Integration of Worldview-2 and Lidar Data to MAP a Subtropical Forest Area: Comparison of Machine Learning Algorithms

This work is committed to explore the integration of airborne LiDAR data and WorldView-2 (WV-2) images to classify land cover and land use in a rural area with the presence of a subtropical forest. Different methods were used for this purpose: two artificial neural networks (ANN) and three decision trees forests. The results demonstrated that the inclusion of LiDAR data significantly improved the classifications in all methods. Excluding the Convolutional Neural Network, the classification algorithms had a nearly similar performance, and none of them achieved the best accuracy for all adopted classes. Forest by Penalizing Attributes (FPA) attained the best general result, with a Kappa index of 0.92, while Rotation Forest obtained the best result in the classification of the two vegetation classes.