In this study, we used images obtained by Unmanned Aerial Vehicles (UAV) and an instance segmentation model based on deep learning (Mask R-CNN) to evaluate the ability to detect and delineate canopies in high density orange plantations. The main objective of the work was to evaluate the improvement acquired by the segmentation model when integrating the Canopy Height Model (CHM) as a fourth band to the images. Two models were evaluated, one with RGB images and the other with RGB + CHM images, and the results indicated that the model with combined images presents better results (overall accuracy from 90.42% to 97.01%). In addition to the comparison, this work suggests a more efficient ground truth mapping method and proposes a methodology for mosaicking the results by Mask R-CNN on remotely sensed images.