Satellite-based deforestation maps are often created using training data from image interpreters. These interpreters will look at high resolution imagery from before and after a date of interest to see if forest loss occurred. However, what they define as forest loss can vary with factors like forest type and their familiarity with the region. This disagreement between interpreters can confuse a machine learning model since we are giving it conflicting information. Here we analyze how prevalent this disagreement is and suggest ways to adapt to this known uncertainty.