Data from "Groundwater changes the climate niche of an Amazonian mound-building termite"

This dataset was collected to test for an interaction between climate and groundwater in shaping the abundance of the Amazonian termite, Amitermes excellens, while accounting for other environmental effects. Termite mound abundance was remotely surveyed (using Google Street View), while environmental data were collected from published products. There are two files: "termite_mound_data.csv", which contains the raw data; and "termite_mound_code.R", which contains the R script used to perform all analyses and plots.

Study area
The study was performed in the Guianan Savanna ecoregion, a complex of savannas in northern Amazonia whose largest continuous part (61.664 km²) occurs between Brazil (known as "Lavrado"), Guyana (known as "Rupununi Savanna"), and Venezuela (known as "Gran Sabana"). Most of the area belongs to the Brazilian state of Roraima (43.358 km² or 70%).

Termite mound abundance
Google Street View was used to count A. excellens mounds along roads in the region. Google Street View is a tool from Google Maps and Google Earth that offers panoramic, high-resolution, colored images along streets and roads around the world, with images taken since 2007. At the time of this study, Street View imagery was available only for the Brazilian road network in the region and for a single year (June do December 2012). The network included the following roads: BR-174 (Brazilian federal road), and RR-203, RR-205, RR-319, RR-342, RR-343, RR-401, and RR-433 (Roraima state roads). Accordingly, sampling sites were regularly distributed every 5 km along these roads, except for Boa Vista and its surroundings where the landscape is highly urbanized. Overall, 162 sites were surveyed over ca. 795 km.
In May 2024, each site was surveyed for A. excellens mounds using Google Street View in Google Earth Pro (version 7.3.6). First, all sites were scored for visual obstruction (0 for low, 1 for moderate, or 2 for high) based on the presence of trees, tall grass or any other physical obstacle adjacent to the road that could interfere with mound detection. Then, all A. excellens mounds within 45 m from the observer were counted across the 360º view (i.e. 90-m diameter circumference or 0.64 ha), with the observer remaining fixed in the site coordinates. This observation distance was determined from the known distance of fences along the roads and assured that mounds were easily recognized. Most sites had low obstruction (n = 131), and visual obstruction did not correlate with any of the predictors used in the analysis (r ≤ 0.25). Thus, only sites with low obstruction were used in this study, which guaranteed easy mound detection without biasing the data relative to predictors.

Environmental data
For each sampling site, the following environmental variables were obtained from published geospatial products: soil sand content, vegetation density, fire frequency, Indigenous Land occurrence, annual rainfall, and water table depth. Soil sand content (%) was obtained from SoilGrids 2.0, which predicts soil features across the world at a 250-m resolution from environmental covariates using machine learning. Vegetation was represented by the Normalized Difference Vegetation Index (NDVI), which measures photosynthetic activity and is a common proxy for plant biomass and productivity in tropical savannas. NDVI ranges from -1 to 1 and is measured each 16 days at a 250-m resolution using spectral data from NASA’s Terra MODIS sensor. NDVI was obtained from NASA’s Application for Extracting and Exploring Analysis Ready Samples (AppEEARS) and averaged for the decade prior to Google Street View images (January 2003 to December 2012). Fire frequency and Indigenous Land occurrence were obtained from MapBiomas, a Brazilian project that provides annual maps of land cover and use for the country since 1985. Fire frequency is estimated as the number of years in which fire scars were detected at a 30-m resolution using Landsat satellite images. Fire frequency was considered for the period prior to Google Street View images (1985 to 2012), averaged to a 250-m resolution for consistency with the other local environmental variables. MapBiomas also provides a shapefile of currently recognized Indigenous Lands in Brazil, which was used to classify each sampling site as inside or outside any of them.
Annual rainfall (mm) was obtained from WorldClim 2.1, which predicts climate variables for the world from global climate model simulations fined-tuned with whether station data. This variable (BIO12 in the WorldClim nomenclature) represents the average from 1970 to 2000 and was extracted at 5-minute resolution (ca. 9.25 km) given the high spatial autocorrelation of climatic variables at finer resolutions. Lastly, water table depth (m) was obtained from a simulated global map. The simulation used a model relating groundwater to surface water and plant water use, constrained by direct well measurements and ran hourly from 2004 to 2014. Values are means over this period at ca. 1-km resolution.