Multibeam acoustic backscatter has considerable utility for remote characterization of spatially heterogeneous bed sediment composition over vegetated and unvegetated riverbeds of mixed sand and gravel. However, the use of high-frequency, decimeter-resolution acoustic backscatter for sediment classification in shallow water is hampered by significant topographic contamination of the signal. In mixed sand-gravel riverbeds, changes in the abiotic composition of sediment (such as homogeneous sand to homogeneous gravel) tend to occur over larger spatial scales than is characteristic of small-scale bedform topography (ripples, dunes, and bars) or biota (such as vascular plants and periphyton). A two-stage method is proposed to filter out the morphological contributions to acoustic backscatter. First, the residual supragrain-scale topographic effects in acoustic backscatter with small instantaneous insonified areas, caused by ambiguity in the local (beam-to-beam) bed-sonar geometry, are removed. Then, coherent scales between high-resolution topography and backscatter are identified using cospectra, which are used to design a frequency domain filter that decomposes backscatter into the (unwanted) high-pass component associated with bedform topography (ripples, dunes, and sand waves) and vegetation, and the (desired) low-frequency component associated with the composition of sediment patches superimposed on the topography. This process strengthens relationships between backscatter and sediment composition. A probabilistic framework is presented for classifying vegetated and unvegetated substrates based on acoustic backscatter at decimeter resolution. This capability is demonstrated using data collected from diverse settings within a 386 km reach of a canyon river whose bed varies among sand, gravel, cobbles, boulders, and submerged vegetation.