Here, intelligent and autonomous robots can enable ultraprecision agriculture, potentially changing the nature of the agrochemicals business. In this process, bulk commodity chemical suppliers will be transformed into speciality chemical companies, whilst many will have to reinvent themselves, learning to view data and artificial intelligence (AI) as a strategic part of their overall crop protection offerings.
In this article key points that are covered in depth in the report “Agricultural Robots and Drones 2017-2027: Technologies, Markets, Players” (www.idtechex.com/agri).
Computer vision
Computer vision is already commercially used in agriculture. In one use case, simple row-following algorithms are employed, enabling a tractor-pulled implement to automatically adjust its position. This relieves the pressure on the driver to maintain an ultra-accurate driving path when weeding to avoid inadvertent damage to the crops.
The computer vision technology is however already evolving past this primitive stage. Now, implements are being equipped with full computer systems, enabling them to image small areas, to detect the presence of plants, and to distinguish between crop and weed. The system can then instruct the implement to take a site-specific precision action to, for example, eliminate the weed. In the future, the system has the potential to recognize different crop and weed types, enabling it to take further targeted precision action.
This technology is already commercial, although at a small scale and only for specific crops. The implements are still very much custom built, assembled and ruggedized for agriculture by the start-ups themselves. This situation will continue until the market is proven, forcing the developers to be both hardware and software specialists. Furthermore, the implements are not yet fully reliable and easy to operate, and the upfront machine costs are high, leading the developers to favour a robotic-as-a-service business model.
Nonetheless, the direction of travel is clear: data will increasingly take on a more prominent (strategic) role in agriculture. This is because the latest image processing techniques, based on deep learning, feed on large datasets to train themselves. Indeed, a time-consuming challenge in applying deep learning techniques to agriculture is in assembling large-scale sets of tagged data as training fodder for the algorithms. The industry needs its equivalents of image databases used for facial recognition and developed with the help of internet images and crowd-sourced manual labelling.
In not too distant a future, a series of image processing algorithms will emerge, each focused on some set of crop or weed type. In time, these capabilities will inevitably expand, allowing the algorithms to become applicable to a wider set of circumstances. In parallel, and in tandem with more accumulated data (not just images but other indicators such NDVA too), algorithms will offer more insight into the status of different plants, laying the foundation of ultra-precision farming on an individual plant basis.
Agriculture is a challenging environment for image processing. Seasons, light, and soil conditions change, whilst the plant themselves transform shape as they progress through their different stages of growth. Nonetheless, the accuracy threshold that the algorithms in agriculture must meet are lower than those found in many other applications such as autonomous general driving. This is because an erroneous recognition will, at worse, result in elimination of a few healthy crops, and not in fatalities. This, of course, matters economically but is a not safety critical issue and is thus not a showstopper.
This lower threshold is important because achieving higher levels of accuracy becomes increasingly challenging. This is because after an initial substantial gain in accuracy improvement the algorithms enter the diminishing returns phase where lots more data will be needed for small accuracy gains. Consequently, algorithms can be commercially rolled out in agriculture far sooner, and based on orders of magnitude lower data sizes and with less accuracy, than in many other applications.
Navigational autonomy
Agriculture is already a leading adapter of autonomous mobility technology. Here, the autosteer and autoguide technology, based on outdoor RTK GPS localization, are already well-established. The technology is however already moving towards full level-5 autonomy. The initial versions are likely to retain the cab, enabling the farmer/driver to stay in charge, ready to intervene, during critical tasks such as harvesting. Unmanned cable versions will also emerge when technology reliability is proven and when users begin to define staying in charge as remote fleet supervision.
The evolution towards full unmanned autonomy has major implications. As we have discussed in previous articles, it may give rise to fleets of small, slow, lightweight agricultural robots (agrobots). These fleets today have limited autonomous navigational capability and suffer from limited productivity, both in individual and fleet forms. This will however ultimately change as designs/components become standardized and as the cost of autonomous mobility hardware inevitably goes down a steep learning curve.
Agrobots of the future
Now we can see the silhouette of the agrobots of the future: small intelligent autonomous mobile robots taking precise action on an individual plant basis. These robots can be connected to the cloud to share learning and data, and to receive updates en mass. These robots can be modular, enabling the introduction of different sensor/actuator units as required. These robots will never be individually as productive as today’s powerful farm vehicles, but can be in fleet form if hardware costs are lowered and the fleet size-to-supervisor ratio is increased.
Now we can also see what this may mean for the agrochemicals business. First, data and AI will become an indispensable part of the general field of crop protection, of which agrochemical supply will become only a subset, albeit still a major one. This will mandate a major rethinking of the chemical companies’ business model and skillsets. Second, non-selective blockbuster agrochemicals (together with engineered herbicide resistant seeds) may lose their total dominance. This is because the robots will apply a custom action for each plant, potentially requiring many specialized selective chemicals.
These will not happen overnight. The current approach is highly productive, particularly over large areas, and off-patent generic chemicals will further drive costs down. The robots are low-lying today, constricting them to short crops. Achieving precision spraying using high boys will be a mechanical and control engineering challenge. But these changes will come, diffusing into general use step by step and plant by plant. True, this is a long term game, but playing it cannot be kicked into the long grass for long.