Gaussian process based c-regression models (GPCRM) is a method to obtain the cluster partition and nonlinear regression models simultaneously. Since the regression model depends on the kernel parameters, it is difficult to obtain a regression model that captures the data structure depending on the kernel parameters. In this paper, we propose maximum marginal likelihood GPCRM(MML-GPCRM) as a method introducing kernel parameter estimation. The experimental results suggest that the MML-GPCRM estimated a better-fitting regression models than the GPCRM.

Deep Learning has expanded the use of AI, particularly in healthcare, where it is used for primary screening to exclude normal samples. However, conventional classification models lack confidence in their judgments, leading to potential misclassifications. Therefore, some of the authors enabled mapping of image ambiguity to specific positions by introducing new parameters into the loss function of Siamese Networks. And they proposed a classification model inspired by radar charts. However, the effectiveness of this approach has not been extensively discussed so far. So, we aim to improve this model by determining endpoints based on class data distribution, ensuring accurate and error-free classification.

Clustering is an unsupervised learning and can be broadly classified into hierarchical and non-hierarchical methods. Among non-hierarchical methods, those based on objective function optimization have been studied extensively from the viewpoint of not only effectiveness but also theoretical depth. Fuzzified Even-sized Clustering Based on Optimization (FECBO) is a method that focuses on clustering with the constraint that the cluster sizes are equal in terms of the fuzzy membership. However, FECBO has the problem of being susceptible to noise. To solve this problem, a method was proposed that introduces the idea of noise clustering into FECBO, called Fuzzy Even-sized Noise Clustering (FENC). Nevertheless, FENC consists of a noise clustering phase and a FECBO phase, each of which is run separately, which makes the algorithm more complex. Also, from the standpoint of run time, better results may be obtained if these two phases are performed at the same time. In this paper, we propose a simpler algorithm that combines these two phases. Furthermore, we attempt to introduce a kernel function to perform classification by nonlinear boundaries. We then compare the proposed method with conventional methods through numerical examples.

Clustering is one of methods of unsupervised learning. However, semi-supervised clustering, labeling a part of data to obtain desired result, is actively researched. On semi-supervised clustering, labeled data set is called semi-supervised dataset. a representive labeling is pairwise constraints which apply constarints to pair of instances. Above all, multiple researches about must-link constraints which demand pair of instances to be in same cluster and cannot-link constraints which demand pair of instances to be in differnt clusters are reported. incidentally, one of us proposed fuzzified even-sizedclustering based on optimization (FECBO) as one of clustering algorithms which classfies data set to same size clusters. It has been applied to delivery planning. In this paper, we propose algorithms that add pairwise constraints to FECBO and invesgate its effectiveness through numerical example.

The transportation problem is the problem of determining the amount of transportation from a supply location to a demand location so as to minimise the total transportation cost, given a given number of supply and demand locations, the amount of transportation from each supply location to the demand location, and the transportation cost according to the distance from the supply location to the demand location. On the other hand, clustering methods for unsupervised division of given data into a given number of clusters can be broadly classified into hierarchical and non-hierarchical. The latter are referred to as ’objective-based clustering’ in this paper, as they generally search for the optimal solution of a set objective function under constraint conditions and obtain it as the clustering result. In the previous paper, several objective-based clustering methods from the perspective of transportation problems were discussed and the correlation between them was examined. Furthermore, based on these considerations, an index for evaluating clustering results was proposed. Based on the previous paper, this paper first discusses the evaluation functions for the clustering results proposed in the previous section through numerical examples. Next, a new objective-based clustering algorithm from the perspective of transportation problems is presented.

It is difficult to understand the structure of the high-dimensional data. Time-series data is an example of such high-dimensional data. Time-series data requires consideration of differences in period and number of sequences. The choice of dissimilarity and the clustering algorithm also affect the generated cluster structure. In particular, it is difficult to understand the cluster structure of time-series data because of its high dimensionality. The cluster validity measures are useful for evaluating the cluster partition of time-series data. In this study, we propose a cluster validity measure based on fuzzy membership for time-series data. It is shown that the proposed method performs as well as or better than existing methods through numerical experiments. It is also shown that the proposed method is useful for data that is close to the cluster center to which each data belongs and far from other cluster centers.

Clustering methods such as k-means and fuzzy c-means methods require the number of clusters to be determined in advance. In addition, conventional clustering methods may not be able to properly classify unbalanced data or data containing outliers. In this study, we propose a sequential cluster extraction method based on controlled-sized sequential possibilisitic clustering for imbalanced data and data with outliers. Furthermore, numerical experiments were conducted to confirm the effectiveness of the proposed method.

In Waste-to-Energy plants, combustion is dependent on many complex factors and a sign of unstable states appears in other locations. We have proposed a fuzzy relational maps of many sensors (FuRMS) for expressing relationships between many sensors as relations between fuzzy sets of sensor values. We use data sets of the same combustion state for constructing a fuzzy relational map. We calculate strengths of relationships between fuzzy sets of sensors based on the distribution of sensor values, considering relationships between sensors as fuzzy rules. The calculation results for all combinations of sensors are called a fuzzy relational map. For evaluating unknown data with the maps, we calculate the weighted sum of degrees of consistency with the relational map. The many sensor pairs with weak relationships get greater weights in an evaluation value. In this paper we propose a method to exclude such data, assumed to be a uniform distribution, from the data set for constructing maps. It makes the evaluation values greater with strong relationships. We construct fuzzy relational maps using data of a certain plant with two methods, one this method and the other the previous method, and compare evaluation values for unknown data.

In this study, we introduce Virtual Reality to practice of painting and sealing material on a car body. We propose to implement this VR application by applying Unity and its assets. As a result, the painting and sealant processes is successfully reproduced as virtual reality. As a result, we successfully reproduce the painting and sealing material processes on a car body as virtual reality.

In recent years, researchers have developed multi-agent reinforcement learning systems to automate luggage transfer. However, these systems often struggle to learn effectively in partially observable environments, such as POMDPs. This paper presents a novel learning approach that leverages relative vectors to address this limitation. The proposed method is compared to conventional approaches, and the results show that it can achieve better performance in POMDPs.

In this study, English sentences of the “The National Examination for certified General Travel Service Supervisor,” a qualification test for being required to sell travel products at a travel agency are examined in terms of metrical linguistics. In short, frequency characteristics of character- and word-appearance are investigated using a program written in C++. These characteristics are approximated by an exponential function. Furthermore, the percentage of Japanese junior high school required vocabulary and American basic vocabulary is calculated to obtain the difficulty-level of each material.

Robot vacuum cleaners, which automatically clean the floor, are becoming popular these days. Products with various functions are on sale from many manufacturers, including iRobot’s “Rumba” that is the forerunner of popularity. In this research, what kind of evaluation is done about these home robot vacuum cleaners using the word of mouth of people who actually used them are investigated.

We conducted an experiment in which personality traits of speakers are inferred using a chat dialogue corpus between human and a chatbot. We analyzed the relationship between the personality impressions obtained in the experiment and various utterance features including the frequency of occurrence of certain categories of vocabulary and the rate of occurrence of dialogue acts. In order to realize a dialogue system with personality, we constructed a Transformer model for response generation in chat dialogues using this corpus as training data, and conducted experiments on the reproducibility of personality.

Recently, a large number of methods have been proposed for detecting sarcasm in Twitter tweets. In English-speaking countries, it is common to use explicit hashtags when tweeting sarcasm, which can be leveraged to collect training data. However, sarcasm-indicative hashtags are not widely used in Japanese tweets, and thus it is difficult to collect a sufficient amount of Japanese sarcastic tweets in this approach.Therefore, in this study, we aim to enhance the classifier performance by increasing the amount of training data through semi-supervised learning. Specifically, based on the assumption that tweets containing the word ’皮肉’ tend to express sarcasm, we collect tweets that meet this criterion. We then extract highly reliable sarcastic sentences by NU learning with the collected data as unlabeled, and train a classifier by adding them to training data. In an evaluation experiment of the classifier, we compared two methods: one utilizing only PU learning and the other combining PU learning with NU learning. The results demonstrated a 3.4% improvement in accuracy rate for the proposed method that integrated PU learning with NU learning, as compared to the method solely using PU learning.

”Business katakana words” such as ”internship” and ”sustainable” have been used rapidly in recent years. However, it has not been investigated why the tendency to use ”business katakana words” has increased in spite of the existence of Japanese expressions with the same meaning. In this study, we analyze a large newspaper database and examine the trend of ”business katakana words” as a time series, as well as the possibility that the occurrence of social events affected the increase in ”business katakana words.” We also attempt to verify the possibility that the occurrence of social events affected the increase of ”business katakana words” by analyzing the co-occurrence network.

Common ground refers to the knowledge, beliefs, and suppositions that the dialogue participants believe they share about the activity. The construction of common ground among the participants is essential for the dialogue to proceed smoothly. In dialogues where science museum staff provide explanations to visitors, the amount of knowledge about the objects to be explained differs greatly between the science museum staff and visitors, making the construction of common ground even more important. In this study, we analyze the speech strategy of science museum staff to build common ground with visitors in the context of exhibit explanations.

This article is a partial reorganization of a master's thesis submitted by the first author (Minagawa) to Tokyo Gakugei University in 1978. There have been few studies on memory retrieval processes and why they occur. Minagawa took a hint from the kinetic theory of gas molecules in physics and the idea of chemical equilibrium in chemistry, corrected the shortcomings of the original equation, and succeeded in reconstructing an equation that is partially useful for "quantitative prediction" of the retrieved quantity. The formula was finally obtained by a flash of inspiration.

In this research, we propose a method of detecting scene change in a narrative in order to realize extractive summarization. First, preprocessing removing dialogue sentences from narrative sentences is performed. Next, candidate sentences for scene change are enumerated by calculating the power of attractiveness of a word, which means the dependency from other words, and the conjunctive expressions. At this time, the characters are automatically extracted based on the subject of a specific verb, and are used as words to be focused on when calculating the power of attractiveness of words. After that, the referential expressions are used to delete sentences that are assumed not to be scene changes from the candidate sentences. As an evaluation of the proposed method, the accuracy of the scene change inference was obtained by manually creating the correct data and comparing it with the estimation result by the proposed method. As a result of the evaluation, the precision was lower than the recall. It was confirmed that short sentences were difficult to extract due to the influence of sentence length when calculating the power of attractiveness of a word. Furthermore, the connective expressions used in scene change sentences may be biased.

Many local governments in Japan publish their council meeting minutes in an electronic format. In a previous study, the authors showed that XAI technology can be applied to a BERT-based classifier for estimating the speaker of an utterance in the minutes to visualize characteristic expressions that are specific to a region or that show the speaker’s political interests. In this study, we describe the implementation of characteristic phrase visualization using XAI in a prefectural assembly meeting minutes search system developed and published by the authors. We also discuss use cases in which residents utilize the characteristic information of regions and speakers obtained as a result of the visualization.

In recent years, online services for gathering information and facilitating interaction related to pregnancy, childbirth, and child rearing have become widespread among women. The search behaviors and posts of users utilizing these services often reflect the concerns of mothers. This paper attempts to analyze the search terms and post contents on mother-targeted websites, with the aim of elucidating the relationship between a mother’s concerns and sentiments. Furthermore, by comparing data from fiscal years 2019 and 2020, we also investigate the impact of the COVID-19 pandemic on mothers.

In the Mini4WD-AI, a non-powerful microcontroller executes the processes of sensor measurement, self-position estimation, and speed control during high-speed laps. Among these, self-position estimation is quite computationally intensive and memory-intensive, making it very difficult to obtain the estimation results online. Standard inertial sensor values contain a large amount of noise. When using them as material for self-position estimation, appropriate design is required for similarity tolerance and filter processing. Since the increase in computational complexity makes it difficult to obtain results online, only sensor threshold processing and short-time running history information can be used for speed control, which requires a high response. Therefore, we propose a one-dimensional motion and inertial information that does not depend on running speed. We also propose a simple map representation and a computationally inexpensive position estimation method. We verify the effectiveness of the method on actual vehicles using a variety of driving data.

The elements necessary for intelligent behavior of mini 4WD are position estimation, machine learning (parameter optimization), and map generation. I introduce the examples implementation of these three elements. Finally, I show that mini4wd ai learned from test running, and that the learning results enable the mini 4WD to clear the difficult course layout.

This paper describes an approach to position estimation using a geomagnetic sensor to control of mini 4WD AI automatically. In our previous work, position estimation was performed using an external camera, but the error between the estimated and actual positions due to communication delays was significant. We attempt to solve the problem of the communication delay by eliminating external communication using the geomagnetic sensor mounted on the mini 4WD AI. Using the magnetism-sensing characteristics of the geomagnetic sensor, we examined the feasibility of position estimation by placing a magnet on the course.

In competitive team sports, individual player decisions significantly impact overall team performance. Designing an appropriate evaluation function for scoring player behavior in complex games like soccer is challenging. Incorporating the team supervisor’s guidance into decision-making processes is essential, but accurately scoring numerous trials poses difficulties. This paper employs a learning to rank method to obtain an evaluation function for action selection, focusing on ball-chasing behavior in soccer. The RoboCup soccer simulator is used for experiments, deriving a ranking model from players’ action logs. Gradient boosting trees are applied, and CatBoost is employed as the implementation of the learning-to-rank model. The study demonstrates that a model with satisfactory performance can be learned when the number of situations exceeds approximately 1,000, even with training data generated by humans.

Inclusion-Exclusion integral neural network is a network model that uses fuzzy measures, which are non-additive measures, and Inclusion-Exclusion integral, which are represented by polynomial operations. The model outputs of this network can be expressed in terms of Inclusion-Exclusion integral formula using the Möbius inversion formula, and the contribution of each feature to the prediction can be calculated using the Shapley values, which is one of the reward calculation methods used in the theory of cooperative games. This allows us to construct a network with interpretability for parameters after training. For this network, we interpret the model using the Shapley values and the Owen values, which is an extended version of the Shapley values when considering a particular coalition structure or division of the set of players. We will also examine model interpretation using SHAP values, where Shapley values are applied to machine learning to allow interpretation of the impact of features on prediction based on input-output relationships.

The Inclusion-Exclusion integral neural network is a neural network model that uses fuzzy measures, which are non-additive measures, and Inclusion-Exclusion integrals. This makes it possible to measure the contribution of each explanatory variable using statistical indices such as Shapley values and has the advantage of preserving interpretability of the parameters after learning. However, to maintain interpretability, the monotonicity of fuzzy measures should be satisfied. In this study, we propose a regularization term that satisfies the monotonicity of the fuzzy measure for the parameters obtained by training an Inclusion-Exclusion integral neural network, and conduct experiments using regression data. Furthermore, we will verify under what conditions learning can be performed with high accuracy while preserving the monotonicity of the fuzzy measure, compared to the l1 regularization and l2 regularization that are currently widely used.

Inclusion-exclusion integral neural network is an explainable network model using inclusion-exclusion integral defined by fuzzy measures and polynomial operations. Inclusion-exclusion integral neural networks increase the expressiveness of the network by specifying a large additivity, but the number of parameters increases exponentially, and the training time increases accordingly. This is one of the challenges. The objective of this research is to accelerate the learning of inclusion-exclusion integral neural networks by comparing the speed between CPU and GPU in terms of additivity and the number of data, and by applying acceleration methods in PyTorch.

The Traveling Salesman Problem (TSP) is a problem of finding the shortest route back to the starting point after visiting each city only once. It is called an NP-complete problem, and many computational methods have been proposed. Referring previous Self-Organizing Maps (SOM) research, we have studied how to shorten the computation time for applying SOM to the TSP while keeping the original form. In this study, we applied and examined the problem of traveling cities on the earth and the optimization of component placement in a chip mounter, which is practical application. The results will be shown at the conference.

The principle of significance was explained using iris data. This method was applied to animal data. Furthermore, the normal value of uric acid level was tested by applying it to medical data.

Growth of the Someiyoshino cherry trees has peaked from 40 to 60 years after planting and the trees tend to decline rapidly later and curing such as fertilization at this peak has a remarkable delayed effect of decline of the trees. It is the most important task to grasp the health condition of the trees. We collected the data of cherry trees and old trees of cherry blossoms in parks, schools, riverbanks, etc. in the west part of Tottori Prefecture, and investigated distribution status and characteristics by using the spherical SOM.

In this study, we propose a method for determining teaching content utilizing rough set theory and formal concept analysis. It is important to connect the concept of multivariate functions and the concept of derivatives, and to have students understand the concept of total derivatives in order to make them understand both. Decision rules are generated to analyze the relationship between student characteristics and comprehensible instructional content of total derivatives. In addition, we develop a prototype VR system for teaching total derivatives to validate the method. It helps students understand the concept of total derivatives graphically.

In this paper, we discuss an improved method to quickly calculate relative reducts after appending some objects to the decision table. In general, when a decision table that has already created relative reducts is updated, it is necessary to calculate all relative reducts again. However, recalculation of all relative reducts requires a lot of time and storage area. Therefore, we propose a new method to calculate relative reducts quickly and easily after appending some objects to the decision table.

Interrelationship mining is a rough set-based data analysis framework using interrelated attributes, which represents characteristics based on comparison between values of two attributes. However, in rough set-based data analysis, the number of attributes affects computational loads exponentially, and so we cannot use all interrelated attributes at the same time. In this paper, we show a necessary and sufficient condition that an interrelated attribute is included in relative reducts.

Based on rough set theory, various data analysis methods have been studied. On the other hand, similar researches are also being conducted in the fields of machine learning and statistics. It is believed that by incorporating the findings of these fields into the rough set approach, it would be possible to develop new data analysis methods and integrate theories. In this study, for the variable precision rough set model, we propose rough membership values based on Bayesian learning. We propose a probability model that reflects the concept of rough set theory.

User interface (UI) is a set of display screen and operation method for exchanging information between a computer system and users. UI in VR devices is an important factor for users to play games comfortably. In this presentation, we focus on the fact that when the user holds the controller, the index and middle fingers are naturally placed on the controller and the triggers can be operated with a motion similar to holding the controller. By having subjects experience operations using triggers on various types of controllers, we will investigate the feel of operating buttons and triggers, and consider interfaces suitable for each type of operation.

In the Analytic Hierarchy Process (AHP), the criteria importance is given by a Pairwise Comparison Matrix (PCM), whose elements represent the relative importance of items corresponding to the row and column. In collective decision problems, the aggregation of multiple PCMs representing opinions from various viewpoints of members in different positions is important for obtaining outcomes desirable for a big majority. The components of PCMs given by the experts on the problem can be intervals of acceptable values reflecting their wide perspectives. In this paper, we investigate the methods for the aggregation of two opinions expressed by PCMs with interval components given by two experts. We assume that those interval PCMs are similar as both experts give them in consideration of various members in the collective. Moreover, as the given interval PCMs are not always consistent themselves, some method for ensuring consistency, i.e., consistentization should be applied for the given interval PCMs together with the aggregation. Under the problem setting described above, four consistently aggregated methods are formulated and examined. They are different in the application stages of aggregation and consistentization.

In this paper, we report our attempt of development of cross-domain recommender systems using extracting emotions in text data. In our system, we focus on the emotions contained in reviews and comments on content, and aim to improve the accuracy of extracted emotions from text data. We constructed an emotion-words dictionary with 13001 words to extract eight emotion features from user’ s preference information obtained from text data. In experiments, we conducted cross-domain recommendation of video games from the game sales platform ”steam” based on extracting user’s emotions from comments on various videos in the video sharing service ”Nicovideo”.

In this paper, we propose a recommender system for user group using rough sets. In recent years, with the advancement of information technology, recommender systems have been used in various services. On the other hand, recommender systems for multiple people are not as mainstream as those for a single person. The proposed method aims at providing recommendations that evenly reflects the opinions of multiple people in the user grope. In experiments, we used Rakuten Recipe data. As a result, the result of the questionnaire administered to the subjects showed that the average satisfaction level was higher than that of content-based filtering. In addition, we found that the recipes recommended by the recommendation method tended to have a large number of tags. analysis of rankings using decision rules.

We have investigated rule generation under uncertain information and proposed the framework of NIS-Apriori-based rule generation. We implicitly employed a pair [attribute,value] of an attribute and an attribute value as a descriptor. This paper focuses on descriptors for tables with continuous values and considers the relation to rule generation.

When extracting any information from an information table with incomplete information, following Lipski we only know the lower and upper bounds of the information. Methods of rough sets that are applied to data tables containing incomplete information are examined from the viewpoint of Lipski’s approach based on possible world semantics. It is clarified that the formula that is first used by Kryszkiewicz, which most of the authors use, only gives the lower bound of the lower approximation and the upper bound of the upper approximation. This is due to that the formula considers only the indiscernibility of missing values with another value. We extend Kryszkiewicz’s formula by considering the discernibility of missing values. As a result, the extended Kryszkiewicz’s formula gives the same approximations as those in terms of Lipski’s approach.

We argue that quotient sets based on equivalence relations, which abstract properties unrelated to the problem at hand, can be regarded as immersion support in current the problem. We create some concrete examples in metaverse spaces for experience.

Currently, one out of three women after childbirth experiences hand or wrist pain, and most of the pains appear from one to two months after childbirth, and hug may increase the burden on the hands and wrists. Therefore, we propose a system for determining poor posture of hug using posture estimation.

Currently, one out of three women after childbirth experiences hand or wrist pain, and most of the pains appear from one to two months after childbirth, and hug may increase the burden on the hands and wrists. Therefore, we analyzed the poor posture determining system for hug posture and musculoskeletal injuries.

This paper aims to analyze the effects of pseudo labels for training pose estimation models using semi-supervised learning. One of the problems in training pose estimation models is the high cost of labeling, which requires keypoint annotation of training data. Pseudo-labels estimated by other pose estimator models are able to improve the performance of another estimator. This paper investigates the trade-off between the number of annotations by humans and pseudo labels given by the model.

Fog obscuring images is a severe problem in observing a city using observation cameras installed at high altitudes, such as at the top of a mountain. In recent years, research has been conducted on Dehazing, which removes fog from an image and converts it into a clear image. However, most of these methods involve fog processing on clear images and transforming them by having a machine learning model learn image pairs with perfectly matched angles of view. It is difficult to collect clear and foggy image pairs with matched angles of view from cameras installed at high elevations due to angle-of-view control and wind effects. In addition, the actual fog is different from the images obtained by fog processing, making it difficult to apply the methods of previous studies to the city observation task. Therefore, this paper uses CycleGAN, which learns a set of sunny day images and a set of fog images and performs domain transformation. Also, images actually taken were used for the training images. As a result, we were able to remove some of the fog in the images captured by the observation camera. In the stitching task between the captured images, the fog images could not be stitched together but could be stitched together after the fog removal process.

Radiation transmission testing (RT) is used to evaluate the quality of internal welds of steel materials. In an existing study, the authors proposed a method to automatically determine the location of defects from film images of radiation transmission tests using a Convolutional Neural Network, and showed that the method is useful for assisting on-site engineers in inspections. However, the information on the object in the film image is manually input by visual reading. In this study, we propose a prototype model that recognizes alphabets, numbers, and symbols from film images using YOLOX.

Early detection and control of diseases, especially downy mildew, are very important for productivity and quality improvement in viticulture. In this study, we propose a simple method to train and infer the location of downy mildew from hyperspectral images using machine learning for the purpose of early detection and labor saving. Specifically, a vineyard is photographed at a fixed point using a spectrographic camera, and the image is divided into smaller partial images, which are then trained and inferred by the Convolutional Neural Network (CNN).In order to label the training data with "locations where downy mildew has occurred," it is necessary to confirm the presence or absence of downy mildew in the field over a long period of time, which requires a great deal of time and effort. In this study, we compare this method with a simple method of "labeling by photographing time". The results showed that the inference of the location of downy mildew by CNN tended to have a low recall rate and a high precision rate. The results also suggest that even a simple method of "labeling by photographing time" may be able to achieve a certain degree of precision. This suggested the possibility of detecting downy mildew even without data on the details of the location of downy mildew.

In fruit growing, there are many operations based on growers’ experience, and one of them is fruit thinning. Since this task is based on the grower’s experience, it is difficult for unskilled workers to determine how much fruit should be thinned. In this study, we propose a method for estimating the position and number of fruits using a three-dimensional model of an orchard based on photogrammetry. The proposed method uses COLMAP, a high-precision photogrammetric method, and YOLOv7, a fast object detection algorithm. The 3D model and camera position information were generated by applying COLMAP to the captured video images, and the fruit recognition images were extracted by applying YOLOv7. A fruit arrangement model was generated by connecting the 3D model, camera position information, and fruit recognition images. To verify the effectiveness of the proposed method, we conducted indoor experiments in which two fruits were placed in an indoor space. As a result of the verification, the number of fruits could be identified in the indoor experiment, but the coordinates of the fruits differed from those in the real world.

In recent years, there has been a great deal of research into recognizing human behavior.We focused on infrared sensor arrays as a method to solve two of these problems: (1) invasion of privacy and (2) high cost of installation and maintenance.The data acquired from infrared sensor arrays has very low resolution, making it difficult to precisely recognize actions.In this study, we propose an action recognition method that takes into account the hierarchy of actions as a method to improve the above-mentioned problem.

Plants can change their growth conditions due to external factors. In the management of crop cultivation, it is important to assess the growth conditions without harvesting the crops to minimize the reduction in productivity. In this study, we propose a method to assess the growth conditions of grapes based on different cultivation methods using hyperspectral imaging. By using a hyperspectral camera, It is possible to detect changes that are difficult to capture with a visible light camera.Specifically, we focused on vineyards with and without the installation of multi-sheet systems as a difference in cultivation methods. We captured images using a hyperspectral camera and calculated the Normalized Difference Vegetation Index (NDVI) from the captured images. By analyzing the temporal changes in NDVI, we evaluated the growth conditions.The results of the experiment suggested that by comparing the NDVI values after correcting for the effects of factors such as solar azimuth, angle, and imaging equipment, we can grasp the differences in growth conditions such as internal water content and sugar content in plants based on the presence or absence of mulch-sheet systems. Furthermore, the observation of NDVI indicated the potential for estimating Brix values.