lacked suitable imagery in most cases or had poor correlation with litterfall (see supplementary materials).Therefore, we focused on MODIS EVI (250 m) and LAI (500 m), because of their extensive spatiotemporal coverage (figure 1).The full syntax of the GEE code can be found at https://github.com/DelliB/NGEEtropics_SULI_spring.
The limitations of MODIS LAI 500 m and EVI 250 m data products include their coarse spatial resolution and time frame of coverage, with data only available after 2000 for EVI 250 m and 2002 for LAI 500 m.We used LAI because it can indicate cyclone-induced changes, including a large litterfall pulse.EVI was selected because of its ability to improve greenness saturation in high-LAI tropical forests (Zheng and Moskal 2009).The MODIS data products for LAI at 500 m and EVI at 250 m are described in the supplementary materials.
The GEE-based data extraction required both site coordinates and the cyclone occurrence date of the Bomfim et al (2022) cases (table S1).The extraction dates represent either pre-cyclone, post-cyclone, or recovery date ranges for each case, as detailed in table 1.To account for phenology, we considered two calculations of cyclone-induced changes in LAI and EVI with alternate pre-cyclone reference periods: (a) annual or a mean of all available pre-cyclone LAI and EVI values for the entire year prior to the cyclone event, and (b) seasonally adjusted or a month-specific pre-cyclone mean (table 1).Season-specific pre-cyclone VIs were used for both LAI 500 m and EVI 250 m, leading to a total of four seasonally-and non-seasonally based calculations of VI change for each case (data available at http://dx.doi.org/10.15486/ngt/1847332).For the seasonally adjusted values, the pre-cyclone value was calculated by averaging (month-specific) VI values from the month of disturbance, 1 month prior to disturbance, and 1 month-post disturbance for all years prior to the cyclone of study and after the prior cyclone occurrences.As we did not detect differences between the annual and seasonally adjusted VI changes, we used the annually based VI changes in this study, where the pre-cyclone values were calculated by averaging all available data for the entire year prior to the day before the cyclone event.
Post-cyclone data were extracted as the mean from 1 to 16 d post-cyclone for each case (table 1).This controls for the temporal decrease in cyclone effects seen in most cyclone-affected sites, and is consistent with the period employed to calculate cyclone-induced changes in litterfall (Bomfim et al 2022).In addition, extracting data prior to phenological changes in forest vegetation is imperative for assessing short-to medium-term cyclone-induced effects (Peereman et al 2020).The recovery was extracted as a 25 month time series for the post-cyclone period, with the monthly mean obtained using all available data for each of the 24 post-cyclone months.We used this period to match the ground litterfall data range, which was up to 24 months after cyclone disturbance.

Calculation of cyclone-induced changes in VIs and litterfall
After extracting the pre-and post-cyclone VIs for all cases, we used equation (1) to calculate the annual cyclone-induced normalized ∆VI: where ∆VI is the change in the VI, postVI is the mean post-cyclone VI value and preVI is the mean pre-cyclone VI value.To calculate the changes in LAI and EVI, the pre-and post-cyclone values were substituted into equation ( 1) for the available cases (see supplementary materials for seasonally adjusted ∆VI results).
Positive ∆LAI or ∆EVI values indicate an increase in vegetation index, which could result from increased forest canopy due to rainfall brought by the cyclones or perhaps error in the observations.Interannual variability in the VI can contribute noise to the estimation of cyclone effects, but a positive or zero ∆LAI or ∆EVI likely means that no or very little cyclone damage occurred.Conversely, a negative ∆LAI or ∆EVI is more likely to indicate damage to the area's vegetation due to the cyclone.The percentage change in LAI and EVI (∆LAI% and ∆EVI%) was calculated by multiplying the result of equation ( 1) by 100%.For clarity, the results were presented as percentages.
We used equation ( 2) to calculate the annual (i.e.not considering seasonality) change in total litterfall (∆TL) and assessed its linear relationship with ∆LAI and ∆EVI.We used the pre-and post-cyclone litterfall means from Bomfim et al (2022), who found no difference between the cyclone-induced ∆TL calculated using annual versus seasonally adjusted pre-cyclone means.Therefore, in our study we used the annual ∆TL (log response ratio) as calculated in Bomfim et al (2022).
where ∆TL is the change in total litterfall, postTL is the site-level mean of the post-cyclone total litterfall mass flux in g m −2 d −1 , and preTL is the mean of all available (i.e. from a few months to a few years before the cyclone) pre-cyclone total litterfall mass flux measurements in g m −2 d −1 .Post-cyclone litterfall samples were collected from the baskets immediately after the cyclone event (between 5 and 14 d after the cyclone; Bomfim et al 2022).
We used Pearson's correlations to assess whether cyclone-induced changes in vegetation indices (∆LAI and ∆EVI) correlated with ground-derived litterfall data (∆TL).Correlation coefficients (r) were considered significant at the 95% confidence level.
The mixed-effects meta-analysis model of Bomfim et al (2022) indicated that wind speed and total soil P concentration negatively affect litterfall resistance to cyclones pantropically.To test the relationship between soil P, wind speed, and cyclone-induced ∆VI, we ran mixed-effects models including ∆LAI and ∆EVI as dependent variables in separate models, first specifying soil P and wind speed as fixed effects, and then cyclone name and site as random effects (∼1 | site, ∼1| cyclone).Soil P was log-transformed to obtain a normal distribution.We ran a mixed-effects model with ∆TL as the dependent variable and the same fixed and random effects used in the ∆VI models for comparison.
We used mixed-effects models because our data contained clusters of non-independent observational units that were hierarchical in nature (Harrison et al 2018).We ran these same models with variables scaled and centered prior to model fitting but the results do not change.We ran alternate models with soil P as a random slope, and the small number of observations (21) prevented us from fitting a model with random slopes.We used the Akaike information criterion (AIC) to select the most parsimonious model for ∆LAI, ∆EVI, and ∆TL, with selection based on the lowest AIC value.The best model for ∆EVI and ∆TL included site as a random effect, and the best model for ∆LAI included the cyclone as the random effect.We used the lmer function and lme4 package (Bates et al 2015), and the sjlabelled, sjPlot and sjmisc packages (Lüdecke 2020(Lüdecke , 2021b, Lu ¨decke et al 2021a).
To investigate the relationships between cyclone regime and environmental variables, and ∆VIs, we ran separate principal component analyses (PCAs) for ∆LAI and ∆EVI.For comparison, we ran a PCA including ∆TL.In each PCA, we included the following variables from Bomfim et al (2021, 2022): Holdridge life zone, total soil P concentration, longitude, wind duration, peak wind speed, years since the last storm, elevation, cyclone frequency, parent material, mean annual temperature to mean annual precipitation ratio, and cyclone rainfall.After running preliminary PCAs, variables with low (<1.0)eigenvalues (Peña-Claros et al 2012) were excluded, and only those high enough for consideration (>1.0) were included in the final PCAs.In each final PCA run separately for ∆LAI, ∆EVI, and ∆TL, the variables with high eigenvalues included soil phosphorus, longitude, cyclone frequency, wind speed, wind duration, elevation, the Holdridge life zone, and cyclone rainfall.The PCAs were run using factoextra (Kassambara and Mundt 2020) and ggbiplot (Vu 2011) packages.
As expected, ∆LAI was negatively correlated with change in total litterfall (r = −0.52),indicating that sites with more cyclone-induced canopy leaf damage also experienced a larger negative ∆LAI (figure 3(A)).∆EVI was also negatively correlated with change in total litterfall (r = −0.60).Forest sites that experienced a high cyclone-induced reduction in EVI, or a more negative ∆EVI, experienced a larger increase in litterfall, indicating more damage (figure 3(B)).
The mixed-effects models indicated non-significant (p-value > 0.05) negative relationships between ∆EVI and ∆LAI and total soil P (tables 2(a) and (b)).The best model (i.e.lowest AIC) predicting ∆EVI as a function of soil P included site as a random effect and showed a conditional coefficient of determination (R 2 ) of 0.55.The conditional R 2 is the variance explained by the entire model, i.e. both fixed effects and random effects.The best model for predicting ∆LAI as a function of soil P included the cyclone as a random effect and showed a conditional R 2 of 0.73.In both mixed-effects models, the conditional R 2 was much larger than the marginal R 2 , which describes the proportion of the variance explained by soil P alone.Although soil P was not a strong predictor of ∆LAI or ∆EVI, a mixed-effects model (with site as a random effect) showed that soil P was a significant predictor of ∆TL (p = 0.02; conditional R 2 = 0.61; table S2).Wind speed had a significant effect on ∆LAI (p-value = 0.002) and ∆TL (p = 0.03; conditional R 2 = 0.52; table S3) but not on ∆EVI (tables 2(c) and (d)).
The ∆LAI PCA (figure 4(A)) explained 84.1% of the variance in the first two principal components.The most robust gradient was defined by cyclone regime variables (wind speed and cyclone frequency), while the Both ∆LAI and ∆EVI PCAs showed a regional grouping of tropical cases, where sites in Taiwan were grouped separately from those in the Caribbean and Mexico and the only site in Australia (figures 4(A) and (B)).In both plots, the Caribbean cases correlated with wind speed, whose vector varied in opposite direction to ∆LAI (i.e.negative correlation).The cases in Taiwan were more spread apart along the first axis and were more closely correlated with ∆LAI and cyclone frequency.The same gradients were observed for the total litterfall PCA (figure 4(C)), whereas the second axis appeared to be oriented in the opposite direction to that in figures 4(A) and (B).
The mean pantropical ∆LAI recovered (i.e.post-cyclone LAI values reached pre-cyclone LAI values) within the first four months, with the ∆LAI confidence interval reaching the pre-cyclone level within three months [1, 5 months 95% CI at or above baseline; 3.0 months average] (figure 5(A)).The mean pantropical ∆EVI recovered in the first month, with the ∆EVI confidence interval reaching the pre-cyclone level two months post-cyclone [1, 9 months 95% CI; 3.19 months average at baseline] (figure 5(B)).Ground observations indicated that the pantropical total litterfall mass flux recovered within the first 11 months.We observed that ∆LAI and ∆EVI were highly variable across the studied tropical regions throughout the 25 month post-cyclone period.

Pantropical changes in LAI and EVI correlate with cyclone-induced litterfall peaks
We compiled and analyzed remotely sensed vegetation indices to test whether satellite data would match field litterfall measurements in tropical forests following cyclones (Bomfim et al 2022).Owing to limitations in satellite data availability, we retrieved pre-and post-cyclone LAI and EVI data from a subset of Bomfim et al (2021Bomfim et al ( , 2022)), including nine tropical forests affected by 12 tropical cyclones between 2004 and 2017.Based on this subset, we found that cyclone-induced changes in LAI and EVI correlated linearly with litterfall changes across tropical forests, with larger reductions in LAI and EVI matching higher litterfall peaks.This finding is important because litterfall is a critical indicator for assessing the impacts of disturbances such as Note: σ 2 is the variability across individuals' 'residual' variance, which is the variability that was unexplained by the fixed effects in the model (the fixed effects); t00 is the random effect variance; ICC is the proportion of the variance explained by the grouping structure in the population; N site is the number of tropical forest sites; N cyclone_name is the name of the cyclone; Observations is the number of cases included in this analysis; Marginal R 2 is the variance explained only by fixed effects; Conditional R 2 is the variance explained by the entire model, i.e. both fixed effects and random effects.
tropical cyclones, droughts, and insect infestations on forest ecosystems (Lovett et al 2002, Lin et al 2003, Brando et al 2008).As remote sensing is most useful when validated with ground observations, the high accessibility of MODIS data enables the testing and application of this correlation with field litterfall data in other cyclone-prone tropical forests.Previous studies have successfully correlated MODIS data to litterfall dynamics.Wang et al (2016), for instance, estimated litterfall dynamics in Taiwanese forests over 10 years using MODIS data and reported interannual variation that reflected typhoon canopy damage and recovery dynamics.However, the current challenge is to generalize the magnitude of cyclone impact across tropical forests.Bomfim et al (2022) found that peaks in total litterfall flux following tropical cyclones varied widely among pantropical cases, corroborating earlier reviews (Lugo 2008, Xi and Peet 2011, Mitchell 2013) indicating highly variable cyclone resistance across tropical forest canopies.We also found large variation in cyclone-induced ∆LAI and ∆EVI across pantropical cases.In certain tropical forests, we observed an expected decrease in LAI following a cyclone (i.e. a negative ∆LAI).Conversely, ∆LAI was positive in the Taiwanese cases, indicating rapid greening of sites following cyclone events.This finding corroborates observations indicating rapid ecosystem recovery in typhoon-prone regions (Lin et al 2020, Huang et al 2022), where lower damage magnitudes are associated with greater cyclone frequency (Peereman et al 2021).
Although the correlation between remote sensing vegetation indices and field-measured litterfall is significant, the ∆EVI and ∆LAI of sites with similar ∆TL can vary significantly.This might be caused by different resolutions of the satellite images and field-measured datasets.In addition, remote sensing vegetation indices tend to saturate over dense canopies and are unable to differentiate newly exposed understory vegetation and regrowth on trees that have been stripped of their leaves during cyclones.Therefore, field-measured litterfall is important to validate remote sensing indices.The resolution of the LAI and EVI MODIS datasets was 250-500 m.The resolution of field-measured ∆TL can be as precise as a few meters, meaning that ∆TL can capture finer-scale cyclone-induced changes compared with remote sensing images.Analysis of ∆TL and Landsat-derived VIs at 30 m resolution may solve this problem, but the lack of cloud-free Landsat images over our study regions prevented us from confirming this relationship.

Environmental, soil and cyclone moderators of ∆LAI and ∆EVI
Understanding which and how environmental factors mediate cyclone impact on tropical forests is critical for predicting differential recovery trajectories under future climate change scenarios (Lugo 2008, Mitchell 2013, Xi 2015).Across 46 pantropical cases and 17 moderators, wind speed and total soil P were the key moderators of pantropical litterfall resistance to cyclones (Bomfim et al 2022).In 19 of these 46 cases, we detected larger ∆LAI and ∆EVI values in forests with higher soil P.However, these relationships were not significant when cyclones and sites were considered random effects.Wind speed significantly explained ∆LAI, but not ∆EVI, in mixed-effects models; however, both VIs were negatively associated with wind speed in the PCAs.Therefore, with more observations, wind speed may significantly explain the cyclone-induced changes in both VIs (Negrón-Juárez et al 2014).In addition, cyclone variables, such as wind speed, wind direction, and rainfall (Hall et al 2020), can vary locally to modulate disturbance effects on tropical forests (Cortés-Ramos et al 2020, Hogan et al 2020, Taillie et al 2020, Peereman et al 2022).
We validated our remote sensing data using field litterfall data from Taiwan, the Caribbean, Mexico, and Australia.The observed regional clustering of cases was explained by their contrasting cyclone regimes, with tradeoffs between frequency and intensity.Forests in the Caribbean were associated with less frequent, higher-damage cyclones (i.e.36 m s −1 average wind speeds), while forests in Taiwan were associated with more frequent, lower-damage cyclones (i.e.25.5 m s −1 average wind speeds) (Bomfim et al 2021, Bloom et al 2022).Taiwan is frequently hit by multiple typhoons (Peereman et al 2022), making it possible to observe the effects of more than one cyclone on the Taiwanese forests.Generally, tropical forests are affected differently by both cyclone regimes (less-damaging, frequent cyclones and more-damaging, less-frequent cyclones).Although frequent cyclones tend to be less damaging, they disrupt forest regeneration processes, potentially impacting long-term productivity and composition.The more intense, but less frequent cyclones can cause immediate and extensive damage to forest ecosystems, resulting in long-term changes to their structure and composition (Lin et al 2020).Cyclones can also influence the resilience (ability of tropical forests to bounce back) (Lin et al 2020), as well as their dynamics and functional composition (Hogan et al 2018).
Research suggests that forests exposed to frequent cyclones have lower canopy heights (Ibanez et al 2019, Lin et al 2020), resulting in less exposure to cyclone winds and reduced damage (Peereman et al 2021).Wind-related damage is more common in trees that are exposed to the wind, including those that extend above the forest canopy (Hogan et al 2018).However, at an intermediate frequency of storms, forest demographic resistance to cyclones is generally lower than those in high-frequency areas (Hogan et al 2018).In terms of intensity, Ramsey et al (2001) found a relationship between different forest types and wind speed using simulated winds associated with tropical cyclones.Taillie et al (2020) noted that wind speed is the most important characteristic of hurricanes in explaining mangrove damage; however, taller forests are more prone to strong winds than shorter forests (Lin et al 2020).The soil P vector was orthogonal to the cyclone regime vectors (figure 4) and co-varied with elevation.While teasing apart the role of soil P is more challenging, based on earlier studies suggesting a negative relationship between soil fertility and forest resistance to cyclones (Bomfim et al 2022), future remote sensing studies can test whether taller forests on more fertile soils are, in fact, less resistant to cyclones when controlling cyclone variables such as wind speed and rainfall.

∆LAI, ∆EVI, and litterfall recovery following cyclone disturbance
We compared LAI and EVI recovery across tropical regions and balanced this with the pantropical recovery of field litterfall measurements.Comparing these vegetation indices with litterfall to assess the post-cyclone recovery of tropical forests corroborates studies reporting that typhoons (or tropical cyclones) affect litterfall (Wang et al 2016) and peak LAI (Lin et al 2017) temporal variations.Based on the ∆LAI and ∆EVI post-cyclone trajectories, forest canopy recovery occurs within two months in the studied forests (figure 5(A)).LAI and EVI recovered faster than litterfall, which occurred close to one-year post-cyclone (Bomfim et al 2022).This finding makes sense, as trees first recover leaf production, detected as LAI or EVI, before litterfall rates resume pre-cyclone values.However, a similar cyclone-induced ∆VI can be observed from defoliation or structural canopy damage (e.g.branch and stem breakage), wherein the VI reduction associated with defoliation recovers faster than that associated with branch breakage or tree fall (Smith et al 1994, Radabaugh et al 2020).Therefore, remote sensing VI data complement field litterfall measurements instead of entirely replacing them to monitor the post-cyclone recovery of tropical forests.
Regional patterns of post-disturbance trajectories are also crucial for understanding the impact of tropical cyclones on forest dynamics at the landscape scale.Previous studies have used remote sensing to monitor cyclone-affected forest ecosystem recovery with contrasting results.For example, Chang et al (2020) showed that in Taiwan, forest LAI after a typhoon (or tropical cyclone) event does not always recover before the next event.Lee et al (2008) attributed the high forest resistance and resilience to Typhoon Herb to the high cyclone frequency in Taiwan, where forests have acclimatized to highly frequent typhoons.Similar to our study, Lee et al (2008) found strong correlations between ground and remote sensing data, which is necessary for extrapolation to larger spatial scales and for a greater understanding of carbon cycle dynamics in cyclone-prone regions.

Study limitations
Our analysis did not include the influence of previous or compound disturbances (Heartsill-Scalley and López-Marrero 2021), such as drought (e.g.Beard et al 2005) and fire (Myers and van Lear 1998), on forest responses to cyclones.These legacies may be impactful in forests where cyclone disturbances occur repeatedly over multiple decades, as the effects on stems and branching architecture may limit forest canopy responses to cyclones.Other factors that explain the variation in forest canopy responses to cyclones across tropical forests may include stand properties such as stem density (Ibañez et al 2019) due to tree acclimation to wind and the buffering effects of high winds (Herbohn andCongdon 1993, Mitchell 2013).In addition, remote sensing indexes including LAI and EVI might be affected by the occurrence of the type of the cyclone impact.For example, remote sensing instruments may detect downed/fallen green leaves, which might take days to weeks to senesce after the arrival of cyclones.Similarly, instruments might pick up the greenness of newly exposed understory vegetation in the early stages of recovery from cyclones.As we used remote sensing data to detect the impact of cyclones on these forests with varied cyclone regimes, we note that remote sensing indices can identify category 4-5 hurricanes more easily than category 3 hurricanes.

Conclusions
Our investigation of the links between field and remote sensing data to quantify and monitor forest responses to cyclones is increasingly important as climate change progresses and catastrophic tropical cyclone seasons are likely to become more frequent (Bakkensen and Mendelsohn 2019, Wehner et al 2018, Knutson et al 2020, Reed et al 2020).We bridged the field-remote-sensing response and recovery data validation challenges by using nine tropical forest sites affected by 12 tropical cyclones, and tested the effects of total soil P, cyclone and environmental variables on tropical forest canopy damage following cyclone disturbances.Field litterfall measurements correlated with ∆LAI and ∆EVI values, suggesting these indices may be used more broadly to assess cyclone-induced canopy effects corresponding to litterfall pulses.Based on the ∆LAI and ∆EVI post-cyclone trajectories, forests recovered within two months, whereas pantropical litterfall resumed pre-cyclone rates nearly one year after the cyclone.Therefore, remote sensing assessments of recovery likely represent canopy recovery but not forest function.
We found a regional clustering of cases that related to their contrasting cyclone regimes, with tradeoffs between frequency and intensity: Caribbean forests were associated with less frequent, higher-damaging cyclones, while Taiwanese forests were associated with more frequent, less-damaging cyclones.We also found higher ∆LAI and ∆EVI in tropical forest sites with higher soil P, but these relationships were not significant when accounting for the non-independence between observations.Further exploration of whether trees in forests on soils with high soil P are generally taller and thus more exposed to wind damage is needed, as changes in the frequency of cyclones where regimes are characterized by low frequency, high-intensity storms can have implications for the resilience and carbon cycling of those forests.Taken together, our findings demonstrate that remote sensing observations complement but do not substitute for ground observations that reveal cyclone damage and post-cyclone recovery in tropical forests, and that soil phosphorus moderates some but not all metrics of stability in response to cyclones.

Figure 1 .
Figure 1.Workflow representing the steps taken from obtaining vegetative indices to statistical analysis of remote sensing and field data.500 m LAI and 250 m EVI were the vegetation indices chosen for further analysis in this study.

Figure 2 .
Figure 2. Changes in LAI (A) and EVI (B) by tropical case (unique combination of site and cyclone event).Points are colored by region.The vertical dashed line in each panel indicates no change in LAI or EVI.

Figure 3 .
Figure 3. Relationships between pantropical cyclone-induced changes in total litterfall (∆TL; y-axis; unitless) and (A) ∆LAI (x-axis; ∆LAI at 500 m resolution; unitless) and (B) ∆EVI (x-axis; ∆EVI at 250 m resolution; unitless).Pearson's correlation coefficients (R) between ∆TL and each vegetation index, with their respective p-values, are shown in each panel, where the solid lines indicate significant correlations (p-value < 0.05).Pre-cyclone values, represented by horizontal and vertical gray lines, indicate no change in LAI, EVI, or total litterfall values.

Figure 4 .
Figure 4. PCA results for (A) ∆LAI, (B) ∆EVI, and (C) ∆TL across the tropical cases.The regions Australia, Caribbean, Mexico, and Taiwan are represented by different colors in the legend key, and the associated ellipses for these groups are shown in the PCA plots.The axes of the PCA plots show the percentage of variability explained by each principal component, while the relational direction and length of the arrows show the direction and strength of the correlations between variables.

Figure 5 .
Figure 5. Post-cyclone recovery of LAI and EVI across tropical regions.Solid lines show the pantropical mean ∆LAI (blue), ∆EVI (purple), and ∆TL (black), and the shaded areas represent their respective confidence intervals.The solid gray line illustrates the pre-cyclone condition, where lines close to zero indicate full recovery following the cyclone event.

Table 1 .
Dates of LAI and EVI extraction for each case.The case names contain unique combinations of sites and cyclones.The annual pre-cyclone, post-cyclone, and cyclone recovery columns have start and end extraction date ranges (start date [to] end date).All of the dates are in year-month-day format.

Table 2 .
Mixed-effects model outputs including ∆LAI and ∆EVI as the dependent variables, log-transformed soil phosphorus (a)-(d) and wind speed (c), (d) as fixed effects, and site (a), (c) and cyclone (b), (d) as random effects, respectively.Two cases did not have wind speed data so models (c) and (d) presented 19 observations instead of 21.