Studies with the purpose to identify and remotely classify native tree species play a strategic role in management, surveillance, administration, and conservation of these species, contributing to assess their cover extent and spatial distribution in a faster way. In this sense, the greater detailing attained by the spatial resolution of the latest generation sensors is crucial for this end, since their images enable the detection and delimitation of individual trees crowns (ITC). In this case, a tree is represented by a group of pixels that refer to smaller portions of the target of interest, presenting a great variability of spectral responses. Hence, it is preferable to merge them by means of segmentation, so as to make each segment corresponding to an ITC. However, in order to have a segmentation algorithm that generates segments accounting for meaningful image objects, such as an ITC, its parameters ought to be properly tuned. Conventionally, this stage supposes a time-consuming trial and error procedure. In this context, some initiatives for the automatic search of segmentation parameters have been developed. Among them, the stochastic optimization method is to be mentioned, which searches within the parameters space those values that maximize the level of agreement between a set of reference segments, manually defined by the user, and the segmentation result. This level of agreement is assessed by a metric that compares the segmentation result with the reference segments. In face of this, the objective of this work is to automatically estimate the segmentation parameters of different segmentation algorithms, and by means of suitability functions, identify the best fit for the delimitation of trees crowns belonging to the native species Araucaria angustifolia, located in an area of the subtropical Atlantic Forest in Southern Brazil. In order to achieve this goal, two datasets were compared: a WorldView-2 scene with eight spectral bands, pansharpened with the panchromatic banda (0.5 m of spatial resolution); and an orthoimage with three spectral bands and 0.39 m of spatial resolution. Initially, the trees were in-situ geocoded. After this stage, some crowns were manually delimited in the image with the aid of the orthoimage, so serve as reference segments. These segments were then subdivided into two sets: training (70%) and test (30%), and drove the software Segmentation Parameter Tuner (SPT 3.9). In the optimization stage, the optimization function Nelder-Mead was employed to accomplish the automatic search of parameters for the Region Growing, MeanShift and Graph-Based algorithms. The Precision & Recall (P&R) metric was used to assess the segmentation quality. Preliminary results with the Region Growing (Multiresolution) algorithm applied on the orthoimage presented a P&R value of 0.13. The result showed that the crowns were well delineated, since the closer to zero the value of P&R, the greater the agreement between the reference and the segmentation result. This study is committed to advance the research in the delimitation of ITCs using VHR multispectral imagery, and hence, contribute to the object-based classification of tree species in subtropical forests.