Including spatial information using texture variables can improve land cover classification. In this study, forests as sensed by Landsat TM sensors were simulated with four different tree crown sizes and four different average tree crown cover densities. Classification using variances computed in a 5 by 5 pixel windows (called 'local variance') had 50 percent higher accuracies as compared to not using the local variances. When local variances were used Poisson sampling of pixels increased accuracies over traditional sampling in blocks of pixels. Each of the four replicates of the 16 different simulated forest types was analyzed for spatial autocorrelation of errors, comparing the distribution of errors with an expected binomial distribution. For images without underlying spatial autocorrelation and classified without local variances, spatial autocorrelation of errors was not higher than expected by chance. However, for images classified with local variances, misclassified pixels were spatially autocorrelated more than 90 percent of the times.