Improving forest type discrimination with mixed lifeform classes using fuzzy classification thresholds informed by field observations

This paper presents a three-stage methodology to mitigate uncertainty in forest lifeform classification using a case study in the mixed hardwood-conifer forest of Massachusetts, USA. First, two fuzzy membership surfaces representing the proportion of conifer and hardwood lifeform dominance were created using a supervised multilayer perceptron neural network algorithm. Second, an index of lifeform membership was generated using a ratio of the membership surfaces of conifer and hardwood forest. Lastly, this index was thresholded using field measurements of forest lifeform proportion to delineate pure conifer, mixed conifer-hardwood, and pure hardwood categories. This methodology produced a map of forest lifeform with 94% overall accuracy (kappa 0.88 for hardwood, 0.97 for conifer, and 0.97 for mixed), an improvement of 10% over a map generated using a top-down method using mixed forest training sites. Per-class accuracies increased approximately 5% for both the pure hardwood class, 26% for the pure conifer class, and 16% for the mixed class. The improvement in map accuracy was due to improved spectral discrimination of lifeforms, which results in a more geographically plausible map.