In recent years, using Unmanned Aerial Vehicles (UAV) has enabled the development of many precision agriculture applications. Among these applications is the location of crop lines at low or medium altitude imagery, which allow us to estimate many attributes of a crop, such as a crop yield, number of plants, and failures in the sowing process. Since crop lines always show an almost constant appearance (greenish plants against reddish soil), this work proposes to evaluate the U-Net under different configurations of the numbers of filters and convolutional blocks. We also tested how different training sets affect its training. Results show that U-Net is a feasible approach to segment crop lines using fewer blocks and filters than traditional U-Net, being the first more important than the latter.