[M-TT46] Introducing metaverse to agriculture. Are we ready?

[MTT46-P04] Creating decision trees for plant species identification in meadows using spectral reflectance

*Yurino Ishida¹, Yukihiro Takahashi¹, Nobuyasu Naruse², Garid Zorigoo¹, Takanori Emaru¹, Kuriki Murahashi¹, Tatsuharu Ono¹, Hokuto Nakata¹ (1.Hokkaido University, 2.Shiga University of Medical Science)

Keywords:Remote sensing, Agriculture, High ground resolution, Meadow

Due to the decline in the agricultural population caused by the aging of agricultural employees and lack of successors, efforts are rapidly being made to establish a labor-saving and efficient production system in large-scale operations and to realize smart agriculture using ICT and other advanced technologies. The same problem is becoming more serious in the dairy industry, and urgent action is required. In this context, while robotization and ICT are progressing in the management of dairy cows such as milking and feeding, meadow management is still a lack of technological innovation, and this automation is required. In order to achieve this, it is necessary to develop technology that can instantly grasp the condition of the meadow and perform appropriate treatment based on the condition. Therefore, in recent years, remote sensing as a technology to grasp the condition of grassland and AI and machine learning as technologies to determine appropriate treatment have been attracting attention. First of all, the development of the former technology needs to be expedited. Currently, about half of the meadows in Hokkaido are occupied by gramineous weeds. High occupancy of weeds is one of the factors that reduce raw milk production. The detection of weeds in meadows using remote sensing technology has been attracting attention in recent years. As for broadleaf weeds, measurement with high ground resolution makes it possible to identify weed locations using image recognition technology. However, most of the weeds look very similar to the grass and are very difficult to identify visually. It is necessary to estimate the occupancy rate of gramineous weeds by observation and to promote meadow renovation. Previous studies have used simple indices such as NDVI with red and near-infrared bands to identify grass plant species, but the accuracy of these indices is only about 30%. In recent years, there has been a lot of research on the automatic generation of classifiers such as neural networks and support vector machines for data classification. In this study, we attempted to classify plant species in a pasture using the decision tree method, which can determine the branching condition by itself, in order to classify using the if-then rule. Therefore, in this study, we attempt to classify plant species in meadow using the decision tree method, which is easy to set up branching conditions, because the classification course is easy to understand among classifiers and the if-then rule is used.
In this study, we developed a decision tree for identifying plant species with 4 narrow-bands. First, the spectral reflectance for 6 species of grass and weeds, which are common in meadows in Hokkaido, was measured by a hyperspectral camera at wavelength resolution of ~4nm in the range of 420-840nm. Next, A tree diagram with two conditional branches was developed. A decision tree for identifying plant species was made based on those 4 band measurements. 4 narrow bands were selected from the whole measured wavelengths. The differences in the visible light were emphasized by normalizing the reflectance as the first conditional branch. In addition, by using timothy as a standard sample and taking the difference from this, the differences between plant species were emphasized as the second conditional branch. From these emphasized differences, two conditional branches were performed and the plant species were successfully identified.
In conclusion, the created decision tree could identify plants in the meadow. In future research, the branching conditions of the decision tree will be determined by AI to create a decision tree that enables the identification of 6 species of grasses in Hokkaido meadows with higher accuracy. In addition, we will collect more data and try to develop a classification method using machine learning and AI such as support vector machines.