Stream restoration practitioners often rely upon empirical models to quantify annual streambank erosion rates and identify streambank erosion hotspots. Such models are designed to be widely applicable by incorporating readily available field measurements, but they must be calibrated to each hydrophysiographic region and may not reflect the dominant streambank erosion processes in a given region. Here, we present statistical models for streambank erosion using physical and environmental data collected at 53 locations throughout the northern Gulf of Mexico coastal plain. The data include channel geometry, bank characteristics, precipitation, above-ground biomass density, and root density, the latter two surveyed using techniques introduced here. We developed a statistical model selection process using Akaike's Information Criterion (AIC) and repeated cross-validation (CV). Models derived from the literature that were applied a priori were only weak predictors of erosion rate, but AIC-CV model selection identified 3 strong statistical models. The best model according to AIC showed a significant correlation to lateral streambank erosion rates (R2 = 0.54) and included the five strongest covariates of our dataset (bank slope, biomass density, curvature index, BEHI, and understory cover). When volumetric erosion rate (m2/year) was predicted, the fit of this model increased (R2 = 0.65). CV-based selection resulted in a more conservative model with the four strongest covariates and a lower fit (R2 = 0.47). The similarity of the AIC and CV models indicates the stability of the two-tier model selection approach, and suggests it has utility for modeling phenomena with many potential variables. Our models also showcase the ability of our biomass survey to quantify root reinforcement of streambanks. Our approach incorporates measurements familiar to the stream restoration community and can be applied throughout the northern Gulf of Mexico coastal plain, a region characterized by low relief fluvial valleys, unconsolidated alluvium and meandering single thread sand bed channels. The approach, which is based on field observations and robust statistical modeling, offers an alternative for stream restoration practitioners to more traditional streambank erosion prediction methods that underperform in the region, and may have applicability elsewhere.