Inter-row cultivation is the cultivation of soil between rows of crop plants and band-spraying. In either form we need to locate the rows and the spacing between them.
Naïve methods to differentiate the background (soil) and the crop plants or those who attempt to model the rows themselves will fail if they would not take into account the various ways weeds can affect their processing.
Inter-row recognition is all about finding the pattern of the rows and just as weeds can break a row’s pattern or create a pseudo-row of weeds on some occasions large patches of missing crop plants could cause confusion in a similar fashion by breaking a row’s integrity.
To battle these distracting elements it is recommended to try and capture as many rows as possible in the image thus making sure our pattern recognition relies on a robust sampling of the field. If necessary this could be achieved by connecting several cameras together, though this will require careful analysis of the geometry of the scene between the cameras.
Once you have a suitable image there are several approaches that can be used – some dependent on the crop plant in question or its size while other are more generic.
A useful generic approach will be to consider the random distribution a weed will no doubt take and look for the repeating frequency of the crop plant’s rows.
Another will be to build a dynamic model of a row’s contour and look for the deviations which will indicate the presence of weed.
This second approach also gives us a simplistic approach to the problem of accurate band-spraying and other situations where we want to treat the row itself from a dynamic distance. In these case we are not interested in the individual plant (unlike intra-row cultivation ) but in the row itself, the location of the plants in it or the local changes in its thickness. To these end having an even simplistic model of the row can give us tremendous control and the ability to accurately spray where needed.
Whichever the approach that is taken it is important to remember that abnormalities in the initial input are pervasive and that updating your solution throughout the operations will oftentimes be critical to your success.