Ecologists often use transect surveys to estimate the density and abundance of animal populations. Errors in species classification are often evident in such surveys, yet few statistical methods exist to properly account for them. In this paper, we examine biases that result from species misidentification when ignored, and we develop statistical models to provide unbiased estimates of density in the face of such errors. Our approach treats true species identity as a latent variable and requires auxiliary information on the misclassification process (such as informative priors, experiments using known species, or a double-observer protocol). We illustrate our approach with simulated census data and with double-observer survey data for ice-associated seals in the Bering Sea. For the seal analysis, we integrated misclassification into a model-based framework for distance-sampling data. The simulated data analysis demonstrated reliable estimation of animal density when there are experimental data to inform misclassification rates; double-observer protocols provided robust inference when there were "unknown" species observations but no outright misclassification, or when misclassification probabilities were symmetric and a symmetry constraint was imposed during estimation. Under our modeling framework, we obtained reasonable apparent densities of seal species even under considerable imprecision in species identification. We obtained more reliable inferences when modeling variation in density among transects. We argue that ecologists should often use spatially explicit models to account for differences in species distributions when trying to account for species misidentification. Our results support using double-observer sampling protocols that guard against species misclassification (i.e., by recording uncertain observations as "unknown").