Comment on bg-2021-102 Anonymous Referee # 2 Referee comment on " Strong temporal variation in treefall and branchfall rates in a tropical forest is explained by rainfall : results from five years of monthly drone data for a 50-ha plot

The authors present a unique analysis of canopy disturbances over the Barro Colorado Island 50-ha plot using a high-temporal density drone dataset. The high temporal resolution of this dataset allows the authors to relate the occurrence of canopy disturbance events to meteorological conditions with far greater precision than was previously possible with 5-year census intervals. The authors (surprisingly) conclude it is not horizontal wind speed, but high rainfall intensity events that cause canopy disturbances. Overall I think this is a very interesting analysis of a unique dataset, but I think it suffers from some analytical pitfalls that limit its utility for forest dynamics. I believe this will be a notable contribution if these issues can be addressed.


112
We used approximately monthly orthomosaics and canopy surface models produced from drone-acquired imagery to 113 analyze temporal variation in canopy disturbance rates in the 50 ha plot between 2 October 2014 and 28 November 2019. RGB 114 imagery was collected using a variety of drones and cameras over the years, with a horizontal spatial resolution of 3-7 cm. Imagery 115 for each sampling date was processed using the photogrammetry software Agisoft Metashape to obtain orthomosaics and surface 116 elevation models, which were then aligned vertically and horizontally (details in Text S1).

117
We defined a canopy disturbance as a substantial decrease in canopy height in a contiguous patch of canopy occurring 118 over one measurement interval, such as typically results from a treefall or branchfall. We identified canopy disturbances through 119 a combination of analysis of the canopy surface model changes and visual interpretation of the orthomosaics (Fig. 1). We first 120 differenced surface elevation models for successive dates to obtain a raster of the canopy height changes for the associated interval 121 ( Fig. 1, Text S1). We then pre-delineated major canopy disturbances by filtering for areas in which canopy height decreased more 122 than 10 m in contiguous areas of at least 25 m 2 (the minimum area for canopy gaps in previous studies by Brokaw (1982) and

137
We calculated the total number and area of canopy disturbances within the BCI 50 ha plot during the five years of the 138 study. In calculating the number and total area of disturbances, we included all disturbed areas that were inside the plot boundaries 139 (if a disturbance was on the boundary, only the area inside the plot was included). Our analyses of temporal variation employed 140 the same definitions for numbers and areas of canopy disturbances within the 50 ha plot. For analyses of the size structure of 141 disturbances, we included the complete areas of disturbances whose centroids were located within the plot (i.e., we excluded 142 disturbances centered outside the plot, and included area outside the plot for disturbances centered inside the plot to avoid artifacts 143 related to reducing disturbance size by trimming at the plot boundaries). 144 145 2.4 Temporal variation in canopy disturbance rates and its relation to climate

146
We calculated canopy disturbances rates for each measurement interval as the percentage of area disturbed per month 147 (i.e., per 30-day period). Specifically, we summed the total area disturbed during the measurement interval, and divided by the 148 total area of the plot and the length of the time interval. We excluded one excessively long interval (237 days) from all analyses of 149 temporal variation; the remaining intervals ranged from 14 to 91 days, with a median of 31.5 days (Table S1). We also calculated 150 an incidence canopy disturbance rate as the number of canopy disturbances per hectare per month. We calculated the mean, minimum, maximum, and the 25 th , 50 th , and 75 th percentiles of interval length in days, number and area of canopy disturbances, 152 and the respective monthly rates.

153
We compared canopy disturbance rates between wet and dry seasons and between early wet and late wet seasons. We 154 defined the dry season as January 1 to April 30 (rainfall < 100 mm mo -1 , Fig. S3), the early wet season as 1 May to 31 August, and 155 the late wet season as 1 September to 31 December. Intervals that straddled more than one season were classified to the season in 156 which they had more days. We tested for homogeneity of variances using the Levene test, and for differences between means using 157 the two-tailed Student's t-test for the log-transformed canopy disturbance rates.

158
We evaluated the relationship of temporal variation in canopy disturbance rates with temporal variation in climate 159 extremes using linear regressions. We regressed canopy disturbance rates (area per time) against the frequency of extreme rainfall

179
We quantified the size distributions of canopy disturbances by fitting three alternative probability distributions: 180 exponential, power, and Weibull. Recognizing that our methods may miss smaller disturbances, we fit these distributions to 181 truncated datasets, excluding disturbances below 2, 5, 10 or 25 m 2 . (Note that 25 m 2 is the minimum area for defining a canopy 182 disturbance in our automated pre-delineation algorithm, and we are confident we captured all disturbances above this area.) We 183 binned the data into 1 m 2 classes, and fitted each distribution to each truncated dataset using maximum likelihood, as described in

187
For the last three years, for which we classified each canopy disturbance as being a branchfall, treefall, or standing dead 188 tree, we evaluated the relative contributions of branchfalls vs. treefalls. We did not include standing dead trees in the analysis 189 because our methods possibly missed many standing dead trees. We separately calculated treefall and branchfall disturbance rates 190 for each interval, and relative contributions to their summed number and area. We regressed branchfall disturbance rates against 191 treefall disturbance rates, for both area-and number-based rates, and calculated their Pearson correlations.  (Fig. S4). The mean rate of canopy disturbance creation was 916 m 2 mo -1 (range of 75 m 2 mo -1 to 8040.9 m 2 mo -1 ) and median 499 m 2 mo -1 (other statistics in Table S1).

209
The single highest disturbance rate was observed between 1 June and 13 July 2016, when 11,257 m 2 of disturbances were created 210 in just 42 days (a rate of 268 m 2 day -1 ). A full 2.3 % of the total area of the plot was converted to new canopy disturbances during

218
Rates of canopy disturbances were higher during the wet season (p = 0.03; Fig. 4a). There was no significant difference 219 in rates between the early and late wet season (p = 0.27, Fig. 4b). Very high rates of disturbance (> 0.3 % per month) were observed 220 only in the wet season.

225
The best predictor of temporal variation in canopy disturbance rates was the frequency of 1-hour rainfall events above the 226 99.4 th percentile, here 35.7 mm hour -1 , which explained 45 % of the variation (Fig. 5a). This threshold outperformed all other tested 227 rainfall thresholds (all percentiles from 90.0 to 99.9, by 0.1 % of the different frequency time scales - Fig. 5b). Only two of these 228 high rainfall events occurred during the same day (Table S2)

245
Canopy disturbances with larger areas tended to have larger mean decreases in canopy height (Pearson r = 0.39, Fig. 6b).

246
The size distribution of canopy disturbances was close to a power function for areas above 25 m 2 , and was relatively flat 247 over the range of 5 to 25 m 2 (Fig. 6c). The fitted exponent of the power function was -1.96 for canopy disturbances above 25 m 2 , 248 but the Weibull distribution provided a better fit than the power function (Table 1). When distributions were fit to data including 249 smaller size classes (> 2 m 2 , > 5 m 2 or > 10 m 2 ), the distribution is further from a power function; the Weibull remains the best fit, 250 the exponential becomes the second-best fit, and the power function the worst fit of the three (Fig. S7, Table S3).

277
The use of high frequency (approximately monthly) drone imagery enabled us to quantify temporal variation in canopy 278 disturbance rates and to quantify the sizes of canopy disturbances at high temporal and spatial resolutions. We found that canopy 279 disturbance rates of the BCI 50 ha plot varied strongly over time, and were higher in the wet season. The frequency of extreme 280 rainfall events was the single best predictor of monthly variation in canopy disturbance rate during the 5-year study period. In  canopy disturbance rates are lower (Fig. 4a, Fig. S1). We hypothesize that extreme high rainfall is associated with both saturated 295 soils, increasing risk of uprooting, and with gusts having high horizontal and vertical windspeeds that increase stresses on tree 296 crowns. Future studies should include high frequency measurements of vertical and horizontal windspeeds and soil moisture to 297 better capture proximate drivers, and evaluate mechanistically formulated predicted models that include multiple variables. methods are expected to capture all treefall and branchfalls above this threshold, but we may increasingly have missed smaller 342 events, especially below ~ 5 m 2 . However, we consider it unlikely that this is a sufficient explanation for the shortfall in small 343 trees, and suggest that it is more likely explained largely by the low frequency of small trees and branches in the canopy of this 344 mature tropical forest, and thus a scarcity of small treefall and branchfall events.

345
Although rarely quantified, branchfall is an important ecological process, with major contributions to woody turnover and 346 necromass production. We found that branchfalls were almost as common as treefalls in number, although they contributed a 347 substantially smaller total area of disturbance. Similarly, a ground survey of 78 canopy turnover events in a Brazilian Amazon forest found that 44 % were branchfalls, and that they accounted for 15 % of the total affected area (Leitold et al., 2018). In contrast, 349 a landscape level analysis of LiDAR data concluded that branchfalls were seven times more frequent than treefalls and accounted repeat drone imagery acquisitions. Here we applied these methods to 50 ha of old-growth tropical forest for five years, and 364 analyzed the resulting products to quantify major drops in canopy height such as those created by branchfalls and treefalls, and 365 thus calculate the canopy disturbance rate. We found that canopy disturbance rates are highly temporally variable, and are well-366 predicted by extreme rainfall events. Even higher temporal resolution canopy dynamics data together with higher frequency three-367 dimensional wind data would enable an even stronger assessment of the link to storm conditions, and additional analyses of the 368 photogrammetry data could shed light on standing tree mortality. The expansion of these methods to additional and larger areas, 369 potentially in part through citizen science initiatives, has great potential to improve our understanding of tropical forest tree 370 mortality, and the future of tropical forests under changing climate regimes.

372
Code and data availability. Analysis codes, input data and output results are available at