Menu
STARS Project
close
STARS Project
close

Refine your search

STARS Project
Knowledge Portal

Automated farm field delineation and crop row detection from satellite images

Agriculture is vital to food security and economic growth, especially for developing countries. Accurate information on field boundaries and crop rows plays an important role in a variety of agricultural applications like precision farming, crop monitoring and yield forecasting and assists land administration systems considerably. Linear features, such as field boundaries and crop rows, could be extracted from remotely sensed data and are important data sources for geospatial information analysis. Edges in the remote sensing image describe the structural information of linear ground objects such as field boundaries and crop rows. These linear features are extracted from remote sensing imagery using the line segment detection (LSD) algorithm taking into account intensity differences between neighboring pixels. One of the most prominent advantages of remote sensing is the multispectral nature of the observations. However, the LSD algorithm takes only a single band as an input to detect linear segments. To make use of the advantage of the multispectral nature of the observations, this research combined all available information present in different bands as a single band before applying LSD algorithm. The capability of the LSD algorithm for extracting linear features was explored on two real applications in the agricultural sector:  field boundary delineation and crop row detection.

Field boundary delineation

For the purpose of field boundary delineation, multispectral WorldView-2 satellite imagery (8 band, 2m resolution, acquired on 29 July, 2014) from Mali was used. To concentrate all the available information as an input for the LSD algorithm, the gradient information of eight bands was combined into a single band by taking the vector sum of image gradients of all eight bands. When comparing the automatically extracted field boundaries with the reference boundaries (Figure 1), it becomes clear that both over- and underdetection are abundant. The ratio of missing detection (RM) and the ratio of false detections (RF) were 0.73 and 0.78 respectively. Using texture bands instead of spectral image bands did not improve the results.

        

Figure 1: Automatically extracted line segments overlaid on the reference dataset. Blue segments are reference segments, white segments are automatically extracted segments from the LSD algorithm. Bands 7, 5 and 3 of the WorldView-2 image displayed as red, green, blue, respectively.

Crop row detection

For crop row detection, panchromatic WorldView-2 satellite imagery (50 cm resolution, acquired on 12 September 2015) from Nigeria was used. In this study, in addition to detecting crop rows, dominant orientation (direction) of crop rows and spacing between crop rows were explored for different subsets of the image and promising results were obtained. Figure 2a shows an input image and Figures 2b and 2c show the automatically detected crop rows overlaid on the reference data set and the rose diagram of the dominant orientation of crop rows respectively. The figures show good agreement between detected- and reference crop rows and direction of rows. The RM and RF were  0.17 and 0.48 respectively. The high RF value is caused by detection of crop rows with two linear boundaries whereas the reference segment contains only single line.

  • Figure 2: Crop row detection. (a) An input image showing crop rows, (b) automatically extracted crop rows overlaid on reference dataset. Green color shows successful detections and red color shows missed detections, (c) rose diagram showing the dominant orientation of crop rows of the given input image. Figure 2: Crop row detection. (a) An input image showing crop rows, (b) automatically extracted crop rows overlaid on reference dataset. Green color shows successful detections and red color shows missed detections, (c) rose diagram showing the dominant orientation of crop rows of the given input image.


The methodology followed for detecting field boundaries from a single image in areas with a heterogeneous landscape was not as successful as expected. The different approaches followed: extracting information from multiple bands, using texture bands or image bands in detecting field boundaries do not give good results. On the other hand, other results of this research show that the methodology followed has a good potential in detecting crop rows, dominant orientation and spacing of crop rows automatically from satellite images. In future studies, in addition to image segmentation, exploring spectrotemporal and phenological information on crops could help to solve the issue of the problematic automatic detection of field boundaries.