 The study shows that random forest algorithm for classification of remotely sensed data is highly sensitive to input data characteristics, training data selection, specific input variables, and class proportions. High dimension datasets should be reduced to only uncorrelated important variables, independent error assessments should be used to evaluate RF results, iterative classifications are recommended to assess stability, and the size of the training dataset also impacts the accuracy of the image classification. The study recommends using randomly distributed or created training datasets with minimal spatial autocorrelation to improve classification results and mitigate inflated estimates of RF out-of-bag classification accuracy. This article was authored by Karine Millard and Murray Richardson.