Guidance technology is the technology that allows us to automatically control a moving element in a context-full environment. In the field of agriculture it could be used to steer heavy machinery through a densely cultivated field in a precise manner that minimizes the possibility to damage the plants, to spread seeds in the precise and efficient manner or even controlling the weed amounts in the field.
In general we can speak of two main fields of cultivations:
- inter-row cultivations where the devices move between rows of crop plants, usually with the intent of cultivating the soil between the rows (for example by weeding it)
- Intra-row cultivations where the devices cultivate the rows of crop plants either directly interacting with the plants or their surroundings – soil and weed.
In either case the first step in the guidance algorithm is to find the rows of crop plants and separate them from the background. This task’s difficulty is usually based on the crop plant itself as the spacing of the rows and the plant’s features used to play a significant role, but with modern computer vision tools it is possible to build a general system that can achieve this task in a completely automated and successful fashion.
After locating the rows and inter-row spaces we align the machine and camera accordingly.
If our aim in achieving guidance technology in precision agriculture is inter-row cultivation, then the guiding part is usually over at this point – we keep the machine in the line and focus on cultivating the inter-row soil. Note that a major exception is when we are doing accurate band-spraying and have to locate the main body of the crop plants in order to estimate the desired distance for the spraying.
Things are much more sophisticated when intra-row cultivation is required and we need to recognize individual plants. Whilst building a species-specific model is an option – it is a bad one. The goal should be to find the plants in an automated and generic manner as possible and to that end there is a vast array of computer vision techniques and tools that can be used – from contour finding to deep learning – but in essence the idea is to use the relative order of the crop plants in contrast to the random distribution of weeds.
For that end we would ideally want to see as many rows as possible in each frame to better capture the pattern of the plants and in some scenarios this might require using more than one camera in parallel.