INVESTIGATING THE EFFECTS OF AUTOCORRELATION IN DATASET CONSTRUCTION ON TREE SPECIES CLASSIFICATION USING FIELD AND AIRBORNE HYPERSPECTRAL DATA

Spectral measurements provide a non-invasive and efficient means of mapping and monitoring vegetation species over large areas thanks to supervised classification. However, the presence of autocorrelation due to the incorporation of spatial or contextual information in the data can lead to overestimated classification performance and bias the assessment of the generalization power of the classification. This study evaluates this bias by comparing the performance obtained with different methods of training and testing data splitting at leaf and image levels for tree species classification. The average overall accuracies obtained from different data splitting methods differed by up to 30% for airborne image data (ranging from 68% with low correlation to 98% with high correlation) and 16% for field data (ranging from 80% to 96%). These differences demonstrate the need to consider the autocorrelation impacts for classification purpose.