The size and organization of bed material, bed texture, is a fundamental attribute of channels and is one component of the physical habitat of aquatic ecosystems. Multiple discipline-specific definitions of texture exist and there is not a universally accepted metric(s) to quantify the spectrum of possible bed textures found in aquatic environments. Moreover, metrics to describe texture are strictly statistical. Recreational-grade side scan sonar systems now offer the possibility of imaging submerged riverbed sediment at resolutions potentially sufficient to identify subtle changes in bed texture with minimal cost,expertise in sonar, or logistical effort. However, inferring riverbed sediment from side scan sonar data is limited because recreational-grade systems were not designed for this purpose and methods to interpret the data have relied on manual and semi-automated routines. Visual interpretation of side scan sonar data is not practically applied to large volumes of data because it is labor intensive and lacks reproducibility. This thesis addresses current limitations associated with visual interpretation with two objectives: 1) objectively quantify side scan sonar imagery texture, and 2) develop an automated texture segmentation algorithm for broad-scale substrate characterization. To address objective 1), I used a time series of imagery collected along a 1.6 km reach of the Colorado River in Marble Canyon, AZ. A statistically based texture analysis was performed on georeferenced side scan sonar imagery to identify objective metrics that could be used to discriminate different sediment types. A Grey Level Co-occurrence Matrix based texture analysis was found to successfully discriminate the textures associated with different sediment types. Texture varies significantly at the scale of ≈ 9 m2 on side scan sonar imagery on a regular 25 cm grid. A minimum of three and maximum of five distinct textures could be observed directly from side scan sonar imagery. To address objective 2), linear least squares and a Gaussian mixture modeling approach were developed and tested. Both sediment classification methods were found to successfully classify heterogeneous riverbeds into homogeneous patches of sand, gravel, and boulders. Gaussian mixture models outperformed the least squares models because they classified gravel with the highest accuracies.Additionally, substrate maps derived from a Gaussian modeling approach were found to be able to better estimate reach averaged proportions of different sediments types when they were compared to similar maps derived from multibeam sonar.