*2.6. Phenological Metrics*

With the data ingested, Prophet separates the seasonal components of the time series from the trend and the residual components. The seasonal component of the GAMs is used as an approximation of phenological cycles and several techniques can be adapted to extract the start, end, and duration of the mangrove "green-up" season. In this study, we adopted similar definitions to those by [44] to identify the Start of Season, End of Season and Peak Growing Season, however, as our study does not involve a sinusoidal curve the definitions vary slightly. For simplicity, we define the start of season and end of season as the lowest points, and peak growing season as the highest points of the de-trended time series (Figure 3). We also define the length of the growing season as the time between the start of season and end of season. Because we use start of season, end of season and peak growing season as phenological metrics of the landscape, they do not represent individual species.

**Figure 3.** Panel (**A**) shows every available Enhanced Vegetation Index (EVI) observation for every pixel in the 17-ha region of interest from February 1995 to December 1996 for the Gladstone region. Panel (**B**) shows the median and standard deviation of the observed EVI values in grey dots and lines respectively, and the apparent phenology (i.e., GAM) in red. Panel (**C**) shows the apparent phenology, the definitions of start and end of season (SOS, EOS), peak growing season (PGS) and length of the growing season (LGS). Shaded areas represent the wet season months.

We extracted the start of season and peak growing season for each pixel in our field study site in the following way: from the seasonal component, we selected the predicted index values from the GAMs that were lower or higher than the 5 or 95 percentile as the potential start of season or peak growing season dates, respectively (Figure 3C). Then we selected the median of the image acquisition dates as the start of season, peak growing season and end of season dates. In case the selected date was not a date in which an image was acquired, we searched for the image with the closest date and used that date instead.

Because we wanted to determine if the GAMs were correlated with biophysical processes described in the literature (i.e., leaf gain, leaf loss, net leaf production), we decided to shift the models (i.e., displace the models along the time axis) by one, two and three months. We then examined if the EVI response was immediate or delayed. An immediate response of EVI to a biophysical process would imply that remote sensing techniques could be used for real-time monitoring. In contrast, a delayed response would help us understand which processes drive the changes in EVI. After comparing the biophysical processes to the EVI, we examined their relationship using linear regressions.

#### *2.7. Validation of the GAMs*

We assessed the precision of our model by running linear and non-linear regressions between the apparent phenology and (1) observed EVI values from satellite imagery; (2) in situ data from [26]; and (3) leaf fall, leaf gain and net leaf production values from published literature. We did this using the Scikit-Learn package for python [45], specifically, we used simple linear regressions and support vector regression with linear, polynomial and radial basis function kernels. We chose support vector regression because it is robust against outliers, it is easily implemented, and it allowed us to compare linear and nonlinear relationships between apparent phenology and the data measured in the field. In addition, we performed 5-fold cross-validation using the performance metrics functionality provided by "Prophet" (see the Supplementary Information section, Figure S2.
