Lidar data processing: We used an object-based image segmentation approach to estimate the aboveground woody biomass from the LiDAR derived CHM 20. The CHM was used to identify treetops using an algorithm based on local maximum filter; a dynamic moving window approach was used to scan the CHM, and if a given cell was found to be the highest within the window, it was tagged as a treetop. The height of each treetop was retrieved from the CHM. The algorithm was trained to ensure that trees with height > 5 m were accurately detected, which contribute most to the aboveground woody biomass. Based on this algorithm, the LiDAR-derived stem density (> 5 m) was highly correlated with field-measured stem density (R2 = 0.90). Next, tree crowns were segmented based on a watershed algorithm guided by the locations of the treetops 53. The area of each tree crown was calculated as the projection of the outlined crown on the ground surface. After computing height and crown area, a regression model developed specifically for savanna trees in Kruger was applied to generate aboveground biomass per individual tree: log10(m)= β_0+ β_1*log10⁡(A_obj)+β_2*log10(A_obj)^2+ β_3*log10⁡(H_obj) eqn (1) Where m is tree biomass (kg), Aobj object-projected crown area (m2), and Hobj is tree height (m). β0, β1, β2, and β3 are least-square regression coefficients and are estimated as 0.115, 0.161, 0.252, 1.73, respectively. In this study, we set a minimum height of 0.5 m to ensure we detected only woody plants. We used the coordinates of the center of each GPR plot (i.e., 10 m × 10 m) to set a buffer of 15 m radius to compute aboveground woody biomass of each plot. The 30 m diameter LiDAR plot allows for small location errors of GPR plot and also matches the area harvested for developing the regression model as presented in equation (1). Only woody plants with their centroids within the buffer were included. In addition, to cross-validate the object-based LiDAR biomass estimation, we also estimated aboveground woody biomass using species-specific allometric equations based on diameter at breast height 54 as well as plot-averaged pixel-based LiDAR calibration 20. However, we chose to report object-based LiDAR biomass estimation in the main text, since the object-based method is better suited to open-canopy systems and provides a more accurate proxy for measured woody biomass than allometric estimates 20. All analyses were performed with the ForestTools and raster packages in R 3.6.1 software. Files included: Canopy Height Model (CHM) for three treatments across four strings. The files are in tiff format. Three treatments include burned annually in August (e.g., aug_b1), burned triennially in August (i.e., aug_b3), and unburned treatment (i.e., control). Four strings include Fayi, Kamebni, Numbi, and Shabeni. Files also include shapfiles for shrub biomass, shrub crown, tree biomass, and tree crown. For more details see the lidar data processing section Parameters in lidar derived woody biomass excel file include four strings (Fayi, Kamebni, Numbi, and Shabeni), three treatments (Annual, Triennial, Unburned), tree biomass (Mg/ha), tree carbon storage (Mg C/ha), shrub biomass (Mg/ha), and shrub carbon storage (Mg C/ha). Parameters in location excel file include X and Y coordinates for longitude and latitude, and four strings and three treatments listed above. Parameters in pixel-based biomass excel file include four strings and three treatments as listed above and woody biomass (Mg/ha) and woody carbon storage (Mg C/ha). Parameters in tree cover and height excel file include four strings and three treatments as listed above and tree cover (%) and tree height (m). GPR data processing: Post-collection data processing was performed with Radan 7 software (Geophysical Survey Systems Inc., NH, USA) using the processing steps shown in Supplementary Fig. 8. Roots were detected as hyperbolic reflectors in the radar profiles. The aim of post-collection processing was to maximize the coherency of root reflectors and differentiate them from the soil background. After basic edits of radar profiles, we applied an exponential gain function to recover amplitude losses from geometric spreading and soil absorption of the radar impulse. A background removal was applied to filter out horizontal reflections resulted from the ground surface, soil horizons, and bands of low frequency noise. A Kirchoff migration was used to decompose and compact the geometry of hyperbolic reflectors to their source points that are closer to the actual features. The Hilbert transform further collapsed the point diffraction amplitudes and improved background clutter removal. After these processing techniques, GPR profiles were converted to image files. In order to estimate coarse lateral root biomass from GPR data, we applied an image analysis technique to quantify root reflections and used root mass from soil cores for calibration. We selected and marked 16 points with a range of root biomass within each 10 m × 10 m plot. Eight points were selected based on their distance to the nearest tree, and the remainder were identified a priori with GPR to exhibit either high (4 points) or low (4 points) incidence of reflections, encompassing the range of root biomass within the plot. Each point was scanned with the 1.6 GHz antenna on both X and Y directions. The location of each point was electronically marked on the radar profile as the antenna was pulled over the center of the point. After the collection of GPR profiles at each point, a large soil core (15 cm in diameter and 50 cm in length) was used to retrieve coarse roots (> 2 mm) to a depth of 50 cm. Roots were oven-dried (65°C for 72 hours) for biomass. GPR profiles were processed and converted to image files. In order to develop the linear regression equation between root biomass and GPR amplitude, image files were cropped to 15 cm wide sections where the antenna was directly over the location of each point. Pixel intensity, a relative measure of how dark or light a pixel is at a grayscale of 0 (black) to 1 (white), was used to differentiate root reflectors and background within each segment. We used an intensity threshold of > 0.8 to delineate roots with minimum illumination of unwanted clutter. We counted pixels with intensity higher than the threshold within each segment to a depth of 50 cm (hereafter, GPR index, pixels with threshold range). For each point, GPR amplitudes from two scanning directions were averaged. Root biomass retrieved from soil cores was then correlated to GPR index for each EBP string (n = 48) to develop regression lines for plot-level biomass estimation. Plot-level image files were sequentially sectioned into 15 cm segments corresponding to the dimensions of the calibration core, hence a 10 m GPR profile yields 67 segments or unique observations. Since the majority of roots were constrained to the top 60 cm, we calculated the GPR index for each segment to a depth of 60 cm and assigned coordinates to each segment. We then applied the regression line to estimate root biomass for each segment. Ordinary kriging was used to interpolate plot-level coarse lateral root biomass based on segment data and their coordinates. The final product was an average of estimate from X and Y scanning directions. The uncertainty of coarse lateral root biomass was estimated from the 95% confidence intervals of the regression lines. All image analyses were performed with the package “EBImage” in R 3.6.1 software. Files included: FA: Fayi string burned annually FN: Fayi string unburned FT: Fayi string burned triennially KA: Kambeni string burned annually KN: Kambeni string unburned KT: Kambeni string burned triennially NA: Numbi string burned annually NN: Numbi string unburned NT: Numbi string burned triennially SA: Shabeni string burned annually SN: Shabeni string unburned ST: Shabeni string burned triennially Files in core biomass folder include GPR calibration data for three treatments across four strings. Parameters within each excel file includes treatment (i.e., FA represents Fayi burned annually, see above for other treatments and strings), soil cores from 1 to 16 sampled across 10 m by 10 m plot, soil depth from 0-50 cm with 10 cm depth increment, coarse root biomass (g, > 2 mm in diameter) and fine root biomass (g, < 2mm in diameter) and total root biomass (g) Files in FA folder include raw images from each ground penetrating radar profiles, see the above GPR data processing for more details Files in FA folder include resized images from each ground penetrating radar profiles, see the above GPR data processing for more details Files in FA folder include results from GPR data processing. FA_1600_calibration: core, number of cores 1-16, X_amplitude, GPR amplitude on X direction, Y_amplitude, GPR amplitude on Y direction, biomass, root biomass (g) FA_depth_summary: depth in cm from 0-60, amplitude indicates GPR amplitude, percent indicate percentage of amplitude at certain depth to the total amplitude throughout the 60 cm soil profile. FA_biomass: X indicates X direction, Y indicates Y direction, amplitude indicates GPR index, biomass indicates estimated from the GPR amplitude, direction includes X direction and Y direction. These parameters applied to all excel files in this folder. Files in vegetation survey folder include four site/string as mentioned above, three treatments as mentioned above, x direction, y direction, dbh, tree diameter at breast height (cm), tree height (m), species name, abbreviations can be found in the specieslist file, which include species name, 6 letter code and 2 letter code.