Page 407 - NGTU_paper_withoutVideo

Modern Geomatics Technologies and Applications

that it is used as an essential step for other PA applications. The importance of this application is caused by its role in many other
applications such as weed detection, mapping, and crop yield prediction [4, 12].
By allowing collecting information closely from above the crop canopy, UAVs have introduced a new perspective for
agricultural surveys, giving rise to several applications. As early as 2008, Nebiker et al.[13] proposed a multispectral sensor
prototype suitable to be mounted on a UAV and conducted experiments with it to assess vegetation health with promising results.
Starting from this experience, a variety of studies can be found in the literature about the effectiveness of UAV surveys conducted
for PAg purposes.
The main applications involve in-field weed mapping, vegetation growth monitoring, crop water stress analysis, and
irrigation management optimization [3, 14]. Different studies can be found in the literature, proposing (semi)automatic methods,
using image-processing techniques on single-band images, maps of Vegetation Indices (VIs), or Digital Elevation Models
(DEMs), to detect crop canopy[15]. Poblete-Echeverría et al. [16] compared the performance of four classification methods,
including standard and well-known methods (i.e., K-means and VIs’ thresholding) and machine learning methods (i.e., artificial
neural networks and random forest), to detect line canopy using ultra-high-resolution RGB imagery acquired with a conventional
camera mounted on a low-cost UAV.
Marques et al. [17] presented a UAV-based automatic method to detect chestnut trees using RGB and CIR (Color Infrared)
orthomosaics combined with the canopy height model. In [18], potato plant objects were extracted from bare soil using the excess
green index and Otsu thresholding methods. Different methodologies are tested and compared, and several segmentation methods,
such as supervised classifications, Bayesian segmentation, and thresholding algorithms developed for this purpose. Extraction
algorithms are applied both on geometric products (i.e., digital elevation model) and vegetation indices’ maps. As already
presented by other studies [16-18], the proposed methods exploit existing indicators, including NDVI and RGB-derived indices.
Giulia et al. [19] focus on crop row detection and extraction by analyzing and post-processing images acquired through a UAV.
Bah et al. [9] used crop row detection to detect the weed plants as an essential step for their work. First, the authors
performed a vegetation segmentation to differentiate vegetation objects from the background soil, where the output from this
process was a binary vegetation image. Then, their methodology used the Normalized Hough transform [20] to detect the linear
objects in the binary vegetation image, which are the crop rows. Such a procedure to detect the linear objects using Hough
transform in the vegetation binary images was highly adopted by other researcher [4, 21].
Following similar steps, José M Peña et al. (2015) used the concept of Object-Based Image Analysis (OBIA) to detect
the linear objects in the binary vegetation image. The authors used the OBIA to analyze the segmented region and classify the
linear object based on the dimensions of the segmented vegetation objects. Finally, the liner objects are segmented to create the
crop rows.
Moreover, different crop row detection methodologies tried to avoid the use of binary vegetation images. For example,
Comba et al. [22] used a combination of Hough and least square to perform a dynamic detection of crop rows to avoid
illumination effects. Also, other techniques were introduced to detect crop rows based on Fast Fourier Transform (FFT) [23, 24] ,
or based on the difference between the texture properties of crop pixels and the non-crop pixels [25]. Furthermore, other
techniques based on wavelets and multi-resolution analysis were proposed and showed 78% accuracy of crop row identification
[26].
Despite the enormous potential of UAV systems in various applications, including precision agriculture, there are still
limitations in this area, such as the high cost of multispectral imaging sensors. Most methods used in precision farming activities,
including plant line detection, use multispectral imaging sensors to perform the vegetation classification step with maximum
accuracy using the NDVI Vegetation Index. Such image sensors are more expensive than RGB image sensors. Although many
attempts have been made to perform the vegetation segmentation process using RGB images, there are still limitations to
achieving proper accuracy in different areas. Such a limitation led researchers to compare multispectral imaging sensors after
conducting various studies [13]. In this paper, a method for extracting plant planting lines is presented based on image processing
techniques and mathematical calculations.

2. Methodology
The proposed method of detecting crop lines is only through RGB images obtained by low-cost UAV imaging systems
and converted to orthophoto images by performing a preprocessing. As shown in Figure (1), it consists of two main phases,
which are described below. The implementation of the above algorithm is done in a Python programming environment.

Phases of the proposed algorithm:
 Phase 1: Identify the initial lines
 Phase 2: Refine and merge lines

2