Temporal analysis

We calculated temporal trends for all the above datasets for each TMCF using observations from 1997 to 2020. For this purpose, hourly observations were aggregated to an annual average. Linear regressions were then performed to calculate trends (i.e., slope or rate of change) (Δ, year-1). Regional studies tend to evaluate climatic trends from ERA5 using monthly average (i.e., Lei et al., 2020; Yilmaz, 2023), here we use annual average instead to provide a global perspective of temporal changes in low-clouds and others ECVs. The latter was performed using the IBM PAIRS Geoscope platform using a pixel-based approach. This cloud-based platform enables the deployment of user-defined functions on ERA-5 and CHIRP datasets without downloading raw data but obtains the regression coefficients directly (Lu et al., 2016). Trends were extracted from 1997 and not previous decades (e.g., 1940) given the availability of data in IBM PAIRS Geoscope and the reliability of products from recent decades that leverage on new available remote sensing and meteorological observations (Yilmaz, 2023). After extracting the trends, we performed Bayes one-sample t-tests (Kruschke, 2013) to compute mean estimates of trends and evaluate the probability of these differing from zero. In addition, we performed an analysis at the realm level to determine how trends in low-clouds and ECV depend on the biogeography of these ecosystems. Bayes one-sample t-testanalyses were performed on R (R Core Team, 2023) using 30000 Markov chain Monte Carlo iterations in the BayesianFirstAid package (Bååth, 2014). All statistical analyses (including those in the following section) were weighed by the pixel projected area according to the TMCFs location to account for the latitudinal variation of the pixel area.