Recordings of deep Tonga earthquakes from two arrays of 12 broadband seismographs each in the Fiji and Tonga islands are stacked and searched for reflections and conversions from upper mantle discontinuities in the Tonga subduction zone. The arrays operated as part of the Seismic Arrays in Fiji and Tonga (SAFT) experiment from July 2001 to August 2002. In comparison with the commonly used teleseismic approaches, the short path lengths for the local data provide smaller Fresnel zones and high-frequency content for precise mapping of discontinuity topography and sharpness. To enhance the low-amplitude discontinuity phases s410p, P660p and S660p, deconvolved seismograms from each event/array pair are aligned on the maximum amplitude of the direct P wave and subsequently slant stacked. For the 410-km discontinuity, the results show no systematic variations in depth with distance to the cold slab. The 660-km discontinuity varies between 656 and 714 km in depth. For the southern and central parts of the subduction zone, the largest depths occur in the core of the Tonga slab. For the northern part, two separate depressions of the 660-km discontinuity are observed. These anomalies are interpreted as being induced by the active, steeply subducting Tonga deep zone and a subhorizontally lying remnant of subducted lithosphere from the fossil Vityaz trench, respectively. Interpreting the deflections of the 660-km discontinuity in terms of local temperatures implies a thermal anomaly of -800¿K to -1200¿K at 660 km depth. Except for the southern region where it may thicken, the width of the depressed 660-km discontinuity region implies that the Tonga slab seems to penetrate the 660-km discontinuity with little deformation. Waveform modeling suggests that both the 410- and 660-km discontinuities are sharp. The 660-km discontinuity is at most 2 km thick in many parts of the region, and a first-order discontinuity cannot be precluded. The 410-km discontinuity thickness shows somewhat more variability and ranges from 2 to 10 km outside the slab and is at most 10 km thick within the slab. This suggests that the subduction process does not produce dramatic effects on the sharpness of the discontinuities.