Replicating the Judgment of Experienced Farmers with AI: An Automatic Watering Control System

The stable production of agricultural crops is being threatened by climate change and extreme weather caused by global warming. With greenhouse horticulture, humans manually control the growing environment, enabling stable yields that are not impacted by soil conditions. While this allows for production of high-value-added crops, it also requires very finely-detailed environmental conditions compared to open-field cultivation, including ventilation, temperature, and humidity controls, resulting in an intense workload and requiring judgments based on experience. This creates extra work for farmers running mixed operations that combine open-field farming. Older farmers must struggle with concerns about physical stamina, while newcomers face issues with barriers to entry, leading to demands for labor-saving, efficiency, and standardization.

• 

  Working hours for vegetable growing in greenhouses are much longer than for other cultivation methods.
  (Ministry of Agriculture, Forestry and Fisheries, “Situation of Greenhouse Horticulture”)

Among the many varieties of vegetables, tomatoes are considered good candidates for greenhouse horticulture. They offer a huge market and excellent profitability, and their quality can be controlled through water and temperature management.

Kubota has turned its focus to high-sugar tomatoes, which have sugar content of 8 degrees Brix or greater. High-sugar tomatoes are a prime example of high-value-added crops for their strong profit potential. However, cultivating them requires advanced controls for both water and temperature. Because watering impacts the tomato’s sugar content, it must be especially precise and demands a greater level of experience and effort than conventional tomato farming.

One unique aspect of tomatoes is that they naturally accumulate sugar as a defensive response when exposed to moderate environmental stress. By limiting water to just the minimum amount, farmers can apply this stress, harnessing the power of nature to raise the fruit’s sugar content. Because carefully restricted watering is the key factor for success in high-sugar tomato production, the practice places considerable physical and mental strain on farmers, who must work to maintain this balance 365 days a year.

To reduce this strain, Kubota embarked on development of an automatic watering control system. Listening to farmers who work in the fields and integrating the latest technologies, Kubota developed an automatic watering control system based on wilt detection (Hamirus) by cultivating its own tomatoes.

Kubota made visits to tomato farms all across Japan to look into the watering process for high-sugar tomatoes. It found that when it came to timing, a crucial part of watering, judgments depended on visual inspection, including touching leaves to determine their softness. This requires expert intuition to detect subtle signs of wilting that an untrained eye cannot see. With this in mind, Kubota took on the challenge of developing a system that could make such judgments and provide automatic watering, applying its legacy of image-sensing technologies for in-factory inspection equipment.

• 

  Appearance of a plant before and after wilting

The issue with image sensing is that while it excels at identifying objects that do not change, its applicability to agricultural crops, whose forms and environments change constantly in short periods of time, remains untested. What is required is high-precision image sensing technology that can make sensitive judgments for watering at the level of experts. The development of such a system posed various challenges.

How can the expert intuition and experience used to judge leaf wilting based on color, shape, and softness and determine the optimum watering be quantified? Kubota installed a fixed-point camera to monitor tomato colonies. Realizing that healthy foliage visible in the images decreases when plants wilt, it quantified the coverage rate of plants in the image and used this to calculate the wilt rate. It then considered a mechanism that would automatically water plants when this wilt rate falls below a certain value.

• 

  The wilt rate is calculated based on the reduced area when comparing photographs taken of leaves at regular intervals.

• 

  Watering schedules are automatically controlled based on the wilt rate.

The most important factor for this system is to correctly detect the leaf regions in order to determine leaf wilting based on differences in leaf area. While cultivating tomatoes year-round at Kubota Farm Itoman, the company’s testing and demonstration farm in Okinawa, Kubota also visited tomato farms across the country to gather insights as it moved forward with development. In doing so, it identified two major challenges to address in order to accurately detect leaf regions.

The first challenge was differences in light conditions. Greenhouse horticulture cultivates crops inside greenhouses, which means light conditions undergo complex changes depending on the material covering the greenhouse and the time of day. The same green leaf can appear in different colors: the green of the leaf itself, the green under artificial lighting, and the green at the time the setting sun shines. Thus, it was necessary to take these changes into account when extracting leaf regions from images.

The second challenge was differences in plant posture^*. Leaf shapes vary depending on the tomato variety, and the overall crop appearance can change with growth and maintenance. Therefore, it was necessary for detection methods to take these changes into account.

Kubota aimed to address these challenges through the use of AI and development of algorithms.

• The shapes of the stems, branches, and leaves on a crop.

The first challenge, the accurate detection of leaf regions under various light conditions, is a critical element for the automatic watering control system that Kubota was seeking to develop. The company believed that building an AI trained on image data from various light conditions would enable a highly flexible detection system that could work under all kinds of light environments, so it set out to create a highly accurate AI applying deep learning techniques. Deep learning can learn features from large amounts of images, which enables highly accurate detection of even subtle changes in a target. This meant that the team had to collect as much image data as possible. To this end, Kubota’s nationwide farmer network built through its history of business played a key role. Thanks to the cooperation of many farmers, the team was able to quickly amass vast amounts of image data from tomato farms spanning from Aomori to Okinawa.

• 

  Various patterns of images were used to learn wilting conditions

All image data was captured from a uniform, top-down perspective. Because they were photographed from the top of the greenhouse directly above, the images contained only the tomatoes and the ground, simplifying AI recognition. Restricting the shooting pattern this way improved the efficiency of AI training data generation. Data scientists then verified that the collected dataset encompassed enough of a variety of patterns, allowing for the rapid development of a reliable AI model using only several thousand images. Although this is a relatively small dataset for deep learning, the successful results speak to the exceptional quality of the data.

Another issue was the differences in plant posture. During plant cultivation periods, farmers must do various kinds of caretaking. This includes vine lowering, in which tall stems are brought down to adjust plant height, and leaf picking, in which unnecessary leaves are removed to improve air circulation and efficiency of photosynthesis. This kind of caretaking significantly changes plant posture, resulting in periods where drastic changes in coverage rates are seen in images taken with the fixed cameras. If the system continues to calculate wilt rates under these conditions, it will find extremely low calculations that lead to unnecessary repeated waterings, which could greatly damage the plants.

• 

  Vine lowering, in which overgrown vines are shifted to make harvesting easier, can change plant posture.

For these reasons, Kubota developed an algorithm capable of adapting to major changes in plant posture. When a sudden drop in coverage rate is detected and the rate remains low for a certain period, the system determines that the change is a result of caretaking activities and adjusts its settings accordingly to prevent malfunction.

• 

  A graph showing automatic watering control adapting to plant posture changes

In addition, when developers visited fields, they noticed numerous potential false positives that they had not anticipated, such as raked leaves being left on paths or workers wearing clothes that matched the color of the crops. The team addressed each of these cases individually to achieve a system with even greater accuracy.

The automatic watering control system uses data collected by Kubota to set standard guideline values for wilt rate thresholds and watering volumes, determining how wilted leaves need to be for watering to begin. This allows even beginners to reliably produce high-value-added crops.

In addition, these settings can be flexibly changed to leverage expert judgments in addition to the automatic watering control system’s guidelines. The system has been designed to leave room for intervention, which means it can also reduce workloads for expert farmers without requiring them to drastically change the ways they have done things in the past.

A ten-month test cultivation was carried out with and without the automatic watering control system. It used cultivation under identical conditions, up to the day-to-day watering and caretaking. The results confirmed that implementing the system achieved 46% reduction in time spent managing watering and an 8% increase in yield. And as for sugar content, it showed that a high content of 8 degrees Brix could be maintained.

The goal of this project was to reduce the physical and mental strain on both beginner and expert farmers who have to watch over crops day after day, 365 days a year. By identifying needs that can only be discovered by working with farmers up-close, and by addressing those needs through its products and solutions, Kubota is contributing to farmers, the agricultural sector, and the stabilization of food production.

Greenhouse horticulture, unlike open-field farming, gives people precise control over environmental elements, such as light, water, humidity, temperature, and CO₂. It is also highly compatible with advanced technologies like ICT, making it possible to grow a diverse range of crops through careful control of those elements. However, realizing this potential requires the integration of multiple fields, from devices to software.

Kubota leverages its strengths as a manufacturer capable of producing and developing both hardware (agricultural machinery) and software (systems that control machinery) and integrates them for greenhouse horticulture, aiming to explore new realms in which it has never ventured before. For example, by using volumes of data accumulated online through cultivation support and linking it with robots that move autonomously within greenhouses to assist with farm work, Kubota will work to realize fully autonomous cultivation.