Kazuki Shibuya’s research while affiliated with Meiji University and other places

Frost causes damage to crops. Predicting the frost occurrence in advance is highly valuable for practitioners taking possible frost-prevention measures. In our previous study, we have applied machine learning to forecast frost occurrence using two methods. To better support user with a potential trend of frost occurrence in a future period, we propose an integrated system of frost forecast taking the advantages of the two methods.

Causal modeling with time delay has been proposed as a method for predicting frost occurrence in a short period of time. In this method, environment factors are considered as cause, and used as input variables for prediction of frost. For coping with the uncertainty of prediction rooted in randomness of environment, a granulation of environment factors offers potential. In this study, we show that the accuracy of predicting frost occurrence can be improved by appropriately granulating each of the input environment factors involved.

Being a kind of natural phenomenon, frost occurrence is influenced by environment factors with an accumulated impact. The relation between environment factors and frost event is of cause-effect governed by a process taking place in time. Having cause-affect concerned, frost forecast is a problem of complex cause-effect rather than a complicated association and problem modelling plays a key role for the success of forecast. With limited data and lack of true physical model, a well-trained model by machine learning from data is only an approximation constructed on a sub-space of problem domain. As a continued study of causal modelling in frost forecast developed previously, this paper proposes an ensemble causal modelling to compensate the performance of individual models. Such an ensemble involves models with different length of time-delay so to provide a spectrum of early alarm of frost occurrence. Experiments are done using sensor data collected from a vineyard in Hokkaido, Japan.

Frost damage to agricultural products in Japan is by no means uncommon. Moreover, since there is no sensor capable of detecting a minute amount of frost with high accuracy, the presence of frost can be determined only by visual observation, making the study of frost damage in theˆeld challenging. Therefore, the actual conditions of frost damage have not been adequately elucidated. Unlike the vibration sensor, the frost sensor based on time domain re‰ectometry (TDR) has no malfunction-causing factors, and it can distinguish dew condensation and frost more clearly than the temperature-sensitive resistance element. However, improvement in the detection sensitivity of the TDR frost sensor is desired. We have developed a frost sensor with an accelerated cooling rate and reduced heat capacity by reducing the thickness of the TDR frost sensor. In this study, the detection sensitivities of TDR frost sensors of 1.6 and 0.075 mm thicknesses were compared under the conditions of dew condensation, frozen dew, and hoar frost generation. The ‰exible printed circuits (FPC) frost sensor that had a small thickness and thus a small heat capacity responded promptly to the environmental changes, such as radiation cooling, and the times of condensation, frozen dew, and hoar frost generation could be detected without delay. The amount of frost on the sensor also increased. Using the FPC frost sensor, dew condensation, frozen dew, and hoar frost could be detected in theˆeld where meteorological conditions were constantly changing. Various frost formations could be detected with high accuracy by combining this FPC frost sensor that can promptly detect environmental changes with a duty-cycle TDR device (manufactured by Kett Electric Laboratory) that can measure minute changes in the dielectric constant.

Crops are subject to frost damage every year. A method for detecting frost occurrence immediately would allow for earlier and more eŠective management strategies to be employed and ultimately reduce crop damage. We have developed a frost detecting sensor based on time domain reflectometry (TDR) that determines relative permittivity by measuring the propagation time of electromagnetic waves. Frost that formed on a sensor could be detected using the diŠereflnce between the relative permittivity of water, 80, and ice, 3.5. Two types of frost sensor, A and B, were developed. Sensor A was made with a glass-epoxy print-circuit board by etching two zigzag electrodes. Sensor B was made with a glass-composite print circuit board by etching two spiral electrodes. Sensor A was placed on Peltier devices to cool to approximately-3°C, artiˆcially inducing frost. This experiment was conducted in an acrylic box, 30×45×30 cm,ˆlled with air at 100 relative humidity. Relative permittivity and temperature of the sensor were measured at 10 s intervals. Sensor B was also placed on the Peltier devices to cool to approximately-3°C, also artiˆcially inducing the occurrence of frost. The experiment with sensor B was conducted in a constant temperature and humidity incubator. The relative humidity varied to 60, 70, and 80 . Relative permittivity and temperature of sensor B were measured with 10 s intervals. The relative permittivity changed when moisture condensed on both sensors and when condensed moisture turned to ice under the conditions of 60, 70, 80, and 100  R.H. The diŠerences of relative permittivity before and after frost formation were between 0.042 and 0.14. The TDR sensors distinguished frost from dew. As R.H. increased, the amount of frost or dew formed, measured as the relative permittivity, increased as well. We need to verify the shape of the sensor and improve the thickness of the printed circuit board in order to increase the sensitivity.

With the recent achievements in real world applications, being able to learn cause-effect has been expected as a new aim of artificial intelligence (AI) and machine learning (ML). As a preliminary attempt, causal modelling has been proposed for capturing relation between past observation of environment factors and future event of frost. This article continues explore methods of modelling and learning cause-effect relation in frost forecast. It first argues that the relation between environment factors and frost event is of cause-effect more than correlation, then discusses the involvement of time in modelling such cause-effect. Methods of modelling are discussed with their assumptions and rational behind. Performance comparison is provided.

To effectively protect plants from frost damage, an early alarm of frost can be helpful for growers. Frost is a localized phenomenon and can be quite variable across a small area, so predictive models developed with local data are preferred. As a climate phenomenon the occurrence of frost is closely related to multiple environment factors including temperature, humidity, radiation and more. This article proposes construction of predictive models using support vector machine approach to capture possible causal relation between these factors and frost. Such models trained with specific local data are expected to help frost forecast in a few hours ahead in the local area. Problem analysis, modeling methodology, and model ensemble are discussed, and experiments with real data are provided.