Humans still vital: The state of AI image analysis for crop pest and disease management

Technological advancements in pest and disease scouting are transforming a labor-intensive sector into a more efficient, data-driven version of itself. Now, as artificial intelligence (AI) is beginning to be developed to aid crop production, growers must be more critical than ever in assessing the benefits of these early-stage solutions. Dr Mikkel Grum, Research and Development Director at global crop pest and disease mapping experts, Scarab Solutions, discusses the latest AI developments and explains why farm and crop protection managers need to continue to focus on the technologies augmenting human labor instead of holding out for the as yet unfulfilled promise of AI.

The Food and Agriculture Organization of the United Nations estimates that between 20 to 40 percent of global crop production is lost annually to pests and diseases, costing the global economy around $220 billion. Pests such as thrips, aphids, leaf miners, mites, whiteflies and caterpillars, and diseases such as blights, mildews, botrytis and stem and root rots are common throughout all climate zones.

It is true that to become even more effective, crop management will require improved techniques as well as technologies. Many believe AI holds the answer.

AI takes its first steps in horticulture
Pest and disease monitoring is a labor-intensive process, requiring scouts to accurately evaluate plant and crop health as they move across the greenhouse, field or farm. AI-driven image analysis aims to help automate the crop surveillance process.

In horticulture, recent developments include a ‘robot scout’ equipped with near-infrared image cameras to detect powdery mildew and image analysis to predict bud and flower yields, and the IRIS Scout Robot. There is a remote pest monitoring system using machine learning (ML) to conduct image analysis on pheromone traps, and a large number proposing drone and satellite imagery as the basis of future crop management operations.

More wide-spread use cases promote the use of smartphone applications to scan photos for signs of pests and diseases, often presented as ready, or nearly ready for prime time use.

At first glance this sounds plausible. Many have heard that Google’s image analysis is now better than humans at recognizing cats and dogs in images, or that in breast cancer research, AI-image analysis now detects cancer on mammograms with more efficiency and accuracy than expert radiologists. So surely, using image analysis to identify crop pests and diseases on photos taken with a smartphone can’t be that far off. Not so fast.

Reality paints a less rosy picture
Current efforts to use image recognition technology in smartphones fall short of their promise to provide both a granular insight and actionable overview of farms and greenhouses.

As highlighted in a recent Scientific American article, the statistics used to present how well image analysis works are often misleading at best. The most common ‘pairing test’ that tests the ability to compare two images and state which of the two has the pest or disease, gives much higher percentages for accuracy than an analysis of multiple images with no knowledge of whether any of them have the pest or disease.

…and can create problems such as false positives.

Using inaccurate or skewed results gleaned from AI as a basis for pesticide use can cause more damage than good, as illustrated by the issue of false positives.

Let’s envisage an imaging system that gives a false positive of blight just five percent of the time, a very conservative figure even by the claims of accuracy of any current app. In a field full of blight, this wouldn't pose a problem, but now let’s take a field that doesn't have any occurrence of the disease. If you took 2,000 images in that field – which is the number of observation points a skilled scout manages in a day – you would get 100 positive results!

Does the grower act on this result, or do they inspect the 100 “positive” locations to check whether they really do have this problem? Multiply this by the other pests and diseases that the image analysis system is also checking for and perhaps has even higher false-positive rates for, and you have the workings of a practical nightmare. The higher the number of false positives, the more resources are required to conduct independent verification of results—meaning all the gains of automation are lost.

Machine versus human
The approach also needs to be put in context. Studies comparing situations where either there is AI or no crop scouting technology at all do not paint a realistic picture, because in some cases there is already a system in place that helps record and analyze data collected by human scouts.

In the breast cancer research case, as tumors are not visible to the human eye, the doctors and AI are looking at the same image. In a greenhouse setting, however, image analysis is much less effective than human attention to detail. A scout can move their head and turn over leaves to see the problem from multiple angles and with a magnifying glass, therefore has a significantly better view of the issue than a smartphone image would have.

Augmenting human skills with mobile technology – smartphones make people smarter
Simply put, farms and greenhouses still need people to walk around, open up the crop canopy and turn over leaves and use a magnifying glass where needed. To build on this requires technologies that enable scouts to do their jobs more accurately, faster and to a greater result—not technologies that ignore their expertise.

Smartphones are and will continue to be key in this process—but not primarily as an AI tool. A more realistic and proven use of mobile applications is for data collection and mapping purposes. Instead of using smartphones to take photos for AI to analyze, crop protection managers should empower scouts to use their inspection skills and record the results as they go.

Training plays an important role in this process, not the least in tightening scouting timelines—another pain point AI expects to solve. Correct identification and scoring of pests and diseases, thorough knowledge of sampling protocol and the technique to speed up the process are all required to harmonize the performance and accuracy of scouts across the entire farm, which is key for success.

AI may help guide scouts towards correctly identify pests or diseases they come across that they don’t know, but most crop scouting is about keeping track of the very dynamic distribution of a well-known set of pests and diseases.

Digital mapping and scouting technologies enable humans to glean new insights
If we couple the data recorded by scouts with geographical information, the results create datasets, providing a clear audit trail for traceability and data visualization options such as digital maps, charts and graphs—and valuable help to easily identify unique and recurring problems and patterns with few, if any, false positives.

Digital mapping is where pest and disease scouting technology meets human expertise to optimize outcomes. At Scarab Solutions, we see this every day as clients use Scarab Precision crop pest and disease scouting and mapping solutions to provide a solid basis to pinpoint infestation hotspots, determine the right pesticide use or biological control agent and reduce crop losses by enhanced farm management.

As their datasets grow, crop protection managers can not only measure their own progress, but in some cases benchmark against pest and disease figures in their region, using anonymized data from other farms.

Not yet AI’s time to shine, but we will always need the human touch
While AI-driven image analysis remains a talking point in the industry, the technology has a long way to go before it can produce reliable, accurate and actionable use cases. Today, GPS-tracking, mobile data collection and interpretation tools are the most effective and lucrative technology solutions for pest and disease management for crops.

As horticulture undergoes a technological transformation, artificial intelligence should not be seen as a substitute for existing processes, but as an extension of human intelligence. AI-driven image analysis will come with drones and robots in some settings, but that's a story for another time.

For more information:
Scarab Solutions 

Publication date:

Receive the daily newsletter in your email for free | Click here

Other news in this sector:

© 2020

Sign up for our daily Newsletter and stay up to date with all the latest news!

Subscribe I am already a subscriber