Tropical savannas encompass 14% of the terrestrial biosphere and are an important regulator of the carbon cycle. Tropical savannas are characterized by rainfall seasonality leading distinct rainy and dry seasons. The dry seasons can become prolonged and lead to extended drought conditions. Climate change is expected to cause more intense, frequent, and longer droughts in the future, which has major implications for ecosystems and human livelihoods that depend on them. In this study, we mapped the long-term sensitivity of gross primary production (GPP) to drought at different temporal scales ranging from short (1 month) to long (24 months) intervals using newly developed GPP data from the Global Inventory Modeling and Mapping Studies (GIMMS) project and the Standardized Precipitation and Evapotranspiration Index (SPEI). The 35-year (1982-2016) trend in the relationship between GPP and SPEI was quantified for the chosen temporal scales. A machine-learning algorithm based on random forests was used to identify the hierarchical importance of environmental covariates in explaining the GPP-SPEI trend. The results show an increase in the sensitivity of GPP to drought conditions at short temporal scales (1-6 months) whereas sensitivity was low to at both intermediate (9-12 months) and long (21-24 months) scales. Surprisingly, we found a resurgent increase in sensitivity at the 15-month time scale. We disaggregated the trends by land cover type and found that the largest increases sensitivity at the 95% level was exhibited by shrubland, grassland, and rainfed cropland. With respect to environmental covariates that could explain the mechanisms behind the observed trends, we found that vapor pressure deficit (VPD) dominated the hierarchy as it contributed the largest increase in mean standard error of the random forest permutations. These results contribute to a deeper understanding of the sensitivity of carbon uptake to droughts and the whether past drought events impact photosynthetic capacity of tropical savannas.This study was funded by the Swedish Research Council, grant number 2018-00430.