April 14, 2022
NEDO (New Energy and Industrial Technology Development Organization)
Farmship Co., Ltd.
NEDO is working on “Realizing an Intelligent Society by Applying Artificial Intelligence Technology”, and within this framework, Farmship Co., Ltd. uses artificial intelligence (AI) to estimate the weight of lettuce in a non-contact, non-contact environment. Destructively, I develop an algorithm. This time, we conducted a verification test of this algorithm in a real factory and confirmed high estimation accuracy against the measured value.
This algorithm analyzes the lettuce image captured to estimate the weight. For this reason, it is possible to effectively measure the weight even during cultivation, and it is possible to improve the productivity of the plant factory by early detection of growth abnormalities, increase in yield by appropriate selection and accurate yield prediction.
NEDO and Farmship Co., Ltd. verify yield changes due to improved sorting accuracy and continue to accumulate crop data. In addition, we aim to achieve a very precise adjustment of supply and demand by combining it with the demand forecasting technology and the growth monitoring technology being developed in this project.
Compared with open-field cultivation, plant factories are not affected by weather conditions and can produce stably in narrow cultivated land, so vegetable production has greatly increased in recent years. However, it is difficult to make the growth environment completely uniform even in a plant factory, and the growth rate varies from individual to individual, so an effective understanding and management of the growth situation in a plant factory is a problem.
Figure 1 is a graph of the weight distribution of lettuce actually grown in a plant production facility, showing that the lettuce in the red circle weighs only about half the weight of the average. Artificial intelligence (AI) can detect and select these slow-growing individuals at an early stage, or move them to a suitable environment to suppress growth variability, making it more stable and efficient.Production is possible . In addition, it is possible to detect significant deviations from the average distribution very early on, and it is possible to prevent production defects.
In this context, NEDO (New Energy and Industrial Technology Development Organization) is working on “Realizing an Intelligent Society by Applying Artificial Intelligence Technology”.※1In the sector, Farmship Co., Ltd. has developed a non-contact / non-destructive weight estimation algorithm for vegetables using AI. Accordingly, we conducted a verification test to estimate the weight of lettuce produced in a real factory using the developed algorithm, and confirmed that the estimation accuracy is high compared to the measured value. .
2. The achievements of this era
The algorithm developed in this project is a convolutional neural network (CNN).※2Object detection using※3And the weight density※4Is combined with a regression analysis method that allows CNN to predict weight density by training CNN. In this verification test, the rectangular (rectangular) area of each lettuce was accurately extracted from the image in which 20 lettuces partially overlapped, and the correlation coefficient was 0.76 (the exact match is 1, 0 if there is no correlation). succeeded in estimating the weight of individual high-stacked lettuces.
Normally, to learn a CNN image, you need to convert it to a square. However, at that time there was a problem of loss of information about the original size of lettuce. Therefore, in this method, the weight density is estimated, and finally the lettuce weight is calculated by multiplying by the rectangular area, so that the rectangle size information can be used to make a more accurate prediction. .. Since several lettuces can be measured simultaneously and without contact and non-destructively, it is possible to effectively estimate the weight of each individual even during cultivation.
3. 3.Future plans
NEDO and Farmship Co., Ltd. use this algorithm to photograph lettuce during replanting work in a plant factory, estimate the growth situation, and aim to realize a system that can grow with the same weight for each individual. In the future, in addition to verifying yield changes due to improved sorting accuracy, we will continue to accumulate crop data (Fig. 3). of development within the framework of this project※5And by combining with growth monitoring technology, we aim to achieve a very precise adjustment of supply and demand.
Farmship Co., Ltd. plans to continue studies for the practical application of this system from 2023. Thanks to the practical application of this system, it will be possible in the future to ship standard vegetables without waste. At the same time, the production cost is lower than that of open-field cultivation, so consumers can get high-quality vegetables produced in the factory from the factory at a lower price. In addition, efficient production can reduce energy waste and contribute to energy savings.
- *1 Realization of an intelligent society by applying artificial intelligence technology
- Trade name: Realization of intelligent society by applying artificial intelligence technology / Research and development of value chain efficiency improvement system such as AI factory
Project period: 2018-2022
Contractor: Farmship Co., Ltd., National University Corporation The University of Tokyo
Subcontractor: Pai Material Design Co., Ltd., National University Corporation Toyohashi University of Technology
Summary of activity: Realization of an intelligent society by applying artificial intelligence technology
Intro Video: NEDO Channel “Achieving an Intelligent Society by Applying Artificial Intelligence Technology”
Intro Video: NEDO Channel “Research and development of value chain efficiency improvement system for factory plants by AI”
- * 2 Convolutional Neural Network (CNN)
- It is a type of neural network that is particularly often used to recognize two-dimensional data such as images. A convolution layer that extracts features by applying discrete convolution with the kernel to the input activity of multiple two-dimensional planar data channels, a clustering layer that reduces input size and absorbs feature misalignment, and a convolution layer. fully connected layers that classify the features extracted by the grouping layer.
- ※3 Object detection
- A technique using deep learning is being actively developed as a technique to detect the existence of a specific object in a moving image and infer the position and distance of each object if it exists. When an object exists, the common method is to estimate the rectangle that surrounds the object.
- ※4 Weight density
- This is the value obtained by dividing the actual weight of lettuce designed in this development by the area of the rectangle (rectangle) extracted from the image.
- *5 Prediction technology is required
- Reference: press release of November 19, 2019 “Development of an algorithm to predict the market price of vegetables using AI”
Four. Contact Information
(Contact for inquiries regarding the content of this press release)
NEDO Robot / AI Department Manager: Kato, Terashita TEL: 044-520-5241
Farmship Co., Ltd. Responsible: Kondo TEL: 03-5829-9601
(Other general inquiries about NEDO activities)
NEDO Public Relations Department Responsible: Suzuki, Sakamoto, Hashimoto, Nemoto
- Please use “NEDO (New Energy and Industrial Technology Development Organization)” or “NEDO” when introducing our organization’s name in newspapers, television, etc.