An automated method for large-scale monitoring of seed dispersal by ants

Myrmecochory is the process of seed dispersal by ants; however, it is highly challenging to study, mainly because of the small size of both partners and the comparatively large range of dispersal. The mutualistic interaction between ants and seeds involves the former retrieving diaspores, consuming their elaiosome (a nutrient-rich appendage), and the rejection of seeds from the nest. Here, we introduce a semi-automated method based on stitching high resolution images together, allowing the study of myrmecochory in a controlled environment over time. We validate the effectiveness of our method in detecting and discriminating seeds and ants. We show that the number of retrieved diaspores varies highly among colonies, and is independent of both their size and activity level, even though the dynamics of diaspore collection are correlated with the arrival of ants at the food source. We find that all retrieved seeds are rejected from the nest in a clustered pattern, and, surprisingly, they are also frequently redispersed within the arena afterwards, despite lacking elaiosome. This finding suggests that the dispersal pattern might be more complex and dynamic than expected. Our method unveils new insights on the mechanisms of myrmecochory, and could be usefully adapted to study other dispersal phenomena.


Initial processing of the data
Using USE Tracker software, diaspores and ants were indistinctly detected by background subtraction algorithm on each frames of the 3 hours movies of the platform (Figure S1 B). Numerical output for each frame of the movie consisted of the total number of pixels detected, corresponding to both diaspores and ants. Values collected on each frame were aggregated on a 10 seconds basis (i.e. total number of pixels detected in 50 consecutive frames, movies being at 5 frames/s; Figure S2 B), before computing the mean number of detected pixels D track (t) using a rolling mean of window k = 300 s (Figure S2 B).
The detected pixels correspond to both diaspores and ants as where D pix (t) and A pix (t) are the number of pixels corresponding to the diaspores and the ants respectively. Computing the number of diaspores remaining on the platform Dp(t) is then executed in two steps: 1. determine the number of ants on the platform 2. determine the relative size of diaspores and ants

Number of ants on the platform
The estimated number of ants located on the platform at each time step A est (t) (Figure S2 A) was determined by using the timing of entrances and exits of workers (data of incoming flow and outgoing flows collected by hand with USE Tracker). The accuracy of A est (t) was then assessed by counting the number of workers located on the platform A count (t) at different time steps t ∈ [30; 60; 90; 120; 150; 180] minutes. In 6 of 9 replicates, the estimated number of ants A est (t) at t=180 min was higher than the number obtained by a manual counting as A est (t) = A count (t) + ∆, with ∆ > 1 ant (Figure S2 A).
The error ∆ was due to ants leaving the platform unnoticed, by falling or exiting the platform while out of sight (e.g. walking on lower face of the access ramp). Since ∆ was linearly related to the total number of ants A total that had entered the platform at the end of the foraging phase (∆ = 0.080 A total − 1.48, R 2 = 0.96, F 1,7 = 203.1, P < 0.0001), we computed a corrected value of the number of ants on the platform A corr (t) ( Figure S2 A) from A est (t) by subtracting an amount of ants (i.e. those leaving the platform unnoticed) proportional to a constant fraction Ψ of the cumulated number of workers A cum (t) that had entered the platform at time t as The value of Ψ was different for each replicate and has been determined by optimisation, so as to minimize the difference between A count (t) and A corr (t) computed as with t ∈ [30; 60; 90; 120; 150; 180]. The values of Ψ ranged from 0.04 to 0.085 (n=6 replicates). The correction with Ψ allowed a reduction of the values of RSS computed with A corr ranging from 60% to 95% compared (n=6 replicates) to that obtained with A est (t) (i.e. ∑ 180 t=30 (A est (t) − A count (t)) 2 ); Figure S2 A). After this step, the number of ants on the platform was available to compute the remaining number of diaspores.

Number of diaspores on the platform
The mean number of pixels δ corresponding to a diaspore was computed at t=0 when only the 200 diaspores were on the platform. By pooling the values of pixels Pix(t), counted diaspores D count (t) and ants A count (t) from each replicate at 6 different times t ∈ [30; 60; 90; 120; 150; 180] minutes, we determined the ratio ρ between diaspores and ants size using the following linear relationship Pix(t)/δ − D count (t) = ρ A count (t) (linear regression: ρ = 3.49, R 2 = 0.94, F 1,53 = 914.4, P < 0.0001). Finally, the number of diaspores Dp(t) on the platform was computed as Dp(t) = Pix(t)/δ − ρ A corr (t).
As a last step, the number of diaspores on the platform was smoothed using a rolling mean of window k = 15 minutes, and the resulting value D track (t) was used in the analysis.  Figure S2. Estimation of (A) ants A corr (t) and (B) diaspores D track (t) on the platform. A Number of ants A est (t) predicted by incoming and outgoing ants flows shows an important error ∆ at the end of the experiment, which is corrected A corr (t) by constantly subtracting a fraction Ψ of the total number of ants A cum (t) that have reached the platform. B Number of pixels detected by the tracker was aggregated (sum) over a 10 s duration, and smoothed with a rolling mean of window k = 300 s (Pix(t)). Estimate of the number of diaspores remaining on the platform D track (t) was computed before comparison with manually counted numbers D count .

2/6
number of ants detected total number of pixels

S2. Simulations of seed dispersal Model
Our analysis indicates that the seeds are released isotropically within the arena (see Seed rejection, in Results). This finding is in agreement with dispersal of other wastes away of the nest, such as soil particles during nest excavation 1 and corpses disposal. 2 The results also show the importance of redispersal all along the rejection stage, but these findings were unexpected and our data do not contain any quantification about the individual behaviours involved. We thus model the seeds rejection and purposely ignoring redispersal behaviours (catch/drop of previously released seeds), to investigate the ability of the model of craters building through soil rejection 1 to generate clustered patterns. The procedure consists of centrifugal seeds dispersal, isotropically distributed around the nest (seeds are expelled from nest in random directions, following straight lines). Two alternate models based on this procedure are investigated : 1. blind drop model. Ants follow a straight exit trajectory from the nest, seeds dropping is independent from previous deposits 2. reactive drop model. Ants follow a straight exit trajectory from the nest, seeds dropping is dependent of previous deposits within a short scale (5 mm): when encountering seeds within this scale, ants drop their load in the vicinity of this previous deposit (1 mm apart).
In each simulations, 200 seeds are released from the nest located at (0;0), into the simulated arena of 2000 mm radius. Each seed is released along a straight trajectory oriented at a random angle θ with θ ∈ [−π;π[, and drop away from the nest at a distance (cm) sampled from the experimental density function of seed distance to the nest. The final spatial pattern is characterized using G(d) function and DCLF test (see Data analysis, in Material and Methods section). We realize 1000 simulations for each model with the same number of seeds (200) to be released.

Simulation results
The simulations indicate that the blind drop model produces clustered pattern in only 5.9% of the simulations (59 of 1000 simulations). On the contrary, the reactive model produces clustered patterns in 97.9% of the simulations (977 of 1000 simulations), what is much closer to the ratio of clustered pattern observed in the experiments (10 of 12 replicates).
The distribution of the nearest neighbor distance reflects these results: experience as well as reactive drop model produce a significant fraction of short distances (between 10 and 30% of all distances are below 1 mm); blind drop model on the contrary exhibits a more uniform distribution of distances ( Figure S4 B & C).
If the blind model shows radial densities of seeds in agreement with the one observed in the experiments (as expected since simulated rejection distances follow the empirical distribution of distances), on the contrary the reactive model exhibits more seeds between 25 and 40 cm from the nest (Figure S4 A). This suggests that most of the clustering in the reactive model occurs in this region ( This indicates that the reactive model partially reproduces the observed patterns: it is able to generate clustered patterns, but the location of the clusters is not in complete agreement with the experimental patterns. We hypothesize that the redispersal behaviours could play a significantly role in the clustering process and its absence in our model could explain the partial disagreement between our experiments and the simulations. → S track = #seeds → θ = angle within arena → = distance from nest → nnDist = seed-to-nearest-seed distances -S end count = #seeds rejected at end (last picture) Table S1. Summary of raw extracted data and variables used in the analysis. Raw collected data are sorted by location (Foraging platform/Arena; rows) and collection method (Automatic/Manual; lines), and the variables extracted from them are detailed after the arrow →. Measurement interval is given in parenthesis.