Effects on Vegetation Heterogeneity

Except for SAVI in relation to Typhoon Herb, typhoons led to a higher heterogeneity in VI values, as indicated by the significantly greater post-typhoon CV of all VIs except for NDVI for Typhoons Aere and Dujuan, wherein the CV did not change significantly (Figure 3.3).

Figure 3.3. Changes in coefficient of variation (CV) of four vegetation indices following the five typhoons. Asterisks (*) indicate significant differences between pre- and post-typhoon CV based on bootstrap comparisons on means (5000 iterations).

53 3.3.4. Topography and Disturbance Severity

Topographical variables differed in their ability to explain the variation of ΔVIs among the typhoons (Table 3.6 for Nari and Herb, Table F1 for other typhoons).

Topography was a better predictor of ΔNDVI for Typhoons Herb and Nari (adjusted R² = 0.52 and 0.47, respectively) than it was for other ΔVIs (adjusted R² ≤ 0.34). For Typhoons Aere, Dujuan, and Soudelor that passed further from the FEF (Figure 3.1), topographical variables were poor predictors of ΔVI values (adjusted R² ≤ 0.16).

Elevation and slope were significant in explaining ΔVI variation in all cases except slope for ΔEVI and ΔSAVI for Typhoon Dujuan (p > 0.05, Table 3.6 and Table F1);

however, coefficients were small (β < 0.0002 for elevation, β < 0.008 for slope). The magnitude of ΔVI decreased with increasing topographic slope for all VI–typhoon combinations, except for ΔEVI with Typhoons Herb and Nari, and ΔSAVI with Typhoon Herb. However, increasing elevation led to either higher or lower damage (i.e., change in VI) depending on the typhoon in question, but the change in direction was consistent among VIs for each typhoon. Except for flat slopes, all TPI positions had positive and statistically significant (p < 0.05) regression slope coefficients for Soudelor, Nari, Herb and Aere. On the other hand, Typhoon Dujuan showed negative relationships for all TPI positions except for flat slopes. Nevertheless, the sign of regression slope coefficients remained consistent among ΔVIs for each typhoon. The relationships between aspect and ΔVI values changed among typhoons, showing no clear pattern.

54

Table 3.6. β Coefficients of ordinary least squares linear models between topographical variables and changes in vegetation index values (ΔVIs) associated with Typhoons Nari and Herb. Sample size of 4593 (equal to the number of pixels from analyzed Landsat images). Significance levels are shown with ^† (p < 0.05),

‡ (p < 0.01), and * (p < 0.001). Only the first non-zero value is shown, complete results for the five typhoons are in Table F1.

Topography Nari Herb steepness. Elevation was negatively related to EVI-based damage frequencies (β = −6.71, p < 0.001, adjusted R² = 0.016) but not to NDII-based frequencies (p = 0.47).

On the other hand, slope was positively related to EVI- and NDII-based frequencies (β = 0.32, adjusted R² = 0.001, p = 0.01 for EVI; β = 2.16, adjusted R² = 0.07, p < 0.001 for NDII). The proportion of damage frequency varied among aspects (Figure 3.5, Table F2). For NDII, all aspects were significantly different in damage frequency (p < 0.001), except for south–east, northeast–north, southwest–northeast, southwest–north, and west–northwest (p > 0.05, Table F2). For NDII, southwestern to northwestern aspects

55

had the highest damage frequencies and southeastern aspects had the lowest disturbance frequency (Figure 3.5A). Considering EVI, northern aspects had more frequent typhoon damage than western and southern aspects, whereas there were no significant differences between other aspects (Table F2, Figure 3.5B). Among slope positions, according to EVI, the flat slope, lower slope, and valley positions had similar typhoon damage frequencies (Table F2); these three topographic positions were more frequently damaged than other positions. For EVI, ridge areas had lower typhoon damage frequency than all other slope positions (Table F2). With NDII-based frequencies, no significant differences were observed between ridges and slope or among other topographical positions (p > 0.05 with adjustment, Table F2).

Figure 3.4. ΔNDII (A) and ΔEVI (B) for five typhoons and across the six damages frequency classes. Different characters above boxes indicate significant differences between frequency classes based on multiple comparison of means with p-adjustment (Herberich et al., 2010).

56

Figure 3.5. Percent of pixels of the eight aspects in different typhoon damage frequency classes (0 to 5) based on NDII (A), and EVI (B).

3.3.6. Recovery

Vegetation cover had recovered in less than a year after Nari, as average VI values measured in June 2002 were equal to or greater than their values in June 2001 (positive 95% CIs, Figure 3.6). Similarly, all Vis, except NDII, registered vegetation recovery in less than one year after typhoons Soudelor–Dujuan (Figure 3.6). In contrast, EVI was the only VI showing recovery after Aere (Figure 3.6).

57

Figure 3.6. One-year recovery of the Fushan Experimental Forest as shown by the vegetation indices measured before typhoon passages and a year later at the same season. Recovery of Typhoons Soudelor and Dujuan were merged as they passed during the same season; no satisfying images were available to study Typhoon Herb recovery. All the differences between pre-typhoon and post-typhoon images are significant based on 95% confidence intervals measured through bootstrapped comparisons on means.

3.4. Discussion

3.4.1. Consistency in the Damage Effects Among Typhoons

The consistent decrease in VI values following all five typhoons (with very few exceptions, see Table 3.4) suggest that all the VIs can generally capture typhoon-induced losses in vegetation cover (Ayala-Silva & Twumasi, 2004; Kang et al., 2005; Lee

58

et al., 2008; Rossi et al., 2013; Tsai & Yang, 2016; Hu & Smith, 2018). However, we find weak correlations in ΔVIs between different typhoon events (ρ < 0.4, Table E2), indicating that typhoons range in their effects on vegetation cover. Although this is not surprising because typhoons differ in intensity, duration, trajectory and occurrence time relative to plant phenology, the inconsistency among vegetation indices highlights that results derived from one or a few disturbance events are unlikely to represent general trends in disturbance effects. Indeed, our results show that the successive Typhoons Soudelor and Dujuan did not have consistent effects on vegetation although they had comparable paths, wind speeds, and directions (Figure 3.1, Figure 3.2, Table 3.1). Additionally, the very weak negative correlations of the ΔVIs between typhoons Soudelor and Dujuan (Table E2) indicate that the two typhoons had different effects on the vegetation cover, as observed in other sites subject to successive cyclones (Elmqvist et al., 1994; Burslem et al., 2000; Liu et al., 2018).

Powerful cyclones, such as Hurricanes Hugo and Maria in Puerto Rico, Typhoon Herb in Taiwan, and Cyclone Larry in northeastern Australia, attract scientific study of their ecological effects (Brokaw & Grear, 1991; Lodge et al., 1991; Walker, 1991; Lin et al., 2003b; Kang et al., 2005; Grimbacher et al., 2008; Lee et al., 2008; Metcalfe et al., 2008; Turton, 2008; Liu et al., 2018; Uriarte et al., 2019; Hall et al., 2020⁾. This high prevalence of studies has advanced our understanding of cyclone ecology, especially in relation to the most powerful and damaging storms (reviewed by Everham &

Brokaw, 1996; Lugo, 2008; Lin et al., 2020). However, in this study, Typhoon Herb was considered to be the most powerful typhoon in several decades (Lin & Jeng, 2000).

Typhoon Herb had the greatest wind speed among the five typhoons studied, while Nari was most intense in terms of precipitation. Considering these two storms, the resultant changes in vegetation index values (ΔVIs) were only weakly correlated.

Future increases in the frequency of the most intense cyclones are predicted (Knutson et al., 2010; Knutson et al., 2015; Walsh et al., 2016). However, our results suggest that conclusions drawn from studies that document the effects of a single or a few intense cyclones are insufficient for predicting the effects of future cyclones.

59

The fact that there were small or no changes in VIs associated with Typhoon Herb is somewhat surprising. In addition to the potential influence of image quality on our analyses, the timing of Typhoon Herb is probably the most important factor. Typhoon Herb occurred in the summer (late July) when the forest was in the middle of the main growing season, and the two images were approximately two months apart. Thus, plant growth could be substantial during the period, potentially confounding the detection of typhoon-caused decreases in vegetation cover by VIs. This contrasts with Typhoon Dujuan, which caused the largest decreases in VIs. Typhoon Dujuan occurred near the end of the main growing season (April to September) so that although the two images were also approximately two months apart, there was little vegetation growth during the period. As a result, typhoon-induced vegetation loss can be better detected with the VIs. Thus, image timing in relation to plant growth phenology should be considered when examining disturbance-induced changes in vegetation cover using VIs.

Not only were the overall effects inconsistent among typhoons, there was large variation in linear regression results, which showed that typhoon damage–topography relationships were inconsistent among typhoons (Table 3.6 and Table F1). Topography was better at explaining disturbance distribution across the landscape for Typhoons Herb and Nari, the typhoons that passed the closest to the FEF, but they were not the storms that caused the greatest degree of vegetation damage. This result suggests that topography–vegetation damage relationships vary with cyclone distance and that topography is a key determinant of vegetation damage only when typhoons are very close to the study site, despite the magnitude of the damage. It also suggests that factors other than cyclone distance determine the severity of typhoon-induced vegetation cover damage.

Nevertheless, there were some consistencies across typhoons. First, the canopy generally recovered quickly as many VIs returned to their pre-disturbance values within a year (Figure 3.6). This result is consistent with the report of the rapid recovery of the FEF observed following Typhoon Bilis using NDVI (Kang et al., 2005). However,

60

recovery of a VI does not imply total canopy recovery as LAI and litterfall typically do not recover within a year of damage in the FEF (Lin et al., 2017). Second, the relationship between high pre-disturbance NDVI and strong NDVI loss observed by Kang et al. (2005) was also detected for all five typhoons in this study despite their differences in paths and intensities (Table E1). This pattern may be explained by the higher aerodynamic drag of dense canopies as suggested by Harrington et al. (1997), who reported a similar relationship between pre-disturbance LAI and LAI loss (see also Herbert et al., 1999). Third, cyclones are disturbance agents which induce heterogeneity in forest landscapes (Lugo, 2008; Lin et al., 2011). Secondary tree falls and defoliation may have led to the increased heterogeneity observed following almost all typhoons examined here. Finally, most ΔVIs were positively correlated between typhoons (Table E2), except for Soudelor–Dujuan, suggesting that different typhoons could have similar effects although the strength of the relationships varied greatly among typhoons and VIs.

3.4.2. ΔVIs in Relation to Topography

The topography only explained a small proportion of the variation in vegetation cover change. As observed by Negrón-Juárez et al. (2014), steeper slopes were associated with greater decreases of all VIs across all typhoons (Table 3.6 and Table F1), perhaps because of different wind exposures and soil stability (e.g., landslides, McEwan et al., 2011). However, our results show that the relationships between vegetation damage and other topographical variables changed among typhoons and VIs, even when typhoons had similar tracks and wind directions. Thus, observations from a single event should be generalized with caution, as they do not necessarily remain true for other cyclones. Indeed, complex interactions between wind and topography have occurred in the FEF, as leeward positions were more affected than eastern positions for Typhoon Nari but not for Dujuan (Figure 3.2, Table 3.6 and Table F1). Less exposed slopes may also offer little protection if the cyclone is particularly strong (Hall et al., 2020). The overall low adjusted-R² values indicate that factors other

61

than topography (e.g., vegetation conditions and timing of typhoon disturbance) likely play a more important role in determining typhoon-induced changes in vegetation cover.

3.4.3. Typhoon Disturbance Frequency

The greater damage for more frequently than less frequently affected cells (Figure 3.4) is consistent with a study from a moist forest of Puerto Rico, in which trees sustaining heavy damages from Hurricane Hugo (1989) were more likely to be damaged again nine years later by Hurricane George (Ostertag et al., 2005). The result is also consistent with the mostly positive correlations between ΔVI values among the different typhoons (Table E2). Although it is possible that sites with certain topographical features within the FEF are more prone to typhoon damage, this claim is not supported by the lack of consistent topographical damage patterns among typhoons (Table 3.6 and Table F1). Possibly, susceptibility to damage is more related to weakened vegetation in these sites (i.e., repeated disturbances). Further field-based research may help to separate the effect of forest characteristics, topography, and typhoon frequency on canopy damage at the FEF.

The greater elevational signal in typhoon damage frequencies at lower than at higher elevations based on EVI fits the pattern of greater damage frequency at lower topographic positions (lower-slope, valley) than other positions (e.g., middle- and upper-slope). It also supports results from field observations of greater typhoon effects at low than at high elevations along a 2300-m elevational gradient in central Taiwan (Chi et al., 2015). It appears that the elevational pattern is consistent across spatial scales including variation in slope across the landscape, possibly because the intensity of typhoons declines quickly in rough topography as they move upward (Chi et al., 2015).

The differences in NDII and EVI sensitivity to vegetation characteristics may explain the different relationships of their damage frequencies with slope. As observed in Puerto Rico (Boose et al., 2004), different aspects had different disturbance frequencies, with northern aspects having significantly higher values. However, the relationship

62

with the windward–leeward direction was less clear here, as western aspects (leeward in the FEF) also had high damage frequencies. The very rough topography at the FEF, with a mean slope of 38%, probably obscures relationships between slope, aspect and typhoon damage.

3.4.4. Consistency Among Vegetation Indices

We did not detect any VI that was consistently the most sensitive across the five typhoons. EVI and NDII were considered particularly functional to measuring vegetation characteristics such as canopy structure (Huete et al., 2002) and water content (Anderson et al., 2004; Cheng et al., 2006). However, in our study, they did not detect strong decreases in vegetation cover relative to pre-typhoon values for Typhoon Herb. Such differences may be the product of different damage variations across disturbance events, as reported for other sites (Elmqvist et al., 1994; Burslem et al., 2000;

Liu et al., 2018). It may also be due to the phenological change in leaf properties between pre- and post-disturbance images (Lin et al., 1997; Chang-Yang et al., 2013), although effort was made to minimize such variation by selecting images from within seven weeks of typhoon events.

Surprisingly, NDVI was the only VI that showed a decreasing canopy cover following Typhoon Herb, whereas it was the least practical index for Dujuan and Nari, despite its lack of sensitivity under high LAI (Huete et al., 1997a; White et al., 1997; Gao et al., 2000). NDVI was also the only VI that did not detect change in vegetation cover heterogeneity following Typhoons Aere and Dujuan (Figure 3.3). Overall, NDVI is less sensitive to vegetation cover change in the FEF than the other Vis, as reported for other sites (Wang et al., 2010; Huete et al., 2011; Townsend et al., 2012; Zhang et al., 2016), and its sensitivity varied with typhoon events or images. Large variability between VIs in their responses to disturbance also took the form of varying correlation strength and direction between and within typhoon events (Table 3.5). However, ΔVIs based on the same spectral bands, such as ΔEVI–ΔSAVI, had stronger correlations than in other cases (e.g., ΔEVI–ΔNDII). The disparity between ΔVIs for a given typhoon suggests

63

that cross-study comparisons of disturbance effects based on several or different VIs could be problematic.

ΔVIs shared some consistencies, nonetheless, as most VIs indicated increasing heterogeneity, and varied somehow similarly with elevation, slope, and TPI-derived classes for each typhoon (Table 3.6, Table F1). In addition, both ΔNDII and ΔEVI detected increasing severity for more frequent canopy damage.

Our study has several constrains. First, although we compared the consistency among different VIs, we could not identify which VI was more accurate in detecting vegetation changes caused by typhoons due to the lack of ground truthing.

Conducting ground truthing is difficult, because although the beginning and the end of typhoon seasons are relatively well understood, typhoon events themselves are unpredictable, hence it is difficult to coordinate ground surveys across the studied landscape without knowing whether a disturbance will occur and where clouds would be present on the satellite images. Although it was possible to conduct ground truthing for past disturbance events, we strongly recommend routine ground surveys on several widely spread plots for studies that aim to assess disturbance effects on vegetation whenever possible. Such ground truthing could help to identify which VI would provide most accurate detection of disturbance-induced vegetation changes.

Second, some of our results may have been affected by cloud obstruction in analyzed images, which is substantial for all images used in this study (17.5%–48.8%).

A previous study showed that cloud contamination does not distribute across the FEF, with more cloud cover at higher elevations (Peereman et al., 2020). Because typhoon-induced vegetation change varied with elevation (Table 3.6), the observed topographical patterns of vegetation change are likely affected to some degree by the non-random cloud contamination. Unfortunately, cloud contamination is common in the FEF and most humid forests. In fact, cloud cover prevented the inclusion of Typhoon Herb in our analyses of vegetation recovery. Third, ideally, images should be derived from the same sensor as different sensors vary in the width of their spectral bands (Irons et al., 2012) and radiometric calibrations (Markham & Helder, 2012;

64

Markham et al., 2014). However, we were constrained by the availability of high-quality (low cloud contamination) images because we studied five typhoons spanning over two decades. With the rapid advancement of remote sensing, high-quality images from the same sensor should be increasingly available.

3.5. Conclusion

Comparison of the effects of five typhoons affecting the Fushan Experimental Forest (FEF) showed substantial differences as well as some consistency in their effects.

First, while typhoons all led to decreases in vegetation index (VI) values, the magnitude of change (ΔVIs) differed among events. The variability of ΔVIs among typhoons may be the product of complex interactions of their characteristics with landscape topography and the biotic conditions when the forest was disturbed (e.g., recovering from previous disturbance, soil moisture). Indeed, topography alone did not explain variation in ΔVIs among typhoons, and its explanatory power varied among the different indices. However, all typhoons shared the same positive relationship between damage severity and pre-disturbance vegetation condition, and all typhoons resulted in increased vegetation heterogeneity. Hence, conclusions drawn from the remote sensing studies of one typhoon may or may not stand for other typhoons in the same landscape, depending on the aspects of the effects under concern. Second, observation of greater disturbance severity for more frequently damaged cells of the FEF shows that some sites are more prone (i.e., have decreased resistance) to disturbances than others. On the other hand, the landscape generally has high resilience in order to maintain its forest cover over the many centuries of typhoon disturbance. Third, the four vegetation indices had different relationships with canopy cover damage, probably because of their different sensitivities across the light reflectance spectrum. The Normalized Difference Vegetation Index (NDVI) was, overall, less sensitive to change, which supports the findings of previous studies pointing to its saturation and overall limited sensitivity under high-biomass forest conditions. The Soil-Adjusted Vegetation Index (SAVI) and the Enhanced Vegetation

65

Index (EVI) were highly correlated, but EVI was less related to the other two indices.

Although the Normalized Difference Infrared Index (NDII) and EVI had the same change in direction after the five typhoons, they differed in the magnitude of change.

Hence, we suggest using them together as complementarity VIs, because NDII and EVI are, respectively, based on short-wave infrared and near-infrared, making them sensitive to different characteristics of vegetation.

Funding

Funding